Generalization and optimization guarantees on the population loss often rely on uniform convergence based analysis, typically based on the Rademacher complexity of the predictors. The rich representation power of modern models has led to concerns about this approach. In this paper, we present generalization and optimization guarantees in terms of the complexity of the gradients, as measured by the Loss Gradient Gaussian Width (LGGW). First, we introduce generalization guarantees directly in terms of the LGGW under a flexible gradient domination condition. Second, we show that sample reuse in ERM does not make the empirical gradients deviate from the population gradients as long as the LGGW is small. Third, focusing on deep networks, we bound their single-sample LGGW in terms of the Gaussian width of the featurizer, i.e., the output of the last-but-one layer. To our knowledge, our generalization and optimization guarantees in terms of LGGW are the first results of its kind, and hold considerable promise towards quantitatively tight bounds for deep models.

Conservative Contextual Bandits (CCBs) address safety in sequential decision making by requiring that an agent's policy, along with minimizing regret, also satisfies a safety constraint: the performance is not worse than a baseline policy (e.g., the policy that the company has in production) by more than $(1+\alpha)$ factor. Prior work developed UCB-style algorithms for this problem in the multi-armed (Wu et al., 2016) and contextual linear (Kazerouni et al., 2017) settings. However, in practice the cost of the arms is often a non-linear function, and therefore existing UCB algorithms are ineffective in such settings. In this paper, we consider CCBs beyond the linear case and develop two algorithms $\mathtt{C\text{-}SquareCB}$ and $\mathtt{C\text{-}FastCB}$, using Inverse Gap Weighting (IGW) based exploration and an online regression oracle. We show that the safety constraint is satisfied in high probability and that the regret for $\mathtt{C\text{-}SquareCB}$ is sub-linear in horizon $T$, while the the regret for $\mathtt{C\text{-}FastCB}$ is first-order and is sub-linear in $L^*$, the cumulative loss of the optimal policy. Subsequently, we use a neural network for function approximation and online gradient descent as the regression oracle to provide $\tilde{\mathcal{O}}\big(\sqrt{KT} + K/\alpha\big) $ and $\tilde{\mathcal{O}}\big(\sqrt{KL^*} + K (1 + 1/\alpha)\big)$ regret bounds respectively. Finally, we demonstrate the efficacy of our algorithms on real world data, and show that they significantly outperform the existing baseline while maintaining the performance guarantee.

Neural Operators that directly learn mappings between function spaces, such as Deep Operator Networks (DONs) and Fourier Neural Operators (FNOs), have received considerable attention. Despite the universal approximation guarantees for DONs and FNOs, there is currently no optimization convergence guarantee for learning such networks using gradient descent (GD). In this paper, we address this open problem by presenting a unified framework for optimization based on GD and applying it to establish convergence guarantees for both DONs and FNOs. In particular, we show that the losses associated with both of these neural operators satisfy two conditions—restricted strong convexity (RSC) and smoothness—that guarantee a decrease on their loss values due to GD. Remarkably, these two conditions are satisfied for each neural operator due to different reasons associated with the architectural differences of the respective models. One takeaway that emerges from the theory is that wider networks benefit optimization convergence guarantees for both DONs and FNOs. We present empirical results on canonical operator learning problems to support our theoretical results and find that larger widths benefit training.

Combining gradient compression with adaptive optimizers is a highly desirable goal in distributed learning, with potential benefits in both fewer communication rounds and less per-round communication. In spite of preliminary empirical promise, certain major challenges in the convergence analysis of such methods have stayed open: handling compression based approximation of both first and second moments (pre-conditioner) which appear as a ratio; avoiding dependence on the number of parameters, which is extremely large in modern deep models; and providing high-probability guarantees instead of in-expectation, which can hide high variance behavior. In this work, we introduce a family of Sketched Adaptive Distributed Learning (SADL) algorithms which can use suitable unbiased gradient sketching for compression with suitable adaptive optimization algorithms. As our main contribution, we provide theoretical convergence guarantees of SADL algorithms which addresses all of the existing challenges. In particular, our guarantees hold with high probability, picks up only a logarithmic dependence on the number of parameters, and the first and second moment approximation is handled precisely yielding a dependence on the intrinsic dimension of the loss Hessian, which is significantly smaller than the full dimensionality of deep learning models. Empirically, the SADL algorithms are shown to be competitive with and often outperform baselines on both vision and language tasks, in both supervised fine-tuning and training-from-scratch regimes. Further, the SADL algorithms are also competitive with the state-of-the-art communication-efficient distributed learning algorithms based on error feedback.

Communication cost and privacy are two major considerations in federated learning (FL). For communication cost, gradient compression by sketching the clients’ transmitted model updates is often used for reducing per‐round communication. For privacy, the Gaussian mechanism (GM), which consists of clipping updates and adding Gaussian noise, is commonly used to guarantee client‐level differential privacy. Existing literature on private FL analyzes privacy of sketching and GM in an isolated manner, illustrating that sketching provides privacy determined by the sketching dimension and that GM has to supply any additional desired privacy. In this paper, we introduce the Sketched Gaussian Mechanism (SGM), which directly combines sketching and the Gaussian mechanism for privacy. Using Rényi-DP tools, we present a joint analysis of SGM's overall privacy guarantee, which is significantly more flexible and sharper compared to isolated analysis of sketching and GM privacy. In particular, we prove that the privacy level of SGM for a fixed noise magnitude is proportional to $1/\sqrt{b}$, where $b$ is the sketching dimension, indicating that (for moderate $b$) SGM can provide much stronger privacy guarantees than the original GM under the same noise budget. We demonstrate the application of SGM to FL with either gradient descent or adaptive server optimizers, and establish theoretical results on optimization convergence, which exhibits only a logarithmic dependence on the number of parameters $d$. Experimental results confirm that at the same privacy level, SGM based FL is at least competitive with non‐sketching private FL variants and outperforms them in some settings. Moreover, using adaptive optimization at the server improves empirical performance while maintaining the privacy guarantees.

Invariant Risk Minimization (IRM) (Arjovsky et al., 2020) proposes an optimization scheme that uses causal features to improve generalization. However, in most realizations, it does not have an explicit feature selection strategy. Prior investigation (Rosenfeld et al., 2020; Zhang et al., 2023) reveals failure cases when searching for causal features, and in light of these concerns, recent work has demonstrated the promise of using sparsity (Zhou et al., 2022; Fan et al., 2024) in IRM, and we make two specific contributions on that theme. First, for the original sparse IRM formulation, we present the first correct non-asymptotic analysis of the effectiveness of sparsity for selecting invariant features. We show that sparse IRM with $L_0$ constraints can select invariant features and ignore spurious and random features. We show that sample complexity depends polynomially on the number of invariant features and otherwise logarithmically on the ambient dimensionality. Second, we present the first invariant feature recovery guarantees with a computationally-efficient implementation of such sparse IRM based on iterative hard thresholding. Prior methods are limited to combinatorially searching over the space of all sparse models, but we present a different loss function. We show this new optimization implies recovery of invariant features under standard assumptions. We present empirical results on standard benchmark datasets to demonstrate the effectiveness and efficiency of the proposed sparse IRM models.

Several application domains, especially in science and medicine, benefit tremendously from acquiring high-resolution images of objects and phenomena of interest. Recognizing this need, generative models for super-resolution (SR) have emerged as a promising approach for such data generation. However, when training data are scarce due to high acquisition costs, such models struggle and often fail to capture the true data distribution due to insufficient data and domain knowledge. While transfer learning, domain adaptation, or few shot learning of such generative models can be a reasonable approach, most existing large scale generative models have been (pre)trained on natural images and it is unclear if such models can be seamlessly transferred to say medical images. In this paper, we propose Multi-Stage Probabilistic Super Resolution (MSP-SR), a cascaded few-shot learning framework for super-resolution through multi-stage transfer learning. At a high level, MSP-SR first transfers a generative model from out-of-domain to in-domain, e.g., from natural to medical images, and then from in-domain to the target application. We present the details based on conditional diffusion models and validate MSP-SR on multiple Magnetic Resonance Imaging (MRI) datasets, demonstrating that MSP-SR persistently and usually significantly outperforms direct fine-tuning (DFT) approaches as well as SR baselines. Further, MSP-SR empirically provides more accurate characterization of uncertainty in SR compared to DFT.

We consider the problem of reward maximization in the dueling bandit setup along with constraints on resource consumption. As in the classic dueling bandits, at each round the learner has to choose a pair of items from a set of $K$ items and observe a relative feedback for the current pair. Additionally, for both items, the learner also observes a vector of resource consumptions. The objective of the learner is to maximize the cumulative reward, while ensuring that the total consumption of any resource is within the allocated budget. We show that due to the relative nature of the feedback, the problem is more difficult than its bandit counterpart and that without further assumptions the problem is not learnable from a regret minimization perspective. Thereafter, by exploiting assumptions on the available budget, we provide an EXP3 based dueling algorithm that also considers the associated consumptions and show that it achieves an $\tilde{\mathcal{O}}\left(\big({\frac{OPT^{(b)}}{B}}+1\big)K^{1/3}T^{2/3}\right)$ regret, where $OPT^{(b)}$ is the optimal value and $B$ is the available budget. Finally, we provide numerical simulations to demonstrate the efficacy of our proposed method.

Recent works have shown a reduction from contextual bandits to online regression under a realizability assumption (Foster and Rakhlin, 2020; Foster and Krishnamurthy, 2021). In this work, we investigate the use of neural networks for such online regression and associated Neural Contextual Bandits (NeuCBs). Using existing results for wide networks, one can readily show a ${\mathcal{O}}(\sqrt{T})$ regret for online regression with square loss, which via the reduction implies a ${\mathcal{O}}(\sqrt{K} T^{3/4})$ regret for NeuCBs. Departing from this standard approach, we first show a $\mathcal{O}(\log T)$ regret for online regression with almost convex losses that satisfy QG (Quadratic Growth) condition, a generalization of the PL (Polyak-\L ojasiewicz) condition, and that have a unique minima. Although not directly applicable to wide networks since they do not have unique minima, we show that adding a suitable small random perturbation to the network predictions surprisingly makes the loss satisfy QG with unique minima. Based on such a perturbed prediction, we show a ${\mathcal{O}}(\log T)$ regret for online regression with both squared loss and KL loss, and subsequently convert these respectively to $\tilde{\mathcal{O}}(\sqrt{KT})$ and $\tilde{\mathcal{O}}(\sqrt{KL^*} + K)$ regret for NeuCB, where $L^*$ is the loss of the best policy. Separately, we also show that existing regret bounds for NeuCBs are $\Omega(T)$ or assume i.i.d. contexts, unlike this work. Finally, our experimental results on various datasets demonstrate that our algorithms, especially the one based on KL loss, persistently outperform existing algorithms.

Contextual bandit algorithms aim to identify the optimal arm with the highest reward among a set of candidates, based on the accessible contextual information. Among these algorithms, neural contextual bandit methods have shown generally superior performances against linear and kernel ones, due to the representation power of neural networks. However, similar to other neural network applications, neural bandit algorithms can be vulnerable to adversarial attacks or corruptions on the received labels (i.e., arm rewards), which can lead to unexpected performance degradation without proper treatments. As a result, it is necessary to improve the robustness of neural bandit models against potential reward corruptions. In this work, we propose a novel neural contextual bandit algorithm named R-NeuralUCB, which utilizes a novel context-aware Gradient Descent (GD) training strategy to improve the robustness against adversarial reward corruptions. Under over-parameterized neural network settings, we provide regret analysis for R-NeuralUCB to quantify reward corruption impacts, without the commonly adopted arm separateness assumption in existing neural bandit works. We also conduct experiments against baselines on real data sets under different scenarios, in order to demonstrate the effectiveness of our proposed R-NeuralUCB.

The high communication cost between the server and the clients is a significant bottleneck in scaling distributed learning for overparametrized deep models. One popular approach for reducing this communication overhead is randomized sketching. However, existing theoretical analyses for sketching-based distributed learning (sketch-DL) either incur a prohibitive dependence on the ambient dimension or need additional restrictive assumptions such as heavy-hitters. Nevertheless, despite existing pessimistic analyses, empirical evidence suggests that sketch-DL is competitive with its uncompressed counterpart, thus motivating a sharper analysis. In this work, we introduce a sharper ambient dimension-independent convergence analysis for sketch-DL using the second-order geometry specified by the loss Hessian. Our results imply ambient dimension-independent communication complexity for sketch-DL. We present empirical results both on the loss Hessian and overall accuracy of sketch-DL supporting our theoretical results. Taken together, our results provide theoretical justification for the observed empirical success of sketch-DL.

