AAAI 2021 Papers

Federated multi-armed bandits (FMAB) is a new bandit paradigm that parallels the federated learning (FL) framework in supervised learning. It is inspired by practical applications in cognitive radio and recommender systems, and enjoys features that are analogous to FL. This paper proposes a general framework of FMAB and then studies two specific federated bandit models. We first study the approximate model where the heterogeneous local models are random realizations of the global model from an unknown distribution. This model introduces a new uncertainty of client sampling, as the global model may not be reliably learned even if the finite local models are perfectly known. Furthermore, this uncertainty cannot be quantified a priori without knowledge of the suboptimality gap. We solve the approximate model by proposing Federated Double UCB (Fed2-UCB), which constructs a novel “double UCB” principle accounting for uncertainties from both arm and client sampling. We show that gradually admitting new clients is critical in achieving an O(log(T)) regret while explicitly considering the communication loss. The exact model, where the global bandit model is the exact average of heterogeneous local models, is then studied as a special case. We show that, somewhat surprisingly, the order-optimal regret can be achieved independent of the number of clients with a careful choice of the update periodicity. Experiments using both synthetic and real-world datasets corroborate the theoretical analysis and demonstrate the effectiveness and efficiency of the proposed algorithms.

In federated learning, models are learned from users’ data that are held private in their edge devices, by aggregating them in the service provider’s “cloud” to obtain a global model. Such global model is of great commercial value in, e.g., improving the customers’ experience. In this paper we focus on two possible areas of improvement of the state of the art. First, we take the difference between user habits into account and propose a quadratic penalty-based formulation, for efficient learning of the global model that allows to personalize local models. Second, we address the latency issue associated with the heterogeneous training time on edge devices, by exploiting a hierarchical structure modeling communication not only between the cloud and edge devices, but also within the cloud. Specifically, we devise a tailored block coordinate descent-based computation scheme, accompanied with communication protocols for both the synchronous and asynchronous cloud settings. We characterize the theoretical convergence rate of the algorithm, and provide a variant that performs empirically better. We also prove that the asynchronous protocol, inspired by multi-agent consensus technique, has the potential for large gains in latency compared to a synchronous setting when the edge-device updates are intermittent. Finally, experimental results are provided that corroborate not only the theory, but also show that the system leads to faster convergence for personalized models on the edge devices, compared to the state of the art.

With the marriage of federated machine learning and recommender systems for privacy-aware preference modeling and personalization, there comes a new research branch called federated recommender systems aiming to build a recommendation model in a distributed way, i.e., each user is represented as a distributed client where his/her original rating data are not shared with the server or the other clients. Notice that, besides the sensitive information of a specific rating score assigned to a certain item by a user, the information of a user's rated set of items shall also be well protected. Some very recent works propose to randomly sample some unrated items for each user and then assign some virtual ratings, so that the server can not identify the scores and the set of rated items easily during the server-client interactions. However, the virtual ratings assigned to the randomly sampled items will inevitably introduce some noise to the model training process, which will then cause loss in recommendation performance. In this paper, we propose a novel lossless federated recommendation method (FedRec++) by allocating some denoising clients (i.e., users) to eliminate the noise in a privacy-aware manner. We further analyse our FedRec++ in terms of security and losslessness, and discuss its generality in the context of existing works. Extensive empirical studies clearly show the effectiveness of our FedRec++ in providing accurate and privacy-aware recommendation without much additional communication cost.

In this paper we give a structural characterization and extend the tractability frontier of the Simple Temporal Problem (STP) by defining the class of the Extended Simple Temporal Problem (ESTP), which augments STP with strict inequalities and monotone Boolean formulae on inequations (i.e., formulae involving the operations of conjunction, disjunction and parenthesization). A polynomial-time algorithm is provided to solve ESTP, faster than previous state-of-the-art algorithms for other extensions of STP that had been considered in the literature, all encompassed by ESTP. We show the practical competitiveness of our approach through a proof-of-concept implementation and an experimental evaluation involving also state-of-the-art SMT solvers.

Stackelberg planning is a recent framework where a leader and a follower each choose a plan in the same planning task, the leader's objective being to maximize plan cost for the follower. This formulation naturally captures security-related (leader=defender, follower=attacker) as well as robustness-related (leader=adversarial event, follower=agent) scenarios. Solving Stackelberg planning tasks requires solving many related planning tasks at the follower level (in the worst case, one for every possible leader plan). Here we introduce new methods to tackle this source of complexity, through sharing information across follower tasks. Our evaluation shows that these methods can significantly reduce both the time needed to solve follower tasks and the number of follower tasks that need to be solved in the first place.

