As the size of modern datasets exceeds the disk and memory capacities of a single computer, machine learning practitioners have resorted to parallel and distributed computing. Given that optimization is one of the pillars of machine learning and predictive modeling, distributed optimization methods have recently garnered ample attention, in particular when either observations or features are distributed, but not both. We propose a general stochastic algorithm where observations, features, and gradient components can be sampled in a double distributed setting, i.e., with both features and observations distributed. Very technical analyses establish convergence properties of the algorithm under different conditions on the learning rate (diminishing to zero or constant). Computational experiments in Spark demonstrate a superior performance of our algorithm versus a benchmark in early iterations of the algorithm, which is due to the stochastic components of the algorithm.

We introduce a mini-batch stochastic variance-reduced algorithm to solve finite-sum scale invariant problems which cover several examples in machine learning and statistics such as principal component analysis (PCA) and estimation of mixture proportions. The algorithm is a stochastic generalization of scale invariant power iteration, specializing to power iteration when full-batch is used for the PCA problem. In convergence analysis, we show the expectation of the optimality gap decreases at a linear rate under some conditions on the step size, epoch length, batch size and initial iterate. Numerical experiments on the non-negative factorization problem with the KullbackLeibler divergence using real and synthetic datasets demonstrate that the proposed stochastic approach not only converges faster than state-of-the-art deterministic algorithms but also produces excellent quality robust solutions.

In recent years, multimodal learning has attracted an increasing interest. A special scenario of multimodal learning, learning on temporal data, is common but has not been well studied. In multimodal temporal data, not all modalities of a sample arrive at the same time. Because of that, different types of samples may have different importance in many use cases, where an early sample with significant modalities may be more valuable than a later one as early predictions can be made to speed up decision-making processes. Besides, sample correlations are very common in multimodal temporal data, as samples accumulate in time and a late sample may contain the same data existing in an earlier sample. Training without the awareness of the importance and correlation yields less effective models. In this work, we define multimodal temporal data, discuss key challenges and propose two methods that improve traditional multimodal training on such data. We demonstrate the effectiveness of the proposed methods on several multimodal temporal datasets, where they show 1% to 3% improvements over the baseline.

Automated feature engineering (AutoFE) is the process of automatically building and selecting new features that help improve downstream predictive performance. While traditional feature engineering requires significant domain expertise and time-consuming iterative testing, AutoFE strives to make feature engineering easy and accessible to all data science practitioners. We introduce a new AutoFE algorithm, IIFE, based on determining which feature pairs synergize well through an information-theoretic perspective called interaction information. We demonstrate the superior performance of IIFE over existing algorithms. We also show how interaction information can be used to improve existing AutoFE algorithms. Finally, we highlight several critical experimental setup issues in the existing AutoFE literature and their effects on performance.

Generative Adversarial Networks (GANs) have been used in many different applications to generate realistic synthetic data. We introduce a novel GAN with Autoencoder (GAN-AE) architecture to generate synthetic samples for variable length, multi-feature sequence datasets as existing GAN models cannot generate synthetic data and associated labels. In this model, we develop a GAN architecture with an additional autoencoder component, where recurrent neural networks (RNNs) are used for each component of the model in order to generate synthetic data to improve classification accuracy for a highly imbalanced medical device dataset. In addition to the medical device dataset, we also evaluate the GAN-AE performance on two additional datasets and demonstrate the application of GAN-AE to a sequence-to-sequence task where both synthetic sequence inputs and sequence outputs must be generated. To evaluate the quality of the synthetic data, we train encoder-decoder models both with and without the synthetic data and compare the classification model performance. We show that a model trained with GANAE generated synthetic data outperforms models trained with synthetic data generated both with standard oversampling techniques such as SMOTE and Autoencoders as well as with state of the art GAN-based models.

While many methods such as Locally Interpretable Model-agnostic Explanation (LIME), Integrated Gradients and Layerwise Relevance Propagation (LRP) have been developed to explain how recurrent neural networks make predictions, the explanations generated by each method often times vary dramatically. There is no consensus about which explainability method most accurately and robustly determine features important for model prediction. We consider a classification task on a sequence of events with different types and apply both gradient-based and attention-based explanation models to compute explanations on the event type level. We show that attention-based models return a higher similarity score between explanations for models initialized with different random seeds. However, there are still significant differences in explanations between model runs. We develop an optimization-based model to find a low-loss, high-accuracy path between two sets of trained weights to understand how model explanations morph between different local minima. We use this low-loss path to provide insight as to why explanations vary on two sentiment datasets.

