The purpose of an online electronic-payment risk detection system is to prevent leakage, i.e., the loss of revenue that occurs when users fail to pay for services or when transactions are reversed.Nonpayment prediction models are trained on datasets comprising of features available when the model is triggered and the corresponding nonpayment labels.The latter are typically only observed several weeks or even months later.Furthermore, behavior indicative of future nonpayment is highly non-stationary, and the true model may drift significantly in the gap between trigger events and label collection.To address these challenges, we use post-transaction signals to generate pseudo-labels, i.e., short-term proxies [23] or surrogate-indices [33].Our framework attains a favorable tradeoff between ameliorating bias due to drift and introducing variance due to pseudo-label noise, as demonstrated by both offline and online experiments on several nonpayment-detection systems at Meta.Our deployment on live user traffic yields a statistically significant improvement in revenue, accounting also for leakage.

Survival analysis is widely deployed in a diverse set of fields, including healthcare, business, ecology, etc. The Cox Proportional Hazard (CoxPH) model is a semi-parametric model often encountered in the literature. Despite its popularity, wide deployment, and numerous variants, scaling CoxPH to large datasets and deep architectures poses a challenge, especially in the high-dimensional regime. We identify a fundamental connection between rank regression and the CoxPH model: this allows us to adapt and extend the so-called spectral method for rank regression to survival analysis. Our approach is versatile, naturally generalizing to several CoxPH variants, including deep models. We empirically verify our method's scalability on multiple real-world high-dimensional datasets; our method outperforms legacy methods w.r.t. predictive performance and efficiency.

We consider a recommender system that takes into account the interplay between recommendations, the evolution of user interests, and harmful content.We model the impact of recommendations on user behavior, particularly the tendency to consume harmful content.We seek recommendation policies that establish a tradeoff between maximizing click-through rate (CTR) and mitigating harm.We establish conditions under which the user profile dynamics have a stationary point, and propose algorithms for finding an optimal recommendation policy at stationarity.We experiment on a semi-synthetic movie recommendation setting initialized with real data and observe that our policies outperform baselines at simultaneously maximizing CTR and mitigating harm.

Recommender systems in social media websites provide value to their communities by recommending engaging content and meaningful connections. Scaling high-quality recommendations to billions of users in real-time requires sophisticated ranking models operating on a vast number of potential items to recommend, becoming prohibitively expensive computationally. A common technique "funnels'' these items through progressively complex models ("multi-stage''), each ranking fewer items but at higher computational cost for greater accuracy. This architecture introduces a trade-off between the cost of ranking items and providing users with the best recommendations. A key observation we make in this paper is that, all else equal, ranking more items indeed improves the overall objective but has diminishing returns. Following this observation, we provide a rigorous formulation through the framework of DR-submodularity, and argue that for a certain class of objectives (reward functions), it is possible to improve the trade-off between performance and computational cost in multi-stage ranking systems with strong theoretical guarantees. We show that this class of reward functions that provide this guarantee is large and robust to various noise models. Finally, we describe extensive experimentation of our method on three real-world recommender systems in Facebook, achieving 8.8% reduction in overall compute resources with no significant impact on recommendation quality, compared to a 0.8% quality loss in a non-personalized budget allocation.

It is challenging for generative models to learn a distribution over graphs because of the lack of permutation invariance: nodes may be ordered arbitrarily across graphs, and standard graph alignment is combinatorial and notoriously expensive. We propose AlignGraph, a group of generative models that combine fast and efficient graph alignment methods with a family of deep generative models that are invariant to node permutations. Our experiments demonstrate that our framework successfully learns graph distributions, outperforming competitors by 25% — 560% in relevant performance scores.