Blackwell approachability is a framework for reasoning about repeated games with vector-valued payoffs. We introduce predictive Blackwell approachability, where an estimate of the next payoff vector is given, and the decision maker tries to achieve better performance based on the accuracy of that estimator. In order to derive algorithms that achieve predictive Blackwell approachability, we start by showing a powerful connection between four well-known algorithms. Follow-the-regularized-leader (FTRL) and online mirror descent (OMD) are the most prevalent regret minimizers in online convex optimization. In spite of this prevalence, the regret matching (RM) and regret matching+ (RM+) algorithms have been preferred in the practice of solving large-scale games (as the local regret minimizers within the counterfactual regret minimization framework). We show that RM and RM+ are the algorithms that result from running FTRL and OMD, respectively, to select the halfspace to force at all times in the underlying Blackwell approachability game. By applying the predictive variants of FTRL or OMD to this connection, we obtain predictive Blackwell approachability algorithms, as well as predictive variants of RM and RM+. In experiments across 18 common zero-sum extensive-form benchmark games, we show that predictive RM+ coupled with counterfactual regret minimization converges vastly faster than the fastest prior algorithms (CFR+, DCFR, LCFR) across all games but two of the poker games, sometimes by two or more orders of magnitude.

Depth-adaptive neural networks can dynamically adjust depths according to the hardness of input words, and thus improve efficiency. The main challenge is how to measure such hardness and decide the required depths (i.e., layers) to conduct. Previous works generally build a halting unit to decide whether the computation should continue or stop at each layer. As there is no specific supervision of depth selection, the halting unit may be under-optimized and inaccurate, which results in suboptimal and unstable performance when modeling sentences. In this paper, we get rid of the halting unit and estimate the required depths in advance, which yields a faster depth-adaptive model. Specifically, two approaches are proposed to explicitly measure the hardness of input words and estimate corresponding adaptive depth, namely 1) mutual information (MI) based estimation and 2) reconstruction loss based estimation. We conduct experiments on the text classification task with 24 datasets in various sizes and domains. Results confirm that our approaches can speed up the vanilla Transformer (up to 7x) while preserving high accuracy. Moreover, efficiency and robustness are significantly improved when compared with other depth-adaptive approaches.

Adversarial attacks by generating examples which are almost indistinguishable from natural examples, pose a serious threat to learning models. Defending against adversarial attacks is a critical element for a reliable learning system. Support vector machine (SVM) is a classical yet still important learning algorithm even in the current deep learning era. Although a wide range of researches have been done in recent years to improve the adversarial robustness of learning models, but most of them are limited to deep neural networks (DNNs) and the work for kernel SVM is still vacant. In this paper, we aim at kernel SVM and propose adv-SVM to improve its adversarial robustness via adversarial training, which has been demonstrated to be the most promising defense techniques. To the best of our knowledge, this is the first work that devotes to the fast and scalable adversarial training of kernel SVM. Specifically, we first build connection of perturbations of samples between original and kernel spaces, and then give a reduced and equivalent formulation of adversarial training of kernel SVM based on the connection. Next, doubly stochastic gradients (DSG) based on two unbiased stochastic approximations (i.e., one is on training points and another is on random features) are applied to update the solution of our objective function. Finally, we prove that our algorithm optimized by DSG converges to the optimal solution at the rate of O(1/$t$) under the constant and diminishing stepsizes. Comprehensive experimental results show that our adversarial training algorithm enjoys robustness against various attacks and meanwhile has the similar efficiency and scalability with classical DSG algorithm.

Bilinear pooling has achieved an excellent performance in many computer vision tasks such as fine-grained classification, scene recognition and texture recognition. However, the high-dimension features from bilinear pooling can sometimes be inefficient and prone to over-fitting. Random Maclaurin (RM) is a widely used GPU-friendly approximation method to reduce the dimensionality of bilinear features. However, to achieve good performance, large projection matrices are usually required in practice, making it costly in computation and memory. In this paper, we propose a Shifted Random Maclaurin (SRM) strategy for fast and compact bilinear pooling. With merely negligible extra computational cost, the proposed SRM provides an estimator with a provably smaller variance than RM, which benefits accurate kernel approximation and thus the learning performance. Using a small projection matrix, the proposed SRM achieves a comparable estimation performance as RM based on a large projection matrix, and thus boosts the efficiency. Furthermore, we upgrade the proposed SRM to SRM+ to further improve the efficiency and make the compact bilinear pooling compatible with fast matrix normalization. Fast and Compact Bilinear Network (FCBN) built upon the proposed SRM+ is devised, achieving an end-to-end training. Systematic experiments conducted on four public datasets demonstrate the effectiveness and efficiency of the proposed FCBN.