The Landsat program is the longest-running Earth observation program in history, with 50+ years of data acquisition by 8 satellites. The multispectral imagery captured by sensors onboard these satellites is critical for a wide range of scientific fields. Despite the increasing popularity of deep learning and remote sensing, the majority of researchers still use decision trees and random forests for Landsat image analysis due to the prevalence of small labeled datasets and lack of foundation models. In this paper, we introduce SSL4EO-L, the first ever dataset designed for Self-Supervised Learning for Earth Observation for the Landsat family of satellites (including 3 sensors and 2 product levels) and the largest Landsat dataset in history (5M image patches). Additionally, we modernize and re-release the L7 Irish and L8 Biome cloud detection datasets, and introduce the first ML benchmark datasets for Landsats 4–5 TM and Landsat 7 ETM+ SR. Finally, we pre-train the first foundation models for Landsat imagery using SSL4EO-L and evaluate their performance on multiple semantic segmentation tasks. All datasets and model weights are available via the TorchGeo library, making reproducibility and experimentation easy, and enabling scientific advancements in the burgeoning field of remote sensing for a multitude of downstream applications.

In this paper we study the problem of lower bounding the minimum eigenvalue of the neural tangent kernel (NTK) at initialization, an important quantity for the theoretical analysis of training in neural networks. We consider feedforward neural networks with smooth activation functions. Without any distributional assumptions on the input, we present a novel result: we show that for suitable initialization variance, $\widetilde{\Omega}(n)$ width, where $n$ is the number of training samples, suffices to ensure that the NTK at initialization is positive definite, improving prior results for smooth activations under our setting. Prior to our work, the sufficiency of linear width has only been shown either for networks with ReLU activation functions, and sublinear width has been shown for smooth networks but with additional conditions on the distribution of the data. The technical challenge in the analysis stems from the layerwise inhomogeneity of smooth activation functions and we handle the challenge using {\em generalized} Hermite series expansion of such activations.

Sub-seasonal forecasting (SSF) is the prediction of key climate variables such as temperature and precipitation on the 2-week to 2-month time horizon. Skillful SSF would have substantial societal value in areas such as agricultural productivity, hydrology and water resource management, and emergency planning for extreme events such as droughts and wildfires. Despite its societal importance, SSF has stayed a challenging problem compared to both short-term weather forecasting and long-term seasonal forecasting. Recent studies have shown the potential of machine learning (ML) models to advance SSF. In this paper, for the first time, we perform a fine-grained comparison of a suite of modern ML models with start-of-the-art physics-based dynamical models from the Subseasonal Experiment (SubX) project for SSF in the western contiguous United States. Additionally, we explore mechanisms to enhance the ML models by using forecasts from dynamical models. Empirical results illustrate that, on average, ML models outperform dynamical models while the ML models tend to generate forecasts with conservative magnitude compared to the SubX models. Further, we illustrate that ML models make forecasting errors under extreme weather conditions, e.g., cold waves due to the polar vortex, highlighting the need for separate models for extreme events. Finally, we show that suitably incorporating dynamical model forecasts as inputs to ML models can substantially improve the forecasting performance of the ML models. The SSF dataset constructed for the work and code for the ML models are released along with the paper for the benefit of the artificial intelligence community.

Remotely sensed geospatial data are critical for applications including precision agriculture, urban planning, disaster monitoring and response, and climate change research, among others. Deep learning methods are particularly promising for modeling many remote sensing tasks given the success of deep neural networks in similar computer vision tasks and the sheer volume of remotely sensed imagery available. However, the variance in data collection methods and handling of geospatial metadata make the application of deep learning methodology to remotely sensed data nontrivial. For example, satellite imagery often includes additional spectral bands beyond red, green, and blue and must be joined to other geospatial data sources that can have differing coordinate systems, bounds, and resolutions. To help realize the potential of deep learning for remote sensing applications, we introduce TorchGeo, a Python library for integrating geospatial data into the PyTorch deep learning ecosystem. TorchGeo provides data loaders for a variety of benchmark datasets, composable datasets for generic geospatial data sources, samplers for geospatial data, and transforms that work with multispectral imagery. TorchGeo is also the first library to provide pre-trained models for multispectral satellite imagery (e.g., models that use all bands from the Sentinel-2 satellites), allowing for advances in transfer learning on downstream remote sensing tasks with limited labeled data. We use TorchGeo to create reproducible benchmark results on existing datasets and benchmark our proposed method for preprocessing geospatial imagery on the fly. TorchGeo is open source and available on GitHub: https://github.com/microsoft/torchgeo.

Recent years have seen advances in generalization bounds for noisy stochastic algorithms, especially stochastic gradient Langevin dynamics (SGLD) based on stability (Mou et al., 2018; Li et al., 2020) and information theoretic approaches (Xu & Raginsky, 2017; Negrea et al., 2019; Steinke & Zakynthinou, 2020). In this paper, we unify and substantially generalize stability based generalization bounds and make three technical contributions. First, we bound the generalization error in terms of expected (not uniform) stability which arguably leads to quantitatively sharper bounds. Second, as our main contribution, we introduce Exponential Family Langevin Dynamics (EFLD), a substantial generalization of SGLD, which includes noisy versions of Sign-SGD and quantized SGD as special cases. We establish data dependent expected stability based generalization bounds for any EFLD algorithm with a O(1/n) sample dependence and dependence on gradient discrepancy rather than the norm of gradients, yielding significantly sharper bounds. Third, we establish optimization guarantees for special cases of EFLD. Further, empirical results on benchmarks illustrate that our bounds are non-vacuous, quantitatively sharper than existing bounds, and behave correctly under noisy labels.

Many bandit problems are characterized by the learner making decisions under constraints. The learner in Linear Contextual Bandits with Knapsacks (LinCBwK) receives a resource consumption vector in addition to a scalar reward in each time step which are both linear functions of the context corresponding to the chosen arm. For a fixed time horizon $T$, the goal of the learner is to maximize rewards while ensuring resource consumptions do not exceed a pre-specified budget. We present algorithms and characterize regret for LinCBwK in the smoothed setting where base context vectors are assumed to be perturbed by Gaussian noise. We consider both the stochastic and adversarial settings for the base contexts, and our analysis of stochastic LinCBwK can be viewed as a warm-up to the more challenging adversarial LinCBwK. For the stochastic setting, we obtain $O(\sqrt{T})$ additive regret bounds compared to the best context dependent fixed policy. The analysis combines ideas for greedy parameter estimation in \cite{kmrw18, siwb20} and the primal-dual paradigm first explored in \cite{agde17, agde14}. Our main contribution is an algorithm with $O(\log T)$ competitive ratio relative to the best context dependent fixed policy for the adversarial setting. The algorithm for the adversarial setting employs ideas from the primal-dual framework \cite{agde17, agde14} and a novel adaptation of the doubling trick \cite{isss19}.

We improve the theoretical and empirical performance of neural-network(NN)-based active learning algorithms for the non-parametric streaming setting. In particular, we introduce two regret metrics by minimizing the population loss that are more suitable in active learning than the one used in state-of-the-art (SOTA) related work. Then, the proposed algorithm leverages the powerful representation of NNs for both exploitation and exploration, has the query decision-maker tailored for $k$-class classification problems with the performance guarantee, utilizes the full feedback, and updates parameters in a more practical and efficient manner. These careful designs lead to an instance-dependent regret upper bound, roughly improving by a multiplicative factor $O(\log T)$ and removing the curse of input dimensionality. Furthermore, we show that the algorithm can achieve the same performance as the Bayes-optimal classifier in the long run under the hard-margin setting in classification problems. In the end, we use extensive experiments to evaluate the proposed algorithm and SOTA baselines, to show the improved empirical performance.

Recent empirical work on stochastic gradient descent (SGD) applied to over-parameterized deep learning has shown that most gradient components over epochs are quite small. Inspired by such observations, we rigorously study properties of Truncated SGD (T-SGD), that truncates the majority of small gradient components to zeros. Considering non-convex optimization problems, we show that the convergence rate of T-SGD matches the order of vanilla SGD. We also establish the generalization error bound for T-SGD. Further, we propose Noisy Truncated SGD (NT-SGD), which adds Gaussian noise to the truncated gradients. We prove that NT-SGD has the same convergence rate as T-SGD for non-convex optimization problems. We demonstrate that with the help of noise, NT-SGD can provably escape from saddle points and requires less noise compared to previous related work. We also prove that NT-SGD achieves better generalization error bound compared to T-SGD because of the noise. Our generalization analysis is based on uniform stability and we show that additional noise in the gradient update can boost the stability. Our experiments on a variety of benchmark datasets with various network architectures verify that NT-SGD matches the speed and accuracy of vanilla SGD while effectively working with sparse gradients, and can successfully escape poor local minima.

Sub-seasonal forecasting (SSF) focuses on predicting key variables such as temperature and precipitation on the 2-week to 2-month time scale. Skillful SSF would have immense societal value in such areas as agricultural productivity, water resource management, and emergency planning for extreme weather events. However, SSF is considered more challenging than either weather prediction or even seasonal prediction, and is still a largely understudied problem. In this paper, we carefully investigate 10 Machine Learning (ML) approaches to sub-seasonal temperature forecasting over the contiguous U.S. on the SSF dataset we collect, including a variety of climate variables from the atmosphere, ocean, and land. Because of the complicated atmosphere-land-ocean couplings and the limited amount of good quality observational data, SSF imposes a great challenge for ML despite the recent advances in various domains. Our results indicate that suitable ML models, e.g., XGBoost, to some extent, capture the predictability on sub-seasonal time scales and can outperform the climatological baselines, while Deep Learning (DL) models barely manage to match the best results with carefully designed architecture. Besides, our analysis and exploration provide insights on important aspects to improve the quality of sub-seasonal forecasts, e.g., feature representation and model architecture. The SSF dataset and code are released with this paper for use by the broader research community.

Subseasonal climate forecasting is the task of predicting climate variables, such as temperature and precipitation, in a two-week to two-month time horizon. The primary predictors for such prediction problem are spatio-temporal satellite and ground measurements of a variety of climate variables in the atmosphere, ocean, and land, which however have rather limited predictive signal at the subseasonal time horizon. We propose a carefully constructed spatial hierarchical Bayesian regression model that makes use of the inherent spatial structure of the subseasonal climate prediction task. We use our Bayesian model to then derive decision-theoretically optimal point estimates with respect to various performance measures of interest to climate science. As we show, our approach handily improves on various off-the-shelf ML baselines. Since our method is based on a Bayesian framework, we are also able to quantify the uncertainty in our predictions, which is particularly crucial for difficult tasks such as the subseasonal prediction, where we expect any model to have considerable uncertainty at different test locations under different scenarios.

Bandit learning algorithms typically involve the balance of exploration and exploitation. However, in many practical applications, worst-case scenarios needing systematic exploration are seldom encountered. In this work, we consider a smoothed setting for structured linear contextual bandits where the adversarial contexts are perturbed by Gaussian noise and the unknown parameter $\theta^*$ has structure, e.g., sparsity, group sparsity, low rank, etc. We propose simple greedy algorithms for both the single- and multi-parameter (i.e., different parameter for each context) settings and provide a unified regret analysis for $\theta^*$ with any assumed structure. The regret bounds are expressed in terms of geometric quantities such as Gaussian widths associated with the structure of $\theta^*$. We also obtain sharper regret bounds compared to earlier work for the unstructured $\theta^*$ setting as a consequence of our improved analysis. We show there is implicit exploration in the smoothed setting where a simple greedy algorithm works.