Social media platforms like Facebook and YouTube connect people with communities that reflect their own values and experiences. People discover new communities either organically or through algorithmic recommendations based on their interests and preferences. We study online journeys users take through these communities, focusing particularly on ones that may lead to problematic outcomes. In particular, we propose and explore the concept of gateways, namely, entities associated with a higher likelihood of subsequent engagement with problematic content. We show, via a real-world application on Facebook groups, that a simple definition of gateway entities can be leveraged to reduce exposure to problematic content by 1% without any adverse impact on user engagement metrics. Motivated by this finding, we propose several formal definitions of gateways, via both frequentist and survival analysis methods, and evaluate their efficacy in predicting user behavior through offline experiments. Frequentist, duration-insensitive methods predict future harmful engagements with an 0.64–0.83 AUC, while survival analysis methods improve this to 0.72–0.90 AUC.

Adversarial pruning compresses models while preserving robustness. Current methods require access to adversarial examples during pruning. This significantly hampers training efficiency. Moreover, as new adversarial attacks and training methods develop at a rapid rate, adversarial pruning methods need to be modified accordingly to keep up. In this work, we propose a novel framework to prune a previously trained robust neural network while maintaining adversarial robustness, without further generating adversarial examples. We leverage concurrent self-distillation and pruning to preserve knowledge in the original model as well as regularizing the pruned model via the Hilbert-Schmidt Information Bottleneck. We comprehensively evaluate our proposed framework and show its superior performance in terms of both adversarial robustness and efficiency when pruning architectures trained on the MNIST, CIFAR-10, and CIFAR-100 datasets against five state-of-the-art attacks..

We study the problem of ranking regression, in which a dataset of rankings is used to learn Plackett–Luce scores as functions of sample features. We propose a novel spectral algorithm to accelerate learning in ranking regression. Our main technical contribution is to show that the Plackett–Luce negative log-likelihood augmented with a proximal penalty has stationary points that satisfy the balance equations of a Markov Chain. This allows us to tackle the ranking regression problem via an efficient spectral algorithm by using the Alternating Directions Method of Multipliers (ADMM). ADMM separates the learning of scores and model parameters, and in turn, enables us to devise fast spectral algorithms for ranking regression via both shallow and deep neural network (DNN) models. For shallow models, our algorithms are up to 579 times faster than the Newton’s method. For DNN models, we extend the standard ADMM via a Kullback–Leibler proximal penalty and show that this is still amenable to fast inference via a spectral approach. Compared to a state-of-the-art siamese network, our resulting algorithms are up to 175 times faster and attain better predictions by up to 26% Top-1 Accuracy and 6% Kendall-Tau correlation over five real-life ranking datasets.

Graph embeddings have been tremendously successful at producing node representations that are discriminative for downstream tasks. In this paper, we study the problem of graph transfer learning: given two graphs and labels in the nodes of the first graph, we wish to predict the labels on the second graph. We propose a tractable, non-combinatorial method for solving the graph transfer learning problem by combining classification and embedding losses with a continuous, convex penalty motivated by tractable graph distances. We demonstrate that our method successfully predicts labels across graphs with almost perfect accuracy; in the same scenarios, training embeddings through standard methods leads to predictions that are no better than random.

We study submodular maximization problems with matroid constraints, in particular, problems where the objective can be expressed via compositions of analytic and multilinear functions. We show that for functions of this form, the so-called continuous greedy algorithm attains a ratio arbitrarily close to $(1-1/e) \approx 0.63$ using a deterministic estimation via Taylor series approximation. This drastically reduces execution time over prior art that uses sampling.

In lifelong learning, we wish to maintain and update a model (e.g., a neural network classifier) in the presence of new classification tasks that arrive sequentially. In this paper, we propose a learn-prune-share (LPS) algorithm which addresses the challenges of catastrophic forgetting, parsimony, and knowledge reuse simultaneously. LPS splits the network into task-specific partitions via an ADMM-based pruning strategy. This leads to no forgetting, while maintaining parsimony. Moreover, LPS integrates a novel selective knowledge sharing scheme into this ADMM optimization framework. This enables adaptive knowledge sharing in an end-to-end fashion. Comprehensive experimental results on two lifelong learning benchmark datasets and a challenging real world radio frequency fingerprinting dataset are provided to demonstrate the effectiveness of our approach. Our experiments show that LPS consistently outperforms multiple state-of-the-art competitors.