Recent works have developed several methods of defending neural networks against adversarial attacks with certified guarantees. However, these techniques can be computationally costly due to the use of certification during training. We develop a new regularizer that is both more efficient than existing certified defenses, requiring only one additional forward propagation through a network, and can be used to train networks with similar certified accuracy. Through experiments on MNIST and CIFAR-10 we demonstrate improvements in training speed and comparable certified accuracy compared to state-of-the-art certified defenses.

We address the problem of performing Principal Component Analysis over a family of probability measures on the real line, using the Wasserstein geometry. We present a novel representation of the 2-Wasserstein space, based on a well known isometric bijection and a B-spline expansion. Thanks to this representation, we are able to reinterpret previous work and derive more efficient optimization routines for existing approaches. As shown in our simulations, the solution of these optimization problems can be costly in practice and thus pose a limit to their usage. We propose a novel definition of Principal Component Analysis in the Wasserstein space that, when used in combination with the B-spline representation, yields a straightforward optimization problem that is extremely fast to compute. Through extensive simulation studies, we show how our PCA performs similarly to the ones already proposed in the literature while retaining a much smaller computational cost. We apply our method to a real dataset of mortality rates due to Covid-19 in the US, concluding that our analyses are consistent with the current scientific consensus on the disease.

Generally, the existing graph-based multi-view clustering models consists of two steps: (1) graph construction; (2) eigen-decomposition on the graph Laplacian matrix to compute a continuous cluster assignment matrix, followed by a post-processing algorithm to get the discrete one. However, both the graph construction and eigen-decomposition are time-consuming, and the two-stage process may deviate from directly solving the primal problem. To this end, we propose Fast Multi-view Discrete Clustering (FMDC) with anchor graphs, focusing on directly solving the spectral clustering problem with a small time cost. We efficiently generate representative anchors and construct anchor graphs on different views. The discrete cluster assignment matrix is directly obtained by performing clustering on the automatically aggregated graph. FMDC has a linear computational complexity with respect to the data scale, which is a significant improvement compared to the quadratic one. Extensive experiments on benchmark datasets demonstrate its efficiency and effectiveness.

The use of language is subject to variation over time as well as across social groups and knowledge domains, leading to differences even in the monolingual scenario. Such variation in word usage is often called lexical semantic change (LSC). The goal of LSC is to characterize and quantify language variations with respect to word meaning, to measure how distinct two language sources are (that is, people or language models). Because there is hardly any data available for such a task, most solutions involve unsupervised methods to align two embeddings and predict semantic change with respect to a distance measure. To that end, we propose a self-supervised approach to model lexical semantic change based on the perturbation of word vectors in the input corpora. We show that our method can be used for the detection of semantic change with any alignment method. Furthermore, it can be used to choose the landmark words to use in alignment and can lead to substantial improvements over the existing techniques for alignment. We illustrate the utility of our techniques using experimental results on three different datasets, involving words with the same or different meanings. Our methods not only provide significant improvements but also can lead to novel findings for the LSC problem.

News recommendation is important for online news services. Existing news recommendation models are usually learned from users' news click behaviors. Usually the behaviors of users with the same sensitive attributes (e.g., genders) have similar patterns and news recommendation models can easily capture these patterns. It may lead to some biases related to sensitive user attributes in the recommendation results, e.g., always recommending sports news to male users, which is unfair since users may not receive diverse news information. In this paper, we propose a fairness-aware news recommendation approach with decomposed adversarial learning and orthogonality regularization, which can alleviate unfairness in news recommendation brought by the biases of sensitive user attributes. In our approach, we propose to decompose the user interest model into two components. One component aims to learn a bias-aware user embedding that captures the bias information on sensitive user attributes, and the other aims to learn a bias-free user embedding that only encodes attribute-independent user interest information for fairness-aware news recommendation. In addition, we propose to apply an attribute prediction task to the bias-aware user embedding to enhance its ability on bias modeling, and we apply adversarial learning to the bias-free user embedding to remove the bias information from it. Moreover, we propose an orthogonality regularization method to encourage the bias-free user embeddings to be orthogonal to the bias-aware one to better distinguish the bias-free user embedding from the bias-aware one. For fairness-aware news ranking, we only use the bias-free user embedding. Extensive experiments on benchmark dataset show that our approach can effectively improve fairness in news recommendation with minor performance loss.