By chaining a sequence of differentiable invertible transformations, normalizing flows (NF) provide an expressive method of posterior approximation, exact density evaluation, and sampling. The trend in normalizing flow literature has been to devise deeper, more complex transformations to achieve greater flexibility. We propose an alternative: Gradient Boosted Normalizing Flows (GBNF) model a density by successively adding new NF components with gradient boosting. Under the boosting framework, each new NF component optimizes a weighted likelihood objective, resulting in new components that are fit to the suitable residuals of the previously trained components. The GBNF formulation results in a mixture model structure, whose flexibility increases as more components are added. Moreover, GBNFs offer a wider, as opposed to strictly deeper, approach that improves existing NFs at the cost of additional training---not more complex transformations. We demonstrate the effectiveness of this technique for density estimation and, by coupling GBNF with a variational autoencoder, generative modeling of images. Our results show that GBNFs outperform their non-boosted analog, and, in some cases, produce better results with smaller, simpler flows.

While stochastic gradient descent (SGD) and variants have been surprisingly successful for training deep nets, several aspects of the optimization dynamics and generalization are still not well understood. In this paper, we present new empirical observations and theoretical results on both the optimization dynamics and generalization behavior of SGD for deep nets based on the Hessian of the training loss and associated quantities. We consider three specific research questions: (1) what is the relationship between the Hessian of the loss and the second moment of stochastic gradients (SGs)? (2) how can we characterize the stochastic optimization dynamics of SGD with fixed step sizes based on the first and second moments of SGs? and (3) how can we characterize a scale-invariant generalization bound of deep nets based on the Hessian of the loss? Throughout the paper, we support theoretical results with empirical observations, with experiments on synthetic data, MNIST, and CIFAR-10, with different batch sizes, and with different difficulty levels by synthetically adding random labels.

An important family of problems in climate science focus on finding predictive relationships between various climate variables. In this paper, we consider the problem of predicting monthly deseasonalized land temperature at different locations worldwide based on sea surface temperature (SST). Contrary to popular belief on the trade-off between (a) simple interpretable but inaccurate models and (b) complex accurate but uninterpretable models, we introduce a weighted Lasso model for the problem which yields interpretable results while being highly accurate. Covariate weights in the regularization of weighted Lasso are pre-determined, and proportional to the spatial distance of the covariate (sea surface location) from the target (land location). We establish finite sample estimation error bounds for weighted Lasso, and illustrate its superior empirical performance and interpretability over complex models such as deep neural networks (Deep nets) and gradient boosted trees (GBT). We also present a detailed empirical analysis of what went wrong with Deep nets here, which may serve as a helpful guideline for application of Deep nets to small sample scientific problems.

Recent years have seen advances in optimizing large scale statistical estimation problems. In statistical learning settings iterative optimization algorithms have been shown to enjoy geometric convergence. While powerful, such results only hold for the original dataset, and may face computational challenges when the sample size is large. In this paper, we study sketched iterative algorithms, in particular sketched-PGD (projected gradient descent) and sketched-SVRG (stochastic variance reduced gradient) for structured generalized linear model, and illustrate that these methods continue to have geometric convergence to the statistical error under suitable assumptions. Moreover, the sketching dimension is allowed to be even smaller than the ambient dimension, thus can lead to significant speed-ups. The sketched iterative algorithms introduced provide an additional dimension to study the trade-offs between statistical accuracy and time.

Several important families of computational and statistical results in machine learning and randomized algorithms rely on uniform bounds on quadratic forms of random vectors or matrices. Such results include the Johnson-Lindenstrauss (J-L) Lemma, the Restricted Isometry Property (RIP), randomized sketching algorithms, and approximate linear algebra. The existing results critically depend on statistical independence, e.g., independent entries for random vectors, independent rows for random matrices, etc., which prevent their usage in dependent or adaptive modeling settings. In this paper, we show that such independence is in fact not needed for such results which continue to hold under fairly general dependence structures. In particular, we present uniform bounds on random quadratic forms of stochastic processes which are conditionally independent and sub-Gaussian given another (latent) process. Our setup allows general dependencies of the stochastic process on the history of the latent process and the latent process to be influenced by realizations of the stochastic process. The results are thus applicable to adaptive modeling settings and also allows for sequential design of random vectors and matrices. We also discuss stochastic process based forms of J-L, RIP, and sketching, to illustrate the generality of the results.

Can machine learning models be easily fooled? Despite the recent surge of interest in learned adversarial attacks in other domains, in the context of recommendation systems this question has mainly been answered using hand-engineered fake user profiles. This paper attempts to reduce this gap. We provide a formulation for learning to attack a recommender as a repeated general-sum game between two players, i.e., an adversary and a recommender oblivious to the adversary's existence. We consider the challenging case of poisoning attacks, which focus on the training phase of the recommender model. We generate adversarial user profiles targeting subsets of users or items, or generally the top-K recommendation quality. Moreover, we ensure that the adversarial user profiles remain unnoticeable by preserving proximity of the real user rating/interaction distribution to the adversarial fake user distribution. To cope with the challenge of the adversary not having access to the gradient of the recommender's objective with respect to the fake user profiles, we provide a non-trivial algorithm building upon zero-order optimization techniques. We offer a wide range of experiments, instantiating the proposed method for the case of the classic popular approach of a low-rank recommender, and illustrating the extent of the recommender's vulnerability to a variety of adversarial intents. These results can serve as a motivating point for more research into recommender defense strategies against machine learned attacks.

Topic modeling enables exploration and compact representation of a corpus. The CaringBridge (CB) dataset is a massive collection of journals written by patients and caregivers during a health crisis. Topic modeling on the CB dataset, however, is challenging due to the asynchronous nature of multiple authors writing about their health journeys. To overcome this challenge we introduce the Dynamic Author-Persona topic model (DAP), a probabilistic graphical model designed for temporal corpora with multiple authors. The novelty of the DAP model lies in its representation of authors by a persona---where personas capture the propensity to write about certain topics over time. Further, we present a regularized variational inference (RVI) algorithm, which we use to encourage the DAP model's personas to be distinct. Our results show significant improvements over competing topic models---particularly after regularization, and highlight the DAP model's unique ability to capture common journeys shared by different authors.

In machine learning and data mining, linear models have been widely used to model the response as parametric linear functions of the predictors. To relax such stringent assumptions made by parametric linear models, additive models consider the response to be a summation of unknown transformations applied on the predictors; in particular, additive isotonic models (AIMs) assume the unknown transformations to be monotone. In this paper, we introduce sparse linear isotonic models (SLIMs) for high-dimensional problems by hybridizing ideas in parametric sparse linear models and AIMs, which enjoy a few appealing advantages over both. In the high-dimensional setting, a two-step algorithm is proposed for estimating the sparse parameters as well as the monotone functions over predictors. Under mild statistical assumptions, we show that the algorithm can accurately estimate the parameters. Promising preliminary experiments are presented to support the theoretical results.

Extracting common narratives from multi-author dynamic text corpora requires complex models, such as the Dynamic Author Persona (DAP) topic model. However, such models are complex and can struggle to scale to large corpora, often because of challenging non-conjugate terms. To overcome such challenges, we adapt new ideas in approximate inference to the DAP model, resulting in the DAP Performed Exceedingly Rapidly (DAPPER) topic model. Specifically, we develop Conjugate-Computation Variational Inference (CVI) based variational Expectation-Maximization (EM) for learning the model, yielding fast, closed form updates for each document, replacing iterative optimization in earlier work. Our results show significant improvements in model fit and training time without needing to compromise the model's temporal structure or the application of Regularized Variation Inference (RVI). We demonstrate the scalability and effectiveness of the DAPPER model on multiple datasets, including the CaringBridge corpus - a collection of 9 million journals written by 200,000 authors during health crises.

Multi-response linear models aggregate a set of vanilla linear models by assuming correlated noise across them, which has an unknown covariance structure. To find the coefficient vector, estimators with a joint approximation of the noise covariance are often preferred than the simple linear regression in view of their superior empirical performance, which can be generally solved by alternating-minimization type procedures. Due to the non-convex nature of such joint estimators, the theoretical justification of their efficiency is typically challenging. The existing analyses fail to fully explain the empirical observations due to the assumption of resampling on the alternating procedures, which requires access to fresh samples in each iteration. In this work, we present a resampling-free analysis for the alternating minimization algorithm applied to the multi-response regression. In particular, we focus on the high-dimensional setting of multi-response linear models with structured coefficient parameter, and the statistical error of the parameter can be expressed by the complexity measure, Gaussian width, which is related to the assumed structure. More importantly, to the best of our knowledge, our result reveals for the first time that the alternating minimization with random initialization can achieve the same performance as the well-initialized one when solving this multi-response regression problem. Experimental results support our theoretical developments.

Learning to interact with users and discover their preferences is central in most web applications, with recommender systems being a notable example. From such a perspective, merging interactive learning algorithms with recommendation models is natural. While recent literature has explored the idea of combining collaborative filtering approaches with bandit techniques, there exist two limitations: (1) they usually consider Gaussian rewards, which are not suitable for implicit feedback data powering most recommender systems, and (2) they are restricted to the one-item recommendation setting while typically a list of recommendations is given. In this paper, to address these limitations, apart from Gaussian rewards we also consider Bernoulli rewards, the latter being suitable for dyadic data. Also, we consider two user click models: the one-item click/no-click model, and the cascade click model which is suitable for top-K recommendations. For these settings, we propose novel machine learning algorithms that learn to interact with users by learning the underlying parameters collaboratively across users and items. We provide an extensive empirical study, which is the first to illustrate all pairwise empirical comparisons across different interactive learning algorithms for recommendation. Our experiments demonstrate that when the number of users and items is large, propagating the feedback across users and items while learning latent features is the most effective approach for systems to learn to interact with the users.

Alzheimer’s disease (AD), the most common type of dementia, not only imposes a huge financial burden on the health care system, but also a psychological and emotional burden on patients and their families. There is thus an urgent need to infer trajectories of cognitive performance over time and identify biomarkers predictive of the progression. In this article, we propose the multi-task learning with fused Laplacian sparse group lasso model, which can identify biomarkers closely related to cognitive measures due to its sparsity-inducing property, and model the disease progression with a general weighted (undirected) dependency graphs among the tasks. An efficient alternative directions method of multipliers based optimization algorithm is derived to solve the proposed non-smooth objective formulation. The effectiveness of the proposed model is demonstrated by its superior prediction performance over multiple state-of-the-art methods and accurate identification of compact sets of cognition-relevant imaging biomarkers that are consistent with prior medical studies.

Future projection of climate is typically obtained by combining outputs from multiple Earth System Models (ESMs) for several climate variables such as temperature and precipitation. While IPCC has traditionally used a simple model output average, recent work has illustrated potential advantages of using a multitask learning (MTL) framework for projections of individual climate variables. In this paper we introduce a framework for hierarchical multitask learning (HMTL) with two levels of tasks such that each super-task, i.e., task at the top level, is itself a multitask learning problem over sub-tasks. For climate projections, each super-task focuses on projections of specific climate variables spatially using an MTL formulation. For the proposed HMTL approach, a group lasso regularization is added to couple parameters across the super-tasks, which in the climate context helps exploit relationships among the behavior of different climate variables at a given spatial location. We show that some recent works on MTL based on learning task dependency structures can be viewed as special cases of HMTL. Experiments on synthetic and real climate data show that HMTL produces better results than decoupled MTL methods applied separately on the super-tasks and HMTL significantly outperforms baselines for climate projection.

In many recommendation settings, the candidate items for recommendation are associated with a maximum capacity, i.e., number of seats in a Point-of-Interest (POI) or number of item copies in the inventory. However, despite the prevalence of the capacity constraint in the recommendation process, the existing recommendation methods are not designed to optimize for respecting such a constraint. Towards closing this gap, we propose Recommendation with Capacity Constraints -- a framework that optimizes for both recommendation accuracy and expected item usage that respects the capacity constraints. We show how to apply our method to three state-of-the-art latent factor recommendation models: probabilistic matrix factorization (PMF), bayesian personalized ranking (BPR) for item recommendation, and geographical matrix factorization (GeoMF) for POI recommendation. Our experiments indicate that our framework is effective for providing good recommendations while taking the limited resources into consideration. Interestingly, our methods are shown in some cases to further improve the top-N recommendation quality of the respective unconstrained models.