Feature acquisition algorithms address the problem of acquiring informative features while balancing the costs of acquisition to improve the learning performances of ML models. Previous approaches have focused on calculating the expected utility values of features to determine the acquisition sequences. Other approaches formulated the problem as a Markov Decision Process (MDP) and applied reinforcement learning based algorithms. In comparison to previous approaches, we focus on 1) formulating the feature acquisition problem as a MDP and applying Monte Carlo Tree Search, 2) calculating the intermediary rewards for each acquisition step based on model predictions and acquisition costs and 3) simultaneously optimizing model prediction and acquisition costs with multi-objective Monte Carlo Tree Search. With Proximal Policy Optimization and Deep Q-Network algorithms as benchmark, we show the effectiveness of our proposed approach with experimental study.

This paper presents a novel semi-supervised deep learning algorithm for retrieving similar 2D and 3D videos based on visual content. The proposed approach combines the power of deep convolutional and recurrent neural networks with dynamic time warping as a similarity measure. The proposed algorithm is designed to handle large video datasets and retrieve the most related videos to a given inquiry video clip based on its graphical frames and contents. We split both the candidate and the inquiry videos into a sequence of clips and convert each clip to a representation vector using an autoencoder-backed deep neural network. We then calculate a similarity measure between the sequences of embedding vectors using a bi-directional dynamic time-warping method. This approach is tested on multiple public datasets, including CC_WEB_VIDEO, Youtube-8m, S3DIS, and Synthia, and showed good results compared to state-of-the-art. The algorithm effectively solves video retrieval tasks and outperforms the benchmarked state-of-the-art deep learning model.

In a recurrent setting, conventional approaches to neural architecture search find and fix a general model for all data samples and time steps. We propose a novel algorithm that can dynamically search for the structure of cells in a recurrent neural network model. Based on a combination of recurrent and recursive neural networks, our algorithm is able to construct customized cell structures for each data sample and time step, leading to an efficient sample-level architecture search model. Experiments on three common datasets show that the algorithm discovers high-performance cell architectures and achieves better performance compared to the GRU structure for language modelling with the same network capacity.

Common domain adaptation techniques assume that the source domain and the target domain share an identical label space, which is problematic since when target samples are unlabeled we have no knowledge on whether the two domains share the same label space. When this is not the case, the existing methods fail to perform well because the additional unknown classes are also matched with the source domain during adaptation. In this paper, we tackle the open set domain adaptation problem under the assumption that the source and the target label spaces only partially overlap, and the task becomes when the unknown classes exist, how to detect the target unknown classes and avoid aligning them with the source domain. We propose to model the tail of the entropy distribution with Extreme Value Theory for unknown class detection, which has shown promising performance on conventional domain adaptation datasets, and the proposed method outperforms the state-of-the-art models in domain adaptation.

Federated learning is a distributed machine learning paradigm where multiple data owners (clients) collaboratively train one machine learning model while keeping data on their own devices. The heterogeneity of client datasets is one of the most important challenges of federated learning algorithms. Studies have found performance reduction with standard federated algorithms, such as FedAvg, on non-IID data. Many existing works on handling non-IID data adopt the same aggregation framework as FedAvg and focus on improving model updates either on the server side or on clients. In this work, we tackle this challenge in a different view by introducing redistribution rounds that delay the aggregation. With delayed aggregations, local models are trained on data that are more representative to the global distribution. The proposed algorithm can also be used as a federated learning paradigm, as an alternative to FedAvg, where other methods can be plugged in. We perform experiments on multiple tasks and show that the proposed framework significantly improves the performance on non-IID data.

We propose incremental (re)training of a neural network model to cope with a continuous flow of new data. As such, this is a life-long learning process. We address two challenges of life-long retraining: catastrophic forgetting and efficient retraining. If we combine all past and new data it can easily become intractable to retrain the neural network model. On the other hand, if the model is retrained using only new data, it can easily suffer catastrophic forgetting and thus it is paramount to strike the right balance. Moreover, if we retrain all weights of the model every time new data is collected, retraining tends to require too many computing resources. To solve these two issues, we propose a novel retraining model that can select important samples and important weights utilizing multi-armed bandits. To further address forgetting, we propose a new regularization term focusing on synapse and neuron importance. We analyze multiple datasets to document the outcome of the proposed retraining methods. Various experiments demonstrate that our retraining methodologies mitigate the catastrophic forgetting problem while boosting model performance.