Metric clustering is fundamental in areas ranging from Combinatorial Optimization and Data Mining, to Machine Learning and Operations Research. However, in a variety of situations we may have additional requirements or knowledge, distinct from the underlying metric, regarding which pairs of points should be clustered together. To capture and analyze such scenarios, we introduce a novel family of stochastic pairwise constraints, which we incorporate into several essential clustering objectives (radius/median/means). Moreover, we demonstrate that these constraints can succinctly model an intriguing collection of applications, including among others Individual Fairness in clustering and Must-link constraints in semi-supervised learning. Our main result consists of a general framework that yields approximation algorithms with provable guarantees for important clustering objectives, while at the same time producing solutions that respect the stochastic pairwise constraints. Furthermore, for certain objectives we devise improved results in the case of Must-link constraints, which are also the best possible from a theoretical perspective. Finally, we present experimental evidence that validates the effectiveness of our algorithms.

In machine learning, training data often capture the behaviour of multiple subgroups of some underlying human population. When the amounts of training data for the subgroups are not controlled carefully, under-representation bias arises. We introduce two natural notions of subgroup fairness and instantaneous fairness to address such under-representation bias in time-series forecasting problems. In particular, we consider the subgroup-fair and instant-fair learning of a linear dynamical system (LDS) from multiple trajectories of varying lengths and the associated forecasting problems. We provide globally convergent methods for the learning problems using hierarchies of convexifications of non-commutative polynomial optimisation problems. Our empirical results on a biased data set motivated by insurance applications and the well-known COMPAS data set demonstrate both the beneficial impact of fairness considerations on statistical performance and the encouraging effects of exploiting sparsity on run time.

We consider the problem of allocating a set on indivisible items to players with private preferences in an efficient and fair way. We focus on valuations that have dichotomous marginals, in which the added value of any item to a set is either 0 or 1, and aim to design truthful allocation mechanisms (without money) that maximize welfare and are fair. For the case that players have submodular valuations with dichotomous marginals, we design such a deterministic truthful allocation mechanism. The allocation output by our mechanism is Lorenz dominating, and consequently satisfies many desired fairness properties, such as being envy-free up to any item (EFX), and maximizing the Nash Social Welfare (NSW). We then show that our mechanism with random priorities is envy-free ex-ante, while having all the above properties ex-post. Furthermore, we present several impossibility results precluding similar results for the larger class of XOS valuations.

A set of divisible resources becomes available over a sequence of rounds and needs to be allocated immediately and irrevocably. Our goal is to distribute these resources to maximize fairness and efficiency. Achieving any non-trivial guarantees in an adversarial setting is impossible. However, we show that normalizing the agent values, a very common assumption in fair division, allows us to escape this impossibility. Our main result is an online algorithm for the case of two agents that ensures the outcome is fair while guaranteeing 91.6% of the optimal social welfare. We also show that this is near-optimal: there is no fair algorithm that guarantees more than 93.3% of the optimal social welfare.

We study the fair division problem of allocating multiple resources among a set of agents with Leontief preferences that are each required to complete a finite amount of work, which we term "limited demands". We examine the behavior of the classic Dominant Resource Fairness (DRF) mechanism in this setting and show it is fair but only weakly Pareto optimal and inefficient in many natural examples. We propose as an alternative the Least Cost Product (LCP) mechanism, a natural adaptation of Maximum Nash Welfare to this setting. We characterize the structure of allocation of the LCP mechanism in this setting, show that it is Pareto efficient, and that it satisfies the relatively weak fairness property of sharing incentives. While we prove that it satisfies the stronger fairness property of (expected) envy freeness in some special cases, we provide a counterexample showing it does not do so in general, a striking contrast to the "unreasonable fairness" of Maximum Nash Welfare in other settings. Simulations suggest, however, that these violations of envy freeness are rare in randomly generated examples.