In this paper, we consider the use of structure learning methods for probabilistic graphical models to identify statistical dependencies in high-dimensional physical processes. Such processes are often synthetically characterized using PDEs (partial differential equations) and are observed in a variety of natural phenomena. In this paper, we present ACLIME-ADMM, an efficient two-step algorithm for adaptive structure learning, which decides a suitable edge specific threshold in a data-driven statistically rigorous manner. Both steps of our algorithm use (inexact) ADMM to solve suitable linear programs, and all iterations can be done in closed form in an efficient block parallel manner. We compare ACLIME-ADMM with baselines on both synthetic data simulated by PDEs that model advection-diffusion processes, and real data of daily global geopotential heights to study information flow in the atmosphere. ACLIME-ADMM is shown to be efficient, stable, and competitive, usually better than the baselines especially on difficult problems. On real data, ACLIME-ADMM recovers the underlying structure of global atmospheric circulation, including switches in wind directions at the equator and tropics entirely from the data.

We consider the problem of modeling data matrices with locally low rank (LLR) structure, a generalization of the popular low rank structure widely used in a variety of real world application domains ranging from medical imaging to recommendation systems. While LLR modeling has been found to be promising in real world application domains, limited progress has been made on the design of scalable algorithms for such structures. In this paper, we consider a convex relaxation of LLR structure, and propose an efficient algorithm based on dual projected gradient descent (D-PGD) for computing the proximal operator. While the original problem is non-smooth, so that primal (sub)gradient algorithms will be slow, we show that the proposed D-PGD algorithm has geometrical convergence rate. We present several practical ways to further speed up the computations, including acceleration and approximate SVD computations. With experiments on both synthetic and real data from MRI (magnetic resonance imaging) denoising, we illustrate the superior performance of the proposed D-PGD algorithm compared to several baselines.

In this paper, we investigate general single-index models (SIMs) in high dimensions. Based on U-statistics, we propose two types of robust estimators for the recovery of model parameters, which can be viewed as generalizations of several existing algorithms for one-bit compressed sensing (1-bit CS). With minimal assumption on noise, the statistical guarantees are established for the generalized estimators under suitable conditions, which allow general structures of underlying parameter. Moreover, the proposed estimator is novelly instantiated for SIMs with monotone transfer function, and the obtained estimator can better leverage the monotonicity. Experimental results are provided to support our theoretical analyses.

Quantile regression aims at modeling the conditional median and quantiles of a response variable given certain predictor variables. In this work we consider the problem of linear quantile regression in high dimensions where the number of predictor variables is much higher than the number of samples available for parameter estimation. We assume the true parameter to have some structure characterized as having a small value according to some atomic norm R(.) and consider the norm regularized quantile regression estimator. We characterize the sample complexity for consistent recovery and give non-asymptotic bounds on the estimation error. While this problem has been previously considered, our analysis reveals geometric and statistical characteristics of the problem not available in prior literature. We perform experiments on synthetic data which support the theoretical results.

We consider the problem of learning high-dimensional multi-response linear models with structured parameters. By exploiting the noise correlations among different responses, we propose an alternating estimation (AltEst) procedure to estimate the model parameters based on the generalized Dantzig selector (GDS). Under suitable sample size and resampling assumptions, we show that the error of the estimates generated by AltEst, with high probability, converges linearly to certain minimum achievable level, which can be tersely expressed by a few geometric measures, such as Gaussian width of sets related to the parameter structure. To the best of our knowledge, this is the first non-asymptotic statistical guarantee for such AltEst-type algorithm applied to estimation with general structures.

Hidden semi-Markov models (HSMMs) are latent variable models which allow latent state persistence and can be viewed as a generalization of the popular hidden Markov models (HMMs). In this paper, we introduce a novel spectral algorithm to perform inference in HSMMs. Unlike expectation maximization (EM), our approach correctly estimates the probability of given observation sequence based on a set of training sequences. Our approach is based on estimating moments from the sample, whose number of dimensions depends only logarithmically on the maximum length of the hidden state persistence. Moreover, the algorithm requires only a few matrix inversions and is therefore computationally efficient. Empirical evaluations on synthetic and real data demonstrate the advantage of the algorithm over EM in terms of speed and accuracy, especially for large data sets.

Data science models, although successful in a number of commercial domains, have had limited applicability in scientific problems involving complex physical phenomena. Theory-guided data science (TGDS) is an emerging paradigm that aims to leverage the wealth of scientific knowledge for improving the effectiveness of data science models in enabling scientific discovery. The overarching vision of TGDS is to introduce scientific consistency as an essential component for learning generalizable models. Further, by producing scientifically interpretable models, TGDS aims to advance our scientific understanding by discovering novel domain insights. Indeed, the paradigm of TGDS has started to gain prominence in a number of scientific disciplines such as turbulence modeling, material discovery, quantum chemistry, bio-medical science, bio-marker discovery, climate science, and hydrology. In this paper, we formally conceptualize the paradigm of TGDS and present a taxonomy of research themes in TGDS. We describe several approaches for integrating domain knowledge in different research themes using illustrative examples from different disciplines. We also highlight some of the promising avenues of novel research for realizing the full potential of theory-guided data science.

Statistical modeling of local precipitation involves understanding local, regional and global factors informative of precipitation variability in a region. Modern machine learning methods for feature selection can potentially be explored for identifying statistically significant features from pool of potential predictors of precipitation. In this work, we consider sparse regression, which simultaneously performs feature selection and regression, followed by random permutation tests for selecting dominant factors. We consider average winter precipitation over Great Lakes Region in order to identify its dominant influencing factors.Experiments show that global climate indices, computed at different temporal lags, offer predictive information for winter precipitation. Further, among the dominant factors identified using randomized permutation tests, multiple climate indices indicate the influence of geopotential height patterns on winter precipitation.Using composite analysis, we illustrate that certain patterns are indeed typical in high and low precipitation years, and offer plausible scientific reasons for variations in precipitation.Thus, feature selection methods can be useful in identifying influential climate processes and variables, and thereby provide useful hypotheses over physical mechanisms affecting local precipitation.

We consider the problem of estimating change in the dependency structure of two p-dimensional Ising models, based on respectively n_1 and n_2 samples drawn from the models. The change is assumed to be structured, e.g., sparse, block sparse, node-perturbed sparse, etc., such that it can be characterized by a suitable (atomic) norm. We present and analyze a norm-regularized estimator for directly estimating the change in structure, without having to estimate the structures of the individual Ising models. The estimator can work with any norm, and can be generalized to other graphical models under mild assumptions. We show that only one set of samples, say n_2, needs to satisfy the sample complexity requirement for the estimator to work, and the estimation error decreases as \fracc\sqrt\min(n_1,n_2), where c depends on the Gaussian width of the unit norm ball. For example, for \ell_1 norm applied to s-sparse change, the change can be accurately estimated with \min(n_1,n_2)=O(s \log p) which is sharper than an existing result n_1= O(s^2 \log p) and n_2 = O(n_1^2). Experimental results illustrating the effectiveness of the proposed estimator are presented.

While considerable advances have been made in estimating high-dimensional structured models from independent data using Lasso-type models, limited progress has been made for settings when the samples are dependent. We consider estimating structured VAR (vector auto-regressive model), where the structure can be captured by any suitable norm, e.g., Lasso, group Lasso, order weighted Lasso, etc. In VAR setting with correlated noise, although there is strong dependence over time and covariates, we establish bounds on the non-asymptotic estimation error of structured VAR parameters. The estimation error is of the same order as that of the corresponding Lasso-type estimator with independent samples, and the analysis holds for any norm. Our analysis relies on results in generic chaining, sub-exponential martingales, and spectral representation of VAR models. Experimental results on synthetic and real data with a variety of structures are presented, validating theoretical results.

In this work we consider the problem of anomaly detection in heterogeneous, multivariate, variable-length time series datasets. Our focus is on the aviation safety domain, where data objects are flights and time series are sensor readings and pilot switches. In this context the goal is to detect anomalous flight segments, due to mechanical, environmental, or human factors in order to identifying operationally significant events and highlight potential safety risks. For this purpose, we propose a framework which represents each flight using a semi-Markov switching vector autoregressive (SMS-VAR) model. Detection of anomalies is then based on measuring dissimilarities between the model's prediction and data observation. The framework is scalable, due to the inherent parallel nature of most computations, and can be used to perform online anomaly detection. Extensive experimental results on simulated and real datasets illustrate that the framework can detect various types of anomalies along with the key parameters involved.

In recent years, structured matrix recovery problems have gained considerable attention for its real world applications, such as recommender systems and computer vision. Much of the existing work has focused on matrices with low-rank structure, and limited progress has been made on matrices with other types of structure. In this paper we present non-asymptotic analysis for estimation of generally structured matrices via the generalized Dantzig selector based on sub-Gaussian measurements. We show that the estimation error can always be succinctly expressed in terms of a few geometric measures such as Gaussian widths of suitable sets associated with the structure of the underlying true matrix. Further, we derive general bounds on these geometric measures for structures characterized by unitarily invariant norms, a large family covering most matrix norms of practical interest. Examples are provided to illustrate the utility of our theoretical development.

High dimensional superposition models characterize observations using parameters which can be written as a sum of multiple component parameters, each with its own structure, e.g., sum of low rank and sparse matrices. In this paper, we consider general superposition models which allow sum of any number of component parameters, and each component structure can be characterized by any norm. We present a simple estimator for such models, give a geometric condition under which the components can be accurately estimated, characterize sample complexity of the estimator, and give non-asymptotic bounds on the componentwise estimation error. We use tools from empirical processes and generic chaining for the statistical analysis, and our results, which substantially generalize prior work on superposition models, are in terms of Gaussian widths of suitable spherical caps.

Structured estimation methods, such as LASSO, have received considerable attention in recent years and substantial progress has been made in extending such methods to general norms and non-Gaussian design matrices. In real world problems, however, covariates are usually corrupted with noise and there have been efforts to generalize structured estimation method for noisy covariate setting. In this paper we first show that without any information about the noise in covariates, currently established techniques of bounding statistical error of estimation fail to provide consistency guarantees. However, when information about noise covariance is available or can be estimated, then we prove consistency guarantees for any norm regularizer, which is a more general result than the state of the art. Next, we investigate empirical performance of structured estimation, specifically LASSO, when covariates are noisy and empirically show that LASSO is not consistent or stable in the presence of additive noise. However, prediction performance improves quite substantially when the noise covariance is available for incorporating in the estimator.

Multi-task learning (MTL) aims to improve generalization performance by learning multiple related tasks simultaneously. While sometimes the underlying task relationship structure is known, often the structure needs to be estimated from data at hand. In this paper, we present a novel family of models for MTL, applicable to regression and classification problems, capable of learning the structure of tasks relationship. In particular, we consider a joint estimation problem of the tasks relationship structure and the individual task parameters, which is solved using alternating minimization. The task relationship revealed by structure learning is founded on recent advances in Gaussian graphical models endowed with sparse estimators of the precision (inverse covariance) matrix. An extension to include flexible Gaussian copula models that relaxes the Gaussian marginal assumption is also proposed. We illustrate the effectiveness of the proposed model on a variety of synthetic and benchmark data sets for regression and classification. We also consider the problem of combining Earth System Model (ESM) outputs for better projections of future climate, with focus on projections of temperature by combining ESMs in South and North America, and show that the proposed model outperforms several existing methods for the problem.

In one-bit compressed sensing (1-bit CS), one attempts to estimate a structured parameter (signal) only using the sign of suitable linear measurements. In this paper, we investigate 1-bit CS problems for sparse signals using the recently proposed k-support norm. We show that the new estimator has a closed-form solution, so no optimization is needed. We establish consistency and recovery guarantees of the estimator for both Gaussian and sub-Gaussian random measurements. For Gaussian measurements, our estimator is comparable to the best known in the literature, along with guarantees on support recovery. For sub-Gaussian measurements, our estimator has an irreducible error which, unlike existing results, can be controlled by scaling the measurement vectors. In both cases, our analysis covers the setting of model misspecification, i.e., when the true sparsity is unknown. Experimental results illustrate several strengths of the new estimator.

Hidden semi-Markov models (HSMMs) are latent variable models which allow latent state persistence and can be viewed as a generalization of the popular hidden Markov models (HMMs). In this paper, we introduce a novel spectral algorithm to perform inference in HSMMs. Our approach is based on estimating certain sample moments, whose order depends only logarithmically on the maximum length of the hidden state persistence. Moreover, the algorithm requires only a few spectral decompositions and is therefore computationally efficient. Empirical evaluations on synthetic and real data demonstrate the promise of the algorithm.