We study an Open-World Class Discovery problem in which, given labeled training samples from old classes, we need to discover new classes from unlabeled test samples. There are two critical challenges to addressing this paradigm: (a) transferring knowledge from old to new classes, and (b) incorporating knowledge learned from new classes back to the original model. We propose Class Discovery Kernel Network with Expansion (CD-KNet-Exp), a deep learning framework, which utilizes the Hilbert Schmidt Independence Criterion to bridge supervised and unsupervised information together in a systematic way, such that the learned knowledge from old classes is distilled appropriately for discovering new classes. Compared to competing methods, CD-KNet-Exp shows superior performance on three publicly available benchmark datasets and a challenging real-world radio frequency fingerprinting dataset.

Recommender systems should adapt to user interests as the latter evolve. A prevalent cause for the evolution of user interests is the influence of their social circle. In general, when the interests are not known, online algorithms that explore the recommendation space while also exploiting observed preferences are preferable. We present online recommendation algorithms rooted in the linear multi-armed bandit literature. Our bandit algorithms are tailored precisely to recommendation scenarios where user interests evolve under social influence. In particular, we show that our adaptations of the classic LinREL and ThompsonSampling algorithms maintain the same asymptotic regret bounds as in the non-social case. We validate our approach experimentally using both synthetic and real datasets.

We propose a deep learning approach for discovering kernels tailored to identifying clusters over sample data. Our neural network produces sample embeddings that are motivated by and are at least as expressive as spectral clustering. Our training objective, based on the Hilbert Schmidt Independence Criterion, can be optimized via gradient adaptations on the Stiefel manifold, leading to significant acceleration over spectral methods relying on eigen-decompositions. Finally, our trained embedding can be directly applied to out-of-sample data. We show experimentally that our approach outperforms several state-of-the-art deep clustering methods, as well as traditional approaches such as k-means and spectral clustering over a broad array of real and synthetic datasets.

Retinopathy of Prematurity (ROP) is a leading cause for childhood blindness worldwide. An automated ROP detection system could significantly improve the chance of a child receiving proper diagnosis and treatment. We propose a means of producing a continuous severity score in an automated fashion, regressed from both (a) diagnostic class labels as well as (b) comparison outcomes. Our generative model combines the two sources, and successfully addresses inherent variability in diagnostic outcomes. In particular, our method exhibits an excellent predictive performance of both diagnostic and comparison outcomes over a broad array of metrics, including AUC, precision, and recall.

Pairwise comparison labels are more informative and less variable than class labels, but generating them poses a challenge: their number grows quadratically in the dataset size. We study a natural experimental design objective, namely, D-optimality, that can be used to identify which K pairwise comparisons to generate. This objective is known to perform well in practice, and is submodular, making the selection approximable via the greedy algorithm. A naïve greedy implementation has O(N2 d2 K) complexity, where N is the dataset size, d is the feature space dimension, and K is the number of generated comparisons. We show that, by exploiting the inherent geometry of the dataset–namely, that it consists of pairwise comparisons–the greedy algorithm's complexity can be reduced to O(N2 (K + d) + N(dK + d2) + d2 K). We apply the same acceleration also to the so-called lazy greedy algorithm. When combined, the above improvements lead to an execution time of less than 1 hour for a dataset with 108 comparisons; the naïve greedy algorithm on the same dataset would require more than 10 days to terminate.

Important data mining problems such as nearest-neighbor search and clustering admit theoretical guarantees when restricted to objects embedded in a metric space. Graphs are ubiquitous, and clustering and classification over graphs arise in diverse areas, including, e.g., image processing and social networks. Unfortunately, popular distance scores used in these applications, that scale over large graphs, are not metrics and thus come with no guarantees. Classic graph distances such as, e.g., the chemical and the CKS distance are arguably natural and intuitive, and are indeed also metrics, but they are intractable: as such, their computation does not scale to large graphs. We define a broad family of graph distances, that includes both the chemical and the CKS distance, and prove that these are all metrics. Crucially, we show that our family includes metrics that are tractable. Moreover, we extend these distances by incorporating auxiliary node attributes, which is important in practice, while maintaining both the metric property and tractability.