Certain type of documents such as tweets are collected by specifying a set of keywords. As topics of interest change with time it is beneficial to adjust keywords dynamically. The challenge is that these need to be specified ahead of knowing the forthcoming documents and the underlying topics. The future topics should mimic past topics of interest yet there should be some novelty in them. We develop a keyword-based topic model that dynamically selects a subset of keywords to be used to collect future documents. The generative process first selects keywords and then the underlying documents based on the specified keywords. The model is trained by using a variational lower bound and stochastic gradient optimization. The inference consists of finding a subset of keywords where given a subset the model predicts the underlying topic-word matrix for the unknown forthcoming documents. We compare the keyword topic model against a benchmark model using viral predictions of tweets combined with a topic model. The keyword-based topic model outperforms this sophisticated baseline model by 67%.

We introduce a new, reliable, and agnostic uncertainty measure for classification tasks called logit uncertainty. It is based on logit outputs of neural networks. We in particular show that this new uncertainty measure yields a superior performance compared to existing uncertainty measures on different tasks, including out of sample detection and finding erroneous predictions. We analyze theoretical foundations of the measure and explore a relationship with high density regions. We also demonstrate how to test uncertainty using intermediate outputs in training of generative adversarial networks. We propose two potential ways to utilize logit-based uncertainty in real world applications, and show that the uncertainty measure outperforms.

In model serving, having one fixed model during the entire often life-long inference process is usually detrimental to model performance, as data distribution evolves over time, resulting in lack of reliability of the model trained on historical data. It is important to detect changes and retrain the model in time. The existing methods generally have three weaknesses: 1) using only classification error rate as signal, 2) assuming ground truth labels are immediately available after features from samples are received and 3) unable to decide what data to use to retrain the model when change occurs. We address the first problem by utilizing six different signals to capture a wide range of characteristics of data, and we address the second problem by allowing lag of labels, where labels of corresponding features are received after a lag in time. For the third problem, our proposed method automatically decides what data to use to retrain based on the signals. Extensive experiments on structured and unstructured data for different type of data changes establish that our method consistently outperforms the state-of-the-art methods by a large margin.

Deep learning models based on CNNs are predominantly used in image classification tasks. Such approaches, assuming independence of object categories, normally use a CNN as a feature learner and apply a flat classifier on top of it. Object classes in many settings have known hierarchical relations, and classifiers exploiting these relations should perform better. We propose hierarchical classification models combining a CNN to extract hierarchical representations of images, and an RNN or sequence-to-sequence model to capture a hierarchical tree of classes. In addition, we apply residual learning to the RNN part in order to facilitate training our compound model and improve generalization of the model. Experimental results on a public and a real world proprietary dataset of images show that our hierarchical networks perform better than state-of-the-art CNNs.

In this paper, we propose new listwise learning-to-rank models that mitigate the shortcomings of existing ones. Existing listwise learning-to-rank models are generally derived from the classical Plackett-Luce model, which has three major limitations. (1) Its permutation probabilities overlook ties, i.e., a situation when more than one document has the same rating with respect to a query. This can lead to imprecise permutation probabilities and inefficient training because of selecting documents one by one. (2) It does not favor documents having high relevance. (3) It has a loose assumption that sampling documents at different steps is independent. To overcome the first two limitations, we model ranking as selecting documents from a candidate set based on unique rating levels in decreasing order. The number of steps in training is determined by the number of unique rating levels. More specifically, in each step, we apply multiple multi-class classification tasks to a document candidate set and choose all documents that have the highest rating from the document set. This is in contrast to taking one document step by step in the classical Plackett-Luce model. Afterward, we remove all of the selected documents from the document set and repeat until the remaining documents all have the lowest rating. We propose a new loss function and associated four models for the entire sequence of weighted classification tasks by assigning high weights to the selected documents with high ratings for optimizing Normalized Discounted Cumulative Gain (NDCG). To overcome the final limitation, we further propose a novel and efficient way of refining prediction scores by combining an adapted Vanilla Recurrent Neural Network (RNN) model with pooling given selected documents at previous steps. We encode all of the documents already selected by an RNN model. In a single step, we rank all of the documents with the same ratings using the last cell of the RNN multiple times. We have implemented our models using three settings: neural networks, neural networks with gradient boosting, and regression trees with gradient boosting. We have conducted experiments on four public datasets. The experiments demonstrate that the models notably outperform state-of-the-art learning-to-rank models.