An important problem in discrete graphical models is the maximum a posterior (MAP) inference problem. Recent research has been focusing on the development of parallel MAP inference algorithm, which scales to graphical models of millions of nodes. In this paper, we introduce a parallel implementation of the recently proposed Bethe-ADMM algorithm using Message Passing Interface (MPI), which allows us to fully utilize the computing power provided by the modern supercomputers with thousands of cores. Experimental results demonstrate that for a broad class of problems, our parallel implementation of Bethe-ADMM scales almost linearly even with thousands of cores.

The restricted eigenvalue (RE) condition characterizes the sample complexity of accurate recovery in the context of high-dimensional estimators such as Lasso and Dantzig selector (Bickel et al., 2009). Recent work has shown that random design matrices drawn from any thin-tailed (sub-Gaussian) distributions satisfy the RE condition with high probability, when the number of samples scale as the square of the Gaussian width of the restricted set (Banerjee et al., 2014; Tropp, 2015). We pose the equivalent question for heavy-tailed distributions: Given a random design matrix drawn from a heavy-tailed distribution satisfying the smallball property (Mendelson, 2015), does the design matrix satisfy the RE condition with the same order of sample complexity as sub-Gaussian distributions? An answer to the question will guide the design of highdimensional estimators for heavy tailed problems.

Data mining algorithms for computing solutions to online resource allocation (ORA) problems have focused on budgeting resources currently in possession, e.g., investing in the stock market with cash on hand or assigning current employees to projects. In several settings, one can leverage borrowed resources with which tasks can be accomplished more efficiently and cheaply. Additionally, a variety of opposing allocation types or positions may be available with which one can hedge the allocation to alleviate risk from external changes. In this paper, we present a formulation for hedging online resource allocations with leverage and propose an efficient data mining algorithm (SHERAL). We pose the problem as a constrained online convex optimization problem. The key novel components of our formulation are (1) a loss function for general leveraging and opposing allocation positions and (2) a penalty function which hedges between structurally dependent allocation positions to control risk. We instantiate the problem in the context of portfolio selection and evaluate the effectiveness of the formulation through extensive experiments on five datasets in comparison with existing algorithms and several variants.

Recent years have seen a revival of interest in the Alternating Direction Method of Multipliers (ADMM), due to its simplicity, versatility, and scalability. As a first order method for general convex problems, the rate of convergence of ADMM is O(1=k) [4, 25]. Given the scale of modern data mining problems, an algorithm with similar properties as ADMM but faster convergence rate can make a big difference in real world applications. In this paper, we introduce the Accelerated Alternating Direction Method of Multipliers (A2DM2) which solves problems with the same structure as ADMM. When the objective function is strongly convex, we show that A2DM2 has a O(1=k2) convergence rate. Unlike related existing literature on trying to accelerate ADMM, our analysis does not need any additional restricting assumptions. Through experiments, we show that A2DM2 converges faster than ADMM on a variety of problems. Further, we illustrate the versatility of the general A2DM2 on the problem of learning to rank, where it is shown to be competitive with the state-of-the-art specialized algorithms for the problem on both scalability and accuracy.

For structured estimation problems with atomic norms, recent advances in the literature express sample complexity and estimation error bounds in terms of certain geometric measures, in particular Gaussian width of the unit norm ball, Gaussian width of a spherical cap induced by a tangent cone, and a restricted norm compatibility constant. However, given an atomic norm, bounding these geometric measures can be difficult. In this paper, we present general upper bounds for such geometric measures, which only require simple information of the atomic norm under consideration, and we establish tightness of these bounds by providing the corresponding lower bounds. We show applications of our analysis to certain atomic norms, especially k-support norm, for which existing result is incomplete.

Matrix completion problems have been widely studied under special low dimensional structures such as low rank or structure induced by decomposable norms. In this paper, we present a unified analysis of matrix completion under general low-dimensional structural constraints induced by {\em any} norm regularization.We consider two estimators for the general problem of structured matrix completion, and provide unified upper bounds on the sample complexity and the estimation error. Our analysis relies on generic chaining, and we establish two intermediate results of independent interest: (a) in characterizing the size or complexity of low dimensional subsets in high dimensional ambient space, a certain \textit{\modified}~complexity measure encountered in the analysis of matrix completion problems is characterized in terms of a well understood complexity measure of Gaussian widths, and (b) it is shown that a form of restricted strong convexity holds for matrix completion problems under general norm regularization. Further, we provide several non-trivial examples of structures included in our framework, notably including the recently proposed spectral $k$-support norm.

We consider the problem of high-dimensional structured estimation with norm-regularized estimators, such as Lasso, when the design matrix and noise are drawn from sub-exponential distributions.Existing results only consider sub-Gaussian designs and noise, and both the sample complexity and non-asymptotic estimation error have been shown to depend on the Gaussian width of suitable sets. In contrast, for the sub-exponential setting, we show that the sample complexity and the estimation error will depend on the exponential width of the corresponding sets, and the analysis holds for any norm. Further, using generic chaining, we show that the exponential width for any set will be at most $\sqrt{\log p}$ times the Gaussian width of the set, yielding Gaussian width based results even for the sub-exponential case. Further, for certain popular estimators, viz Lasso and Group Lasso, using a VC-dimension based analysis, we show that the sample complexity will in fact be the same order as Gaussian designs. Our general analysis and results are the first in the sub-exponential setting, and are readily applicable to special sub-exponential families such as log-concave and extreme-value distributions.

A variety of modern data analysis problems, ranging from finance to job scheduling, can be considered as online resource allocation (ORA) problems. A key consideration in such ORA problems is some notion of risk, and suitable ways to alleviate risk. In several settings, the risk is structured so that groups of assets, such as stocks, are exposed to similar risks. In this paper, we present a formulation for online resource allocation with structured diversification (ORASD) as a constrained online convex optimization problem. The key novel component of our formulation is a constraint on the L(∞,1) group norm of the resource allocation vector, which ensures that no single group gets a large share of the resource and, unlike L(1,p) norms used for overlapping group Lasso, does not impose sparsity structures over groups. We instantiate the problem in the context of portfolio selection, propose an efficient ADMM algorithm, and illustrate the effectiveness of the formulation through extensive experiments on two benchmark datasets.

Link prediction is an important problem in online social and collaboration networks, for recommending friends and future collaborators. Most of the existing approaches for link prediction are focused on building unsupervised or supervised classification models based on the availability of accepts and rejects of the past recommendations. Several of these methods are feature-based and they construct a large number of network-level features to make the prediction more effective. A more flexible approach is to allow the model to learn the required features from the network for a specific task, rather than explicit feature engineering. In addition, most of the social and collaboration relationships do not happen instantly and rather build slowly over time through several low cost interactions, such as email and chat. The existing approaches often ignore the availability of such auxiliary networks to make link prediction more robust and effective. The main focus of this work is to build a robust and effective classifier for link prediction using multiple auxiliary networks. We develop a supervised random walk model, that does not require any explicit feature construction, and can be personalized to each user based on the past accept and reject behavior. Our approach consistently outperforms several popular baselines in terms of precision and recall in multiple real-life data sets. Also, our approach is robust to noise and sparsity in auxiliary networks, while several popular baselines, specifically feature-based ones, are inconsistent in their performance under such conditions.

The goal of collaborative filtering is to get accurate recommendations at the top of the list for a set of users. From such a perspective, collaborative ranking based formulations with suitable ranking loss functions are natural. While recent literature has explored the idea based on objective functions such as NDCG or Average Precision, such objectives are difficult to optimize directly. In this paper, building on recent advances from the learning to rank literature, we introduce a novel family of collaborative ranking algorithms which focus on accuracy at the top of the list for each user while learning the ranking functions collaboratively. We consider three specific formulations, based on collaborative p-norm push, infinite push, and reverse-height push, and propose efficient optimization methods for learning these models. Experimental results illustrate the value of collaborative ranking, and show that the proposed methods are competitive, usually better than existing popular approaches to personalized recommendation.

In portfolio selection, it often might be preferable to focus on a few top performing industries/sectors to beat the market. These top performing sectors however might change over time. In this paper, we propose an online portfolio selection algorithm that can take advantage of sector information through the use of a group sparsity inducing regularizer while making lazy updates to the portfolio. The lazy updates prevent changing ones portfolio too often which otherwise might incur huge transaction costs. The proposed formulation leads to a non-smooth constrained optimization problem at every step, with the constraint that the solution has to lie in a probability simplex. We propose an efficient primal-dual based alternating direction method of multipliers algorithm and demonstrate its effectiveness for the problem of online portfolio selection with sector information. We show that our algorithm OLU-GS has sub-linear regret w.r.t. the best fixed and best shifting solution in hindsight. We successfully establish the robustness and scalability of OLU-GS by performing extensive experiments on two real-world datasets.

We consider the problem of estimating sparse precision matrix of Gaussian copula distributions using samples with missing values in high dimensions. Existing approaches, primarily designed for Gaussian distributions, suggest using plugin estimators by disregarding the missing values. In this paper, we propose double plugin Gaussian (DoPinG) copula estimators to estimate the sparse precision matrix corresponding to \emphnon-paranormal distributions. DoPinG uses two plugin procedures and consists of three steps: (1) estimate nonparametric correlations based on observed values, including Kendall’s tau and Spearman’s rho; (2) estimate the non-paranormal correlation matrix; (3) plug into existing sparse precision estimators. We prove that DoPinG copula estimators consistently estimate the non-paranormal correlation matrix at a rate of O(\frac1(1-δ)\sqrt\frac\log pn), where δis the probability of missing values. We provide experimental results to illustrate the effect of sample size and percentage of missing data on the model performance. Experimental results show that DoPinG is significantly better than estimators like mGlasso, which are primarily designed for Gaussian data.

Multi-task learning (MTL) aims to improve generalization performance by learning multiple related tasks simultaneously. While sometimes the underlying task relationship structure is known, often the structure needs to be estimated from data at hand. In this paper, we present a novel family of models for MTL, applicable to regression and classification problems, capable of learning the structure of task relationships. In particular, we consider a joint estimation problem of the task relationship structure and the individual task parameters, which is solved using alternating minimization. The task relationship structure learning component builds on recent advances in structure learning of Gaussian graphical models based on sparse estimators of the precision (inverse covariance) matrix. We illustrate the effectiveness of the proposed model on a variety of synthetic and benchmark datasets for regression and classification. We also consider the problem of combining climate model outputs for better projections of future climate, with focus on temperature in South America, and show that the proposed model outperforms several existing methods for the problem.

Low-rank matrix completion methods have been successful in a variety of settings such as recommendation systems. However, most of the existing matrix completion methods only provide a point estimate of missing entries, and do not characterize uncertainties of the predictions. In this paper, we propose a Bayesian hierarchical probabilistic matrix factorization (BHPMF) model to 1) incorporate hierarchical side information, and 2) provide uncertainty quantified predictions. The former yields significant performance improvements in the problem of plant trait prediction, a key problem in ecology, by leveraging the taxonomic hierarchy in the plant kingdom. The latter is helpful in identifying predictions of low confidence which can in turn be used to guide field work for data collection efforts. A Gibbs sampler is designed for inference in the model. Further, we propose a multiple inheritance BHPMF (MI-BHPMF) which can work with a general directed acyclic graph (DAG) structured hierarchy, rather than a tree. We present comprehensive experimental results on the problem of plant trait prediction using the largest database of plant traits, where BHPMF shows strong empirical performance in uncertainty quantified trait prediction, outperforming the state-of-the-art based on point estimates. Further, we show that BHPMF is more accurate when it is confident, whereas the error is high when the uncertainty is high.

A major challenge for stochastic optimization is the cost of updating model parameters especially when the number of parameters is large. Updating parameters frequently can prove to be computationally or monetarily expensive. In this paper, we introduce an efficient primal-dual based online algorithm that performs lazy updates to the parameter vector and show that its performance is competitive with reasonable strategies which have the benefit of hindsight. We demonstrate the effectiveness of our algorithm in the online portfolio selection domain where a trader has to pay proportional transaction costs every time his portfolio is updated. Our Online Lazy Updates (OLU) algorithm takes into account the transaction costs while computing an optimal portfolio which results in sparse updates to the portfolio vector. We successfully establish the robustness and scalability of our lazy portfolio selection algorithm with extensive theoretical and experimental results on two real-world datasets.