Large-scale optimization problems abound in data mining and machine learning applications, and the computational challenges they pose are often addressed through parallelization. We identify structural properties under which a convex optimization problem can be massively parallelized via map-reduce operations using the Frank-Wolfe (FW) algorithm. The class of problems that can be tackled this way is quite broad and includes experimental design, AdaBoost, and projection to a convex hull. Implementing FW via map-reduce eases parallelization and deployment via commercial distributed computing frameworks. We demonstrate this by implementing FW over Spark, an engine for parallel data processing, and establish that parallelization through map-reduce yields significant performance improvements: we solve problems with 10 million variables using 350 cores in 44 minutes; the same operation takes 133 hours when executed serially.

How can we design a product or movie that will attract, for example, the interest of Pennsylvania adolescents or liberal newspaper critics? What should be the genre of that movie and who should be in the cast? In this work, we seek to identify how we can design new movies with features tailored to a specific user population. We formulate the movie design as an optimization problem over the inference of user-feature scores and selection of the features that maximize the number of attracted users. Our approach, PNP, is based on a heterogeneous, tripartite graph of users, movies, and features (e.g. actors, directors, genres), where users rate movies and features contribute to movies. We learn the preferences by leveraging user similarities defined through different types of relations, and show that our method outperforms state-of-the-art approaches, including matrix factorization and other heterogeneous graph-based analysis. We evaluate PNP on publicly available real-world data and show that it is highly scalable and effectively provides movie designs oriented towards different groups of users, including men, women, and adolescents.

People's interests are dynamically evolving, often affected by external factors such as trends promoted by the media or adopted by their friends. In this work, we model interest evolution through dynamic interest cascades: we consider a scenario where a user's interests may be affected by (a) the interests of other users in her social circle, as well as (b) suggestions she receives from a recommender system. In the latter case, we model user reactions through either attraction or aversion towards past suggestions. We study this interest evolution process, and the utility accrued by recommendations, as a function of the system's recommendation strategy. We show that, in steady state, the optimal strategy can be computed as the solution of a semi-definite program (SDP). Using datasets of user ratings, we provide evidence for the existence of aversion and attraction in real-life data, and show that our optimal strategy can lead to significantly improved recommendations over systems that ignore aversion and attraction.

Recommender systems leverage user demographic information, such as age, gender, etc., to personalize recommendations and better place their targeted ads. Oftentimes, users do not volunteer this information due to privacy concerns, or due to a lack of initiative in filling out their online profiles. We illustrate a new threat in which a recommender learns private attributes of users who do not voluntarily disclose them. We design both passive and active attacks that solicit ratings for strategically selected items, and could thus be used by a recommender system to pursue this hidden agenda. Our methods are based on a novel usage of Bayesian matrix factorization in an active learning setting. Evaluations on multiple datasets illustrate that such attacks are indeed feasible and use significantly fewer rated items than static inference methods. Importantly, they succeed without sacrificing the quality of recommendations to users.

User demographics, such as age, gender and ethnicity, are routinely used for targeting content and advertising products to users. Similarly, recommender systems utilize user demographics for personalizing recommendations and overcoming the cold-start problem. Often, privacy-concerned users do not provide these details in their online profiles. In this work, we show that a recommender system can infer the gender of a user with high accuracy, based solely on the ratings provided by users (without additional metadata), and a relatively small number of users who share their demographics. Focusing on gender, we design techniques for effectively adding ratings to a user's profile for obfuscating the user's gender, while having an insignificant effect on the recommendations provided to that user.

Recommender systems predict user preferences based on a range of available information. For systems in which users generate streams of content (e.g., blogs, periodically-updated newsfeeds), users may rate the produced content that they read, and be given accurate predictions about future content they are most likely to prefer. We design a distributed mechanism for predicting user ratings that avoids the disclosure of information to a centralized authority or an untrusted third party: users disclose the rating they give to certain content only to the user that produced this content.