Many activation functions have been proposed in the past, but selecting an adequate one requires trial and error. We propose a new methodology of designing activation functions within a neural network at each layer. We call this technique an “activation ensemble” because it allows the use of multiple activation functions at each layer. This is done by introducing additional variables, α, at each activation layer of a network to allow for multiple activation functions to be active at each neuron. By design, activations with larger α values at a neuron is equivalent to being “chosen” by the network. We implement the activation ensembles on a variety of datasets using an array of FFNs and CNNs. By using the activation ensemble, we achieve superior results compared to traditional techniques. In addition, because of the flexibility of this methodology, we more deeply explore activation functions and the features that they capture.

The problem of missing values in multivariable time series is a key challenge in many applications such as clinical data mining. Although many imputation methods show their effectiveness in many applications, few of them are designed to accommodate clinical multivariable time series. In this work, we propose a multiple imputation model that capture both cross-sectional information and temporal correlations. We integrate Gaussian processes with mixture models and introduce individualized mixing weights to handle the variance of predictive confidence of Gaussian process models. The proposed model is compared with several state-of-the-art imputation algorithms on both real-world and synthetic datasets. Experiments show that our best model can provide more accurate imputation than the benchmarks on all of our datasets.

As teams of professional leagues are becoming more and more analytically driven, the interest in effective data management and access of sports plays has dramatically increased. In this article, we present a retrieval system that can quickly find the most relevant plays from historical games given an input query. To search through a large number of games at an interactive speed, our system is built upon a distributed framework so that each query-result pair is evaluated in parallel. We also propose a pairwise learning to rank approach to improve search ranking based on users’ clickthrough behavior. The similarity metric in training the rank function is based on automatically learnt features from a convolutional autoencoder. Finally, we showcase the efficacy of our learning to rank approach by demonstrating rank quality in a user study.

Principal component analysis (PCA) is often used to reduce the dimension of data by selecting a few orthonormal vectors that explain most of the variance structure of the data. L1 PCA uses the L1 norm to measure error, whereas the conventional PCA uses the L2 norm. For the L1 PCA problem minimizing the fitting error of the reconstructed data, we propose an exact reweighted and an approximate algorithm based on iteratively reweighted least squares. We provide convergence analyses, and compare their performance against benchmark algorithms in the literature. The computational experiment shows that the proposed algorithms consistently perform best.

Massive radio frequency identification (RFID) data sets are expected to become commonplace in supply chain management systems. Warehousing and mining this data is an essential problem with great potential benefits for inventory management, object tracking, and product procurement processes. Since RFID tags can be used to identify each individual item, enormous amounts of location-tracking data are generated. With such data, object movements can be modeled by movement graphs, where nodes correspond to locations and edges record the history of item transitions between locations. In this study, we develop a movement graph model as a compact representation of RFID data sets. Since spatiotemporal as well as item information can be associated with the objects in such a model, the movement graph can be huge, complex, and multidimensional in nature. We show that such a graph can be better organized around gateway nodes, which serve as bridges connecting different regions of the movement graph. A graph-based object movement cube can be constructed by merging and collapsing nodes and edges according to an application-oriented topological structure. Moreover, we propose an efficient cubing algorithm that performs simultaneous aggregation of both spatiotemporal and item dimensions on a partitioned movement graph, guided by such a topological structure.

Radio Frequency Identification (RFID) applications are set to play an essential role in object tracking and supply chain management systems. In the near future, it is expected that every major retailer will use RFID systems to track the movement of products from suppliers to warehouses, store backrooms and eventually to points of sale. The volume of information generated by such systems can be enormous as each individual item (a pallet, a case, or an SKU) will leave a trail of data as it moves through different locations. As a departure from the traditional data cube, we propose a new warehousing model that preserves object transitions while providing significant compression and path-dependent aggregates, based on the following observations: (1) items usually move together in large groups through early stages in the system (e.g., distribution centers) and only in later stages (e.g., stores) do they move in smaller groups, and (2) although RFID data is registered at the primitive level, data analysis usually takes place at a higher abstraction level. Techniques for summarizing and indexing data, and methods for processing a variety of queries based on this framework are developed in this study. Our experiments demonstrate the utility and feasibility of our design, data structure, and algorithms.