We consider the problem of sparse precision matrix estimation in high dimensions using the CLIME estimator, which has several desirable theoretical properties. We present an inexact alternating direction method of multiplier (ADMM) algorithm for CLIME, and establish rates of convergence for both the objective and optimality conditions. Further, we develop a large scale distributed framework for the computations, which scales to millions of dimensions and trillions of parameters, using hundreds of cores. The proposed framework solves CLIME in column-blocks and only involves elementwise operations and parallel matrix multiplications. We evaluate our algorithm on both shared-memory and distributed-memory architectures, which can use block cyclic distribution of data and parameters to achieve load balance and improve the efficiency in the use of memory hierarchies. Experimental results show that our algorithm is substantially more scalable than state-of-the-art methods and scales almost linearly with the number of cores.

While Bregman divergences have been used for clustering and embedding problems in recent years, the facts that they are asymmetric and do not satisfy triangle inequality have been a major concern. In this paper, we investigate the relationship between two families of symmetrized Bregman divergences and metrics that satisfy the triangle inequality. The first family can be derived from any well-behaved convex function. The second family generalizes the Jensen-Shannon divergence, and can only be derived from convex functions with certain conditional positive definiteness structure. We interpret the required structure in terms of cumulants of infinitely divisible distributions, and related results in harmonic analysis. We investigate kmeans-type clustering problems using both families of symmetrized divergences, and give efficient algorithms for the same.

There are several Global Climate Models (GCM) reported by various countries to the Intergovernmental Panel on Climate Change (IPCC). Due to the varied nature of the GCM model assumptions, the future projections of the GCMs show high variability which makes it difficult to come up with confident projections into the future. Climate scientists combine these multiple GCM models to minimize the variability and the prediction error. Most of these model combinations are specifically for a location, or at a global scale. They do not consider regional or local smoothing (including the IPCC model). In this paper, we address this problem of combining multiple GCM model outputs with spatial smoothing as an important desired criterion. The problem formulation takes the form of multiple least squares regression for each geographic location with graph Laplacian based smoothing amongst the neighboring locations. Unlike the existing Laplacian regression frameworks, our formulation has both inner and outer products of the coefficient matrix, and has Sylvester equations as its special case. We discuss a few approaches to solve the problem, including a closed-form by solving a large linear system, as well as gradient descent methods which turn out to be more efficient. We establish the superiority of our approach in terms of model accuracy and smoothing compared to several popular baselines on real GCM climate datasets.

Covariance matrices have found success in several computer vision applications, including activity recognition, visual surveillance, and diffusion tensor imaging. This is because they provide an easy platform for fusing multiple features compactly. An important task in all of these applications is to compare two covariance matrices using a (dis)similarity function, for which the common choice is the Riemannian metric on the manifold inhabited by these matrices. As this Riemannian manifold is not flat, the dissimilarities should take into account the curvature of the manifold. As a result, such distance computations tend to slow down, especially when the matrix dimensions are large or gradients are required. Further, suitability of the metric to enable efficient nearest neighbor retrieval is an important requirement in the contemporary times of big data analytics. To alleviate these difficulties, this paper proposes a novel dissimilarity measure for covariances, the Jensen-Bregman LogDet Divergence (JBLD). This divergence enjoys several desirable theoretical properties and at the same time is computationally less demanding (compared to standard measures). Utilizing the fact that the square root of JBLD is a metric, we address the problem of efficient nearest neighbor retrieval on large covariance datasets via a metric tree data structure. To this end, we propose a K-Means clustering algorithm on JBLD. We demonstrate the superior performance of JBLD on covariance datasets from several computer vision applications.

Extracting sentiment from Twitter data is one of the fundamental problems in social media analytics. Twitter's length constraint renders determining the positive/negative sentiment of a tweet difficult, even for a human judge. In this work we present a general framework for per-tweet (in contrast with batches of tweets) sentiment analysis which consists of: (1) extracting tweets about a desired target subject, (2) separating tweets with sentiment, and (3) setting apart positive from negative tweets. For each step, we study the performance of a number of classical and new machine learning algorithms. We also show that the intrinsic sparsity of tweets allows performing classification in a low dimensional space, via random projections, without losing accuracy. In addition, we present weighted variants of all employed algorithms, exploiting the available labeling uncertainty, which further improve classification accuracy. Finally, we show that spatially aggregating our per-tweet classification results produces a very satisfactory outcome, making our approach a good candidate for batch tweet sentiment analysis.

In this paper, we consider the $\ell_1$ regularized sparse inverse covariance matrix estimation problem with a very large number of variables. Even in the face of this high dimensionality, and with limited number of samples, recent work has shown this estimator to have strong statistical guarantees in recovering the true structure of the sparse inverse covariance matrix, or alternatively the underlying graph structure of the corresponding Gaussian Markov Random Field. Our proposed algorithm divides the problem into smaller sub-problems, and uses the solutions of the sub-problems to build a good approximation for the original problem. We derive a bound on the distance of the approximate solution to the true solution. Based on this bound, we propose a clustering algorithm that attempts to minimize this bound, and in practice, is able to find effective partitions of the variables. We further use the approximate solution, i.e., solution resulting from solving the sub-problems, as an initial point to solve the original problem, and achieve a much faster computational procedure. As an example, a recent state-of-the-art method, QUIC requires 10 hours to solve a problem (with 10,000 nodes) that arises from a climate application, while our proposed algorithm, Divide and Conquer QUIC (DC-QUIC) only requires one hour to solve the problem.

Given their pervasive use, social media, such as Twitter, have become a leading source of breaking news. A key task in the automated identification of such news is the detection of novel documents from a voluminous stream of text documents in a scalable manner. Motivated by this challenge, we introduce the problem of online L1-dictionary learning where unlike traditional dictionary learning, which uses squared loss, the L1-penalty is used for measuring the reconstruction error. We present an efficient online algorithm for this problem based on alternating directions method of multipliers, and establish a sublinear regret bound for this algorithm. Empirical results on news-stream and Twitter data, shows that this online L1-dictionary learning algorithm for novel document detection gives more than an order of magnitude speedup over the previously known batch algorithm, without any significant loss in quality of results. Our algorithm for online L1-dictionary learning could be of independent interest.

The design of statistical predictive models for climate data gives rise to some unique challenges due to the high dimensionality and spatio-temporal nature of the datasets, which dictate that models should exhibit parsimony in variable selection. Recently, a class of methods which promote structured sparsity in the model have been developed, which is suitable for this task. In this paper, we prove theoretical statistical consistency of estimators with tree-structured norm regularizers. We consider one particular model, the Sparse Group Lasso (SGL), to construct predictors of land climate using ocean climate variables. Our experimental results demonstrate that the SGL model provides better predictive performance than the current state-of-the-art, remains climatologically interpretable, and is robust in its variable selection.

Droughts are one of the most damaging climate-related hazards. The late 1960s Sahel drought in Africa and the North American Dust Bowl of the 1930s are two examples of severe droughts that have an impact on society and the environment. Due to the importance of understanding droughts, we consider the problem of their detection based on gridded datasets of precipitation. We formulate the problem as the one of finding the most likely configuration of a Markov Random Field and propose an efficient inference algorithm. We apply this algorithm to the Climate Research Unit precipitation dataset spanning 106 years. The empirical results show that the algorithm successfully identifies the major droughts of the twentieth century in different regions of the world.

We propose a new matrix completion algorithm—Kernelized Probabilistic Matrix Factorization (KPMF), which effectively incorporates external side information into the matrix factorization process. Unlike Probabilistic Matrix Factorization (PMF) [14], which assumes an independent latent vector for each row (and each column) with Gaussian priors, KMPF works with latent vectors spanning all rows (and columns) with Gaussian Process (GP) priors. Hence, KPMF explicitly captures the underlying (nonlinear) covariance structures across rows and columns. This crucial difference greatly boosts the performance of KPMF when appropriate side information, e.g., users' social network in recommender systems, is incorporated. Furthermore, GP priors allow the KPMF model to fill in a row that is entirely missing in the original matrix based on the side information alone, which is not feasible for standard PMF formulation. In our paper, we mainly work on the matrix completion problem with a graph among the rows and/or columns as side information, but the proposed framework can be easily used with other types of side information as well. Finally, we demonstrate the efficacy of KPMF through two different applications: 1) recommender systems and 2) image restoration.

This survey attempts to provide a comprehensive and structured overview of the existing research for the problem of detecting anomalies in discrete/symbolic sequences. The objective is to provide a global understanding of the sequence anomaly detection problem and how existing techniques relate to each other. The key contribution of this survey is the classification of the existing research into three distinct categories, based on the problem formulation that they are trying to solve. These problem formulations are: 1) identifying anomalous sequences with respect to a database of normal sequences; 2) identifying an anomalous subsequence within a long sequence; and 3) identifying a pattern in a sequence whose frequency of occurrence is anomalous. We show how each of these problem formulations is characteristically distinct from each other and discuss their relevance in various application domains. We review techniques from many disparate and disconnected application domains that address each of these formulations. Within each problem formulation, we group techniques into categories based on the nature of the underlying algorithm. For each category, we provide a basic anomaly detection technique, and show how the existing techniques are variants of the basic technique. This approach shows how different techniques within a category are related or different from each other. Our categorization reveals new variants and combinations that have not been investigated before for anomaly detection. We also provide a discussion of relative strengths and weaknesses of different techniques. We show how techniques developed for one problem formulation can be adapted to solve a different formulation, thereby providing several novel adaptations to solve the different problem formulations. We also highlight the applicability of the techniques that handle discrete sequences to other related areas such as online anomaly detection and time series anomaly detection.

Streaming user-generated content in the form of blogs, microblogs, forums, and multimedia sharing sites, provides a rich source of data from which invaluable information and insights maybe gleaned. Given the vast volume of such social media data being continually generated, one of the challenges is to automatically tease apart the emerging topics of discussion from the constant background chatter. Such emerging topics can be identified by the appearance of multiple posts on a unique subject matter, which is distinct from previous online discourse. We address the problem of identifying emerging topics through the use of dictionary learning. We propose a two stage approach respectively based on detection and clustering of novel user-generated content. We derive a scalable approach by using the alternating directions method to solve the resulting optimization problems. Empirical results show that our proposed approach is more effective than several baselines in detecting emerging topics in traditional news story and newsgroup data. We also demonstrate the practical application to social media analysis, based on a study on streaming data from Twitter.

Covariance matrices provide compact, informative feature descriptors for use in several computer vision applications, such as people-appearance tracking, diffusion-tensor imaging, activity recognition, among others. A key task in many of these applications is to compare different covariance matrices using a (dis)similarity function. A natural choice here is the Riemannian metric corresponding to the manifold inhabited by covariance matrices. But computations involving this metric are expensive, especially for large matrices and even more so, in gradient-based algorithms. To alleviate these difficulties, we advocate a novel dissimilarity measure for covariance matrices: the Jensen-Bregman LogDet Divergence. This divergence enjoys several useful theoretical properties, but its greatest benefits are: (i) lower computational costs (compared to standard approaches); and (ii) amenability for use in nearest-neighbor retrieval. We show numerous experiments to substantiate these claims.

Several data mining algorithms use iterative optimization methods for learning predictive models. It is not easy to determine upfront which optimization method will perform best or converge fast for such tasks. In this paper, we analyze Meta Algorithms (MAs) which work by adaptively combining iterates from a pool of base optimization algorithms. We show that the performance of MAs are competitive with the best convex combination of the iterates from the base algorithms for online as well as batch convex optimization problems. We illustrate the effectiveness of MAs on the problem of portfolio selection in the stock market and use several existing ideas for portfolio selection as base algorithms. Using daily S\&P500 data for the past 21 years and a benchmark NYSE dataset, we show that MAs outperform existing portfolio selection algorithms with provable guarantees by several orders of magnitude, and match the performance of the best heuristics in the pool.

We consider the problem of finding a suitable common low dimensional subspace for accurately representing a given set of covariance matrices. With one covariance matrix, this is principal component analysis (PCA). For multiple covariance matrices, we term the problem Common Component Analysis (CCA). While CCA can be posed as a tensor decomposition problem, standard approaches to tensor decompositions have two critical issues: (i) tensor decomposition methods are iterative and rely on the initialization; (ii) for a given level of approximation error, it is difficult to choose a suitable low dimensionality. In this paper, we present a detailed analysis of CCA that yields an effective initialization and iterative algorithms for the problem. The proposed methodology has provable approximation guarantees w.r.t. the global maximum and also allows one to choose the dimensionality for a given level of approximation error. We also establish conditions under which the methodology will achieve the global maximum. We illustrate the effectiveness of the proposed method through extensive experiments on synthetic data as well as on two real stock market datasets, where major financial events can be visualized in low dimensions.

Co-clustering can be viewed as a two-way (bilinear) factorization of a large data matrix into dense/uniform and possibly overlapping sub-matrix factors (co-clusters). This combinatorially complex problem emerges in several applications, including behavior inference tasks encountered with social networks. Existing co-clustering schemes do not exploit the fact that overlapping factors are often sparse, meaning that their dimension is considerably smaller than that of the data matrix. Based on plaid models which allow for overlapping submatrices, the present paper develops a sparsity-cognizant overlapping co-clustering (SOC) approach. Numerical tests demonstrate the ability of the novel SOC scheme to globally detect multiple overlapping co-clusters, outperforming the original plaid model algorithms which rely on greedy search and ignore sparsity.

This paper presents a generalized version of the linear threshold model for simulating multiple cascades on a network while allowing nodes to switch between them. The proposed model is shown to be a rapidly mixing Markov chain and the corresponding steady state distribution is used to estimate highly likely states of the cascades' spread in the network. Results on a variety of real world networks demonstrate the high quality of the estimated solution.

Probabilistic matrix factorization (PMF) methods have shown great promise in collaborative filtering. In this paper, we consider several variants and generalizations of PMF framework inspired by three broad questions: Are the prior distributions used in existing PMF models suitable, or can one get better predictive performance with different priors? Are there suitable extensions to leverage side information? Are there benefits to taking into account row and column biases? We develop new families of PMF models to address these questions along with efficient approximate inference algorithms for learning and prediction. Through extensive experiments on movie recommendation datasets, we illustrate that simpler models directly capturing correlations among latent factors can outperform existing PMF models, side information can benefit prediction accuracy, and accounting for row/column biases leads to improvements in predictive performance.

In recent years, matrix approximation for missing value prediction has emerged as an important problem in a variety of domains such as recommendation systems, e-commerce and online advertisement. While matrix factorization based algorithms typically have good approximation accuracy, such algorithms can be slow especially for large matrices. Further, such algorithms cannot naturally make prediction on new rows or columns. In this paper, we propose residual Bayesian co-clustering (RBC), which learns a generative model corresponding to the matrix from the non-missing entries, and uses the model to predict the missing entries. RBC is an extension of Bayesian co-clustering by taking row and column bias into consideration. The model allows mixed memberships of rows and columns to multiple clusters, and can naturally handle the prediction on new rows and columns which are not used in the training process, given only a few non-missing entries in them. We propose two variational inference based algorithms for learning the model and predicting missing entries. One of the proposed algorithms leads to a parallel RBC which can achieve significant speed-ups. The efficacy of RBC is demonstrated by extensive experimental comparisons with state-of-the-art algorithms on real world datasets.

Topic detection (TD) is a fundamental research issue in the Topic Detection and Tracking (TDT) community with practical implications; TD helps analysts to separate the wheat from the chaff among the thousands of incoming news streams. In this paper, we propose a simple and effective topic detection model called the temporal Discriminative Probabilistic Model (DPM), which is shown to be theoretically equivalent to the classic vector space model with feature selection and temporally discriminative weights. We compare DPM to its various probabilistic cousins, ranging from mixture models like von-Mises Fisher (vMF) to mixed membership models like Latent Dirichlet Allocation (LDA). Benchmark results on the TDT3 data set show that sophisticated models, such as vMF and LDA, do not necessarily lead to better results; in the case of LDA, notably worst performance was obtained under variational inference, which is likely due to the significantly large number of LDA model parameters involved for document-level topic detection. On the contrary, using a relatively simple time-aware probabilistic model such as DPM suffices for both offline and online topic detection tasks, making DPM a theoretically elegant and effective model for practical topic detection.

Given a data matrix, the problem of finding dense/uniform sub-blocks in the matrix is becoming important in several applications. The problem is inherently combinatorial since the uniform sub-blocks may involve arbitrary subsets of rows and columns and may even be overlapping. While there are a few existing methods based on co-clustering or subspace clustering, they typically rely on local search heuristics and in general do not have a systematic model for such data. We present a Bayesian Overlapping Subspace Clustering (BOSC) model which is a hierarchical generative model for matrices with potentially overlapping uniform sub-block structures. The BOSC model can also handle matrices with missing entries. We propose an EM-style algorithm based on approximate inference using Gibbs sampling and parameter estimation using coordinate descent for the BOSC model. Through experiments on both simulated and real datasets, we demonstrate that the proposed algorithm outperforms the state-of-the-art.

Although mixed-membership models have achieved great success in unsupervised learning, they have not been widely applied to classification problems. In this paper, we propose a family of discriminative mixed-membership models for classification by combining unsupervised mixed-membership models with multi-class logistic regression. In particular, we propose two variants respectively applicable to text classification based on latent Dirichlet allocation and usual feature vector classification based on mixed-membership naive Bayes models. The proposed models allow the number of components in the mixed membership to be different from the number of classes. We propose two variational inference based algorithms for learning the models, including a fast variational inference which is substantially more efficient than mean-field variational approximation. Through extensive experiments on UCI and text classification benchmark datasets, we show that the models are competitive with the state of the art, and can discover components not explicitly captured by the class labels.

Cluster ensembles provide a framework for combining multiple base clusterings of a dataset to generate a stable and robust consensus clustering. There are important variants of the basic cluster ensemble problem, notably including cluster ensembles with missing values, as well as row-distributed or column-distributed cluster ensembles. Existing cluster ensemble algorithms are applicable only to a small subset of these variants. In this paper, we propose Bayesian Cluster Ensembles (BCE), which is a mixed-membership model for learning cluster ensembles, and is applicable to all the primary variants of the problem. We propose two methods, respectively based on variational approximation and Gibbs sampling, for learning a Bayesian cluster ensemble. We compare BCE extensively with several other cluster ensemble algorithms, and demonstrate that BCE is not only versatile in terms of its applicability, but also outperforms the other algorithms in terms of stability and accuracy.

The formation of secure transportation corridors, where cargoes and shipments from points of entry can be dispatched safely to highly sensitive and secure locations, is a high national priority. One of the key tasks of the program is the detection of

The problem of overlapping clustering, where a point is allowed to belong to multiple clusters, is becoming increasingly important in a variety of applications. In this paper, we present an overlapping clustering algorithm based on multiplicative mixture models. We analyze a general setting where each component of the multiplicative mixture is from an exponential family, and present an efficient alternating maximization algorithm to learn the model and infer overlapping clusters. We also show that when each component is assumed to be a Gaussian, we can apply the kernel trick leading to non-linear cluster separators and obtain better clustering quality. The efficacy of the proposed algorithms is demonstrated using experiments on both UCI benchmark datasets and a microarray gene expression dataset.

In recent years, co-clustering has emerged as a powerful data mining tool that can analyze dyadic data connecting two entities. However, almost all existing co-clustering techniques are partitional, and allow individual rows and columns of a data matrix to belong to only one cluster. Several current applications, such as recommendation systems and market basket analysis, can substantially benefit from a mixed membership of rows and columns. In this paper, we present Bayesian co-clustering (BCC) models, that allow a mixed membership in row and column clusters. BCC maintains separate Dirichlet priors for rows and columns over the mixed membership and assumes each observation to be generated by an exponential family distribution corresponding to its row and column clusters. We propose a fast variational algorithm for inference and parameter estimation. The model is designed to naturally handle sparse matrices as the inference is done only based on the non-missing entries. In addition to finding a co-cluster structure in observations, the model outputs a low dimensional co-embedding, and accurately predicts missing values in the original matrix. We demonstrate the efficacy of the model through experiments on both simulated and real data.

In spite of the popularity of probabilistic mixture models for latent structure discovery from data, mixture models do not have a natural mechanism for handling sparsity, where each data point only has a few non-zero observations. In this paper, we introduce conditional naive-Bayes (CNB) models, which generalize naive-Bayes mixture models to naturally handle sparsity by conditioning the model on observed features. Further, we present latent Dirichlet conditional naive-Bayes (LD-CNB) models, which constitute a family of powerful hierarchical Bayesian models for latent structure discovery from sparse data. The proposed family of models are quite general and can work with arbitrary regular exponential family conditional distributions. We present a variational inference based EM algorithm for learning along with special case analyses for Gaussian and discrete distributions. The efficacy of the proposed models are demonstrated by extensive experiments on a wide variety of different datasets.

Logistic models are arguably one of the most widely used data analysis techniques. In this paper, we present analyses focussing on two important aspects of logistic models—its relationship with exponential family based generative models, and its performance in online and potentially adversarial settings. In particular, we present two new theoretical results on logistic models focusing on the above two aspects. First, we establish an exact connection between logistic models and exponential family based generative models, resolving a long-standing ambiguity over their relationship. Second, we show that online Bayesian logistic models are competitive to the best batch models, even in potentially adversarial settings. Further, we discuss relevant connections of our analysis to the literature on integral transforms, and also present a new optimality result for Bayesian models. The analysis makes a strong case for using logistic models and partly explains the success of such models for a wide range of practical problems. 1

Topic modeling techniques have widespread use in text data mining applications. Some applications use batch models, which perform clustering on the document collection in aggregate. In this paper, we analyze and compare the performance of three recently-proposed batch topic models—Latent Dirichlet Allocation (LDA), Dirichlet Compound Multinomial (DCM) mixtures and von-Mises Fisher (vMF) mixture models. In cases where offline clustering on complete document collections is infeasible due to resource and response-rate constraints, online unsupervised clustering methods that process incoming data incrementally are necessary. To this end, we propose online variants of vMF, EDCM and LDA. Experiments on large real-world document collections, in both the offline and online settings, demonstrate that though LDA is a good model for finding word-level topics, vMF finds better document-level topic clusters more efficiently, which is often important in text mining applications. Finally, we propose a practical heuristic for hybrid topic modeling, which learns online topic models on streaming text and intermittently runs batch topic models on aggregated documents offline. Such a hybrid model is useful for several applications (e.g., dynamic topic-based aggregation of user-generated content in social networks) that need a good tradeoff between the performance of batch offline algorithms and efficiency of incremental online algorithms.

A number of real-world domains such as social networks and e-commerce involve heterogeneous data that describes relations between multiple classes of entities.Understanding the natural structure of this type of heterogeneous relational data is essential both for exploratory analysis and for performing various predictive modeling tasks.In this paper, we propose a principled multi-way clustering framework for relational data, wherein different types of entities are simultaneously clustered based not only on their intrinsic attribute values, but also on the multiple relations between the entities.To achieve this, we introduce a relation graph model that describes all the known relations between the different entity classes, in which each relation between a given set of entity classes is represented in the form of multi-modal tensor over an appropriate domain.Our multi-way clustering formulation is driven by the objective of capturing the maximal "information" in the original relation graph, i.e., accurately approximating the set of tensors corresponding to the various relations.This formulation is applicable to all Bregman divergences (a broad family of loss functions that includes squared Euclidean distance, KL-divergence), and also permits analysis of mixed data types using convex combinations of appropriate Bregman loss functions.Furthermore, we present a large family of structurally different multi-way clustering schemes that preserve various linear summary statistics of the original data.We accomplish the above generalizations by extending a recently proposed key theoretical result, namely the minimum Bregman information principle [1], to the relation graph setting.We also describe an efficient multi-way clustering algorithm based on alternate minimization that generalizes a number of other recently proposed clustering methods.Empirical results on datasets obtained from real-world domains (e.g., movie recommendations, newsgroup articles) demonstrate the generality and efficacy of our framework.

We show that several important Bayesian bounds studied in machine learning, both in the batch as well as the online setting, arise by an application of a simple compression lemma. In particular, we derive (i) PAC-Bayesian bounds in the batch setting, (ii) Bayesian log-loss bounds and (iii) Bayesian bounded-loss bounds in the online setting using the compression lemma. Although every setting has different semantics for prior, posterior and loss, we show that the core bound argument is the same. The paper simplifies our understanding of several important and apparently disparate results, as well as brings to light a powerful tool for developing similar arguments for other methods.

A perceptual image hash function maps an image to a short binary string based on an image's appearance to the human eye. Perceptual image hashing is useful in image databases, watermarking, and authentication. In this paper, we decouple image hashing into feature extraction (intermediate hash) followed by data clustering (final hash). For any perceptually significant feature extractor, we propose a polynomial-time heuristic clustering algorithm that automatically determines the final hash length needed to satisfy a specified distortion. We prove that the decision version of our clustering problem is NP complete. Based on the proposed algorithm, we develop two variations to facilitate perceptual robustness versus fragility tradeoffs. We validate the perceptual significance of our hash by testing under Stirmark attacks. Finally, we develop randomized clustering algorithms for the purposes of secure image hashing.

While the vast majority of clustering algorithms are partitional, many real world datasets have inherently overlapping clusters. Several approaches to finding overlapping clusters have come from work on analysis of biological datasets. In this paper, we interpret an overlapping clustering model proposed by Segal et al. [23] as a generalization of Gaussian mixture models, and we extend it to an overlapping clustering model based on mixtures of any regular exponential family distribution and the corresponding Bregman divergence. We provide the necessary algorithm modifications for this extension, and present results on synthetic data as well as subsets of 20-Newsgroups and EachMovie datasets.

Several large scale data mining applications, such as text categorization and gene expression analysis, involve high-dimensional data that is also inherently directional in nature. Often such data is L2 normalized so that it lies on the surface of a unit hypersphere. Popular models such as (mixtures of) multi-variate Gaussians are inadequate for characterizing such data. This paper proposes a generative mixture-model approach to clustering directional data based on the von Mises-Fisher (vMF) distribution, which arises naturally for data distributed on the unit hypersphere. In particular, we derive and analyze two variants of the Expectation Maximization (EM) framework for estimating the mean and concentration parameters of this mixture. Numerical estimation of the concentration parameters is non-trivial in high dimensions since it involves functional inversion of ratios of Bessel functions. We also formulate two clustering algorithms corresponding to the variants of EM that we derive. Our approach provides a theoretical basis for the use of cosine similarity that has been widely employed by the information retrieval community, and obtains the spherical kmeans algorithm (kmeans with cosine similarity) as a special case of both variants. Empirical results on clustering of high-dimensional text and gene-expression data based on a mixture of vMF distributions show that the ability to estimate the concentration parameter for each vMF component, which is not present in existing approaches, yields superior results, especially for difficult clustering tasks in high-dimensional spaces.

A wide variety of distortion functions, such as squared Euclidean distance, Mahalanobis distance, Itakura-Saito distance and relative entropy, have been used for clustering. In this paper, we propose and analyze parametric hard and soft clustering algorithms based on a large class of distortion functions known as Bregman divergences. The proposed algorithms unify centroid-based parametric clustering approaches, such as classical kmeans, the Linde-Buzo-Gray (LBG) algorithm and information-theoretic clustering, which arise by special choices of the Bregman divergence. The algorithms maintain the simplicity and scalability of the classical kmeans algorithm, while generalizing the method to a large class of clustering loss functions. This is achieved by first posing the hard clustering problem in terms of minimizing the loss in Bregman information, a quantity motivated by rate distortion theory, and then deriving an iterative algorithm that monotonically decreases this loss. In addition, we show that there is a bijection between regular exponential families and a large class of Bregman divergences, that we call regular Bregman divergences. This result enables the development of an alternative interpretation of an efficient EM scheme for learning mixtures of exponential family distributions, and leads to a simple soft clustering algorithm for regular Bregman divergences. Finally, we discuss the connection between rate distortion theory and Bregman clustering and present an information theoretic analysis of Bregman clustering algorithms in terms of a trade-off between compression and loss in Bregman information.

We consider the problem of predicting a random variable X from observations, denoted by a random variable Z. It is well known that the conditional expectation E[X|Z] is the optimal L/sup 2/ predictor (also known as "the least-mean-square error" predictor) of X, among all (Borel measurable) functions of Z. In this orrespondence, we provide necessary and sufficient conditions for the general loss functions under which the conditional expectation is the unique optimal predictor. We show that E[X|Z] is the optimal predictor for all Bregman loss functions (BLFs), of which the L/sup 2/ loss function is a special case. Moreover, under mild conditions, we show that the BLFs are exhaustive, i.e., if for every random variable X, the infimum of E[F(X,y)] over all constants y is attained by the expectation E[X], then F is a BLF.

An important task in unsupervised learning is maximum likelihood mixture estimation (MLME) for exponential families. In this paper, we prove a mathematical equivalence between this MLME problem and the rate distortion problem for Bregman divergences. We also present new theoretical results in rate distortion theory for Bregman divergences. Further, an analysis of the problems as a trade-off between compression and preservation of information is presented that yields the information bottleneck method as an interesting special case.

This paper provides necessary and sufficient conditions for general loss functions under which the conditional expectation is the unique optimal predictor of a random variable. Further, using such loss functions, we give an exact characterization of the difference between the two sides of Jensen's inequality.

Co-clustering is a powerful data mining technique with varied applications such as text clustering, microarray analysis and recommender systems. Recently, an information-theoretic co-clustering approach applicable to empirical joint probability distributions was proposed. In many situations, co-clustering of more general matrices is desired. In this paper, we present a substantially generalized co-clustering framework wherein any Bregman divergence can be used in the objective function, and various conditional expectation based constraints can be considered based on the statistics that need to be preserved. Analysis of the co-clustering problem leads to the minimum Bregman information principle, which generalizes the maximum entropy principle, and yields an elegant meta algorithm that is guaranteed to achieve local optimality. Our methodology yields new algorithms and also encompasses several previously known clustering and co-clustering algorithms based on alternate minimization.

We propose and test an objective criterion for evaluation of clustering performance: How well does a clustering algorithm run on unlabeled data aid a classification algorithm? The accuracy is quantified using the PAC-MDL bound [3] in a semisupervised setting. Clustering algorithms which naturally separate the data according to (hidden) labels with a small number of clusters perform well. A simple extension of the argument leads to an objective model selection method. Experimental results on text analysis datasets demonstrate that this approach empirically results in very competitive bounds on test set performance on natural datasets.

A wide variety of distortion functions are used for clustering, e.g., squared Euclidean distance, Mahalanobis distance and relative entropy. In this paper, we propose and analyze parametric hard and soft clustering algorithms based on a large class of distortion functions known as Bregman divergences. The proposed algorithms unify centroid-based parametric clustering approaches, such as classical kmeans and information-theoretic clustering, which arise by special choices of the Bregman divergence. The algorithms maintain the simplicity and scalability of the classical kmeans algorithm, while generalizing the basic idea to a very large class of clustering loss functions. There are two main contributions in this paper. First, we pose the hard clustering problem in terms of minimizing the loss in Bregman information, a quantity motivated by rate-distortion theory, and present an algorithm to minimize this loss. Secondly, we show an explicit bijection between Bregman divergences and exponential families. The bijection enables the development of an alternative interpretation of an efficient EM scheme for learning models involving mixtures of exponential distributions. This leads to a simple soft clustering algorithm for all Bregman divergences.

Semi-supervised clustering uses a small amount of supervised data to aid unsupervised learning. One typical approach specifies a limited number of must-link and cannot-link constraints between pairs of examples. This paper presents a pairwise constrained clustering framework and a new method for actively selecting informative pairwise constraints to get improved clustering performance. The clustering and active learning methods are both easily scalable to large datasets, and can handle very high dimensional data. Experimental and theoretical results confirm that this active querying of pairwise constraints significantly improves the accuracy of clustering when given a relatively small amount of supervision.

Competitive learning mechanisms for clustering, in general, suffer from poor performance for very high-dimensional (>1000) data because of "curse of dimensionality" effects. In applications such as document clustering, it is customary to normalize the high-dimensional input vectors to unit length, and it is sometimes also desirable to obtain balanced clusters, i.e., clusters of comparable sizes. The spherical kmeans (spkmeans) algorithm, which normalizes the cluster centers as well as the inputs, has been successfully used to cluster normalized text documents in 2000+ dimensional space. Unfortunately, like regular kmeans and its soft expectation-maximization-based version, spkmeans tends to generate extremely imbalanced clusters in high-dimensional spaces when the desired number of clusters is large (tens or more). This paper first shows that the spkmeans algorithm can be derived from a certain maximum likelihood formulation using a mixture of von Mises-Fisher distributions as the generative model, and in fact, it can be considered as a batch-mode version of (normalized) competitive learning. The proposed generative model is then adapted in a principled way to yield three frequency-sensitive competitive learning variants that are applicable to static data and produced high-quality and well-balanced clusters for high-dimensional data. Like kmeans, each iteration is linear in the number of data points and in the number of clusters for all the three algorithms. A frequency-sensitive algorithm to cluster streaming data is also proposed. Experimental results on clustering of high-dimensional text data sets are provided to show the effectiveness and applicability of the proposed techniques. Index Terms-Balanced clustering, expectation maximization (EM), frequency-sensitive competitive learning (FSCL), high-dimensional clustering, kmeans, normalized data, scalable clustering, streaming data, text clustering.

Automated clustering of text documents such as Web pages is becoming increasingly important for organizing the vast amounts of information available over the Internet. This problem is also very challenging since typically text is represented by very high dimensional (> 1000), normalized (unit length) vectors. Moreover documents are continually being created and their statistics also change with time because of changing new-stories etc, so one needs incremental learning algorithms that can adapt to non-stationary environments. We model high-dimensional, normalized data using a mixture of von Mises-Fisher distributions, and then modify this generative model in a principled way to yield frequency sensitive competitive learning mechanisms that are applicable to streaming data, and produce balanced clusters. Experimental results on clustering of high-dimensional text data sets are provided to show the effectiveness and applicability of the proposed techniques.

High dimensional directional data is becoming increasingly important in contemporary applications such as analysis of text and gene-expression data. A natural model for multi-variate directional data is provided by the von Mises-Fisher (vMF) distribution on the unit hypersphere that is analogous to the multi-variate Gaussian distribution in Rd. In this paper, we propose modeling complex directional data as a mixture of vMF distributions. We derive and analyze two variants of the Expectation Maximization (EM) framework for estimating the parameters of this mixture. We also propose two clustering algorithms corresponding to these variants. An interesting aspect of our methodology is that the spherical kmeans algorithm (kmeans with cosine similarity) can be shown to be a special case of both our algorithms. Thus, modeling text data by vMF distributions lends theoretical validity to the use of cosine similarity which has been widely used by the information retrieval community. As part of experimental validation, we present results on modeling high-dimensional text and gene-expression data as a mixture of vMF distributions. The results indicate that our approach yields superior clusterings especially for difficult clustering tasks in high-dimensional spaces.

1 Introduction The past few years have witnessed a growing interest in clustering algorithms that are suitable for data-mining problems [15, 14, 9]. Clustering algorithms for data-mining problems must be extremely scalable. In addition, several data mining applications demand that the clusters obtained be balanced, i.e., be of approximately the same size or importance. There are several notable approaches that address the scalability issue. Some approaches try to build the clusters dynamically by maintaining sufficient statistics and other summarized information in main memory while minimizing the number of database scans involved. For example, Bradley et al. [4, 5] propose out-of-core methods that scan the database once to form a summarized model (for instance, the size, sum and sum-squared values of potential cluster, and well as a small number of unallocated data-points) in main memory. Subsequent refinement based on this summarized information is then restricted to main memory operations without resorting to further disk scans. Another method with a similar flavor [24] compresses the data objects into many small subclusters using modified index trees and performs clustering with these subclusters. A different approach is to subsample the original data before applying the actual clustering algorithms [6, 12]. Ways of effectively sampling large datasets have also been proposed [20]. A recent work [8] suggests using less number of points in each step of an iterative relocation optimization algorithm like k-means as long as the model produced does not differ significantly from the one that would be obtained with full data.