We study the problem of clustering nodes in a dynamic graph, where the connections between nodes and nodes' cluster memberships may change over time, e.g., due to community migration. We first propose a dynamic stochastic block model that captures these changes, and a simple decay-based clustering algorithm that clusters nodes based on weighted connections between them, where the weight decreases at a fixed rate over time. This decay rate can then be interpreted as signifying the importance of including historical connection information in the clustering. However, the optimal decay rate may differ for clusters with different rates of turnover. We characterize the optimal decay rate for each cluster and propose a clustering method that achieves almost exact recovery of the true clusters. We then demonstrate the efficacy of our clustering algorithm with optimized decay rates on simulated graph data. Recurrent neural networks (RNNs), a popular algorithm for sequence learning, use a similar decay-based method, and we use this insight to propose two new RNN-GCN (graph convolutional network) architectures for semi-supervised graph clustering. We finally demonstrate that the proposed architectures perform well on real data compared to state-of-the-art graph clustering algorithms.
We study the problem of learning data representations that are private yet informative, i.e., providing information about intended "ally" targets while obfuscating sensitive "adversary" attributes. We propose a novel framework, Exclusion-Inclusion Generative Adversarial Network (EIGAN), that generalizes existing adversarial privacy-preserving representation learning (PPRL) approaches to generate data encodings that account for multiple possibly overlapping ally and adversary targets. Preserving privacy is even more difficult when the data is collected across multiple distributed nodes, which for privacy reasons may not wish to share their data even for PPRL training. Thus, learning such data representations at each node in a distributed manner (i.e., without transmitting source data) is of particular importance. This motivates us to develop D-EIGAN, the first distributed PPRL method, based on federated learning with fractional parameter sharing to account for communication resource limitations. We theoretically analyze the behavior of adversaries under the optimal EIGAN and D-EIGAN encoders and consider the impact of dependencies among ally and adversary tasks on the encoder performance. Our experiments on real-world and synthetic datasets demonstrate the advantages of EIGAN encodings in terms of accuracy, robustness, and scalability; in particular, we show that EIGAN outperforms the previous state-of-the-art by a significant accuracy margin (47% improvement). The experiments further reveal that D-EIGAN's performance is consistent with EIGAN under different node data distributions and is resilient to communication constraints.
While Deep Neural Networks (DNNs) are becoming the state-of-the-art for many tasks including reinforcement learning (RL), they are especially resistant to human scrutiny and understanding. Input attributions have been a foundational building block for DNN expalainabilty but face new challenges when applied to deep RL. We address the challenges with two novel techniques. We define a class of \emph{behaviour-level attributions} for explaining agent behaviour beyond input importance and interpret existing attribution methods on the behaviour level. We then introduce \emph{$\lambda$-alignment}, a metric for evaluating the performance of behaviour-level attributions methods in terms of whether they are indicative of the agent actions they are meant to explain. Our experiments on Atari games suggest that perturbation-based attribution methods are significantly more suitable to deep RL than alternatives from the perspective of this metric. We argue that our methods demonstrate the minimal set of considerations for adopting general DNN explanation technology to the unique aspects of reinforcement learning and hope the outlined direction can serve as a basis for future research on understanding Deep RL using attribution.
Online influence maximization has attracted much attention as a way to maximize influence spread through a social network while learning the values of unknown network parameters. Most previous works focus on single-item diffusion. In this paper, we introduce a new Online Competitive Influence Maximization (OCIM) problem, where two competing items (e.g., products, news stories) propagate in the same network and influence probabilities on edges are unknown. We adapt the combinatorial multi-armed bandit (CMAB) framework for the OCIM problem, but unlike the non-competitive setting, the important monotonicity property (influence spread increases when influence probabilities on edges increase) no longer holds due to the competitive nature of propagation, which brings a significant new challenge to the problem. We prove that the Triggering Probability Modulated (TPM) condition for CMAB still holds, and then utilize the property of competitive diffusion to introduce a new offline oracle, and discuss how to implement this new oracle in various cases. We propose an OCIM-OIFU algorithm with such an oracle that achieves logarithmic regret. We also design an OCIM-ETC algorithm that has worse regret bound but requires less feedback and easier offline computation. Our experimental evaluations demonstrate the effectiveness of our algorithms.
Traditional federated learning algorithms impose strict requirements on the participation rates of devices, which limit the potential reach of federated learning. In this paper, we extend the current learning paradigm and consider devices that may become inactive, compute incomplete updates, and leave or join in the middle of training. We derive analytical results to illustrate how the flexible participation of devices could affect the convergence when data is not independently and identically distributed (IID), and when devices are heterogeneous. This paper proposes a new federated aggregation scheme that converges even when devices may be inactive or return incomplete updates. We finally discuss practical research questions an operator would encounter during the training, and provide answers based on our convergence analysis.
Due to the massive size of the neural network models and training datasets used in machine learning today, it is imperative to distribute stochastic gradient descent (SGD) by splitting up tasks such as gradient evaluation across multiple worker nodes. However, running distributed SGD can be prohibitively expensive because it may require specialized computing resources such as GPUs for extended periods of time. We propose cost-effective strategies to exploit volatile cloud instances that are cheaper than standard instances, but may be interrupted by higher priority workloads. To the best of our knowledge, this work is the first to quantify how variations in the number of active worker nodes (as a result of preemption) affects SGD convergence and the time to train the model. By understanding these trade-offs between preemption probability of the instances, accuracy, and training time, we are able to derive practical strategies for configuring distributed SGD jobs on volatile instances such as Amazon EC2 spot instances and other preemptible cloud instances. Experimental results show that our strategies achieve good training performance at substantially lower cost.
We consider the stochastic multi-armed bandit (MAB) problem in a setting where a player can pay to pre-observe arm rewards before playing an arm in each round. Apart from the usual trade-off between exploring new arms to find the best one and exploiting the arm believed to offer the highest reward, we encounter an additional dilemma: pre-observing more arms gives a higher chance to play the best one, but incurs a larger cost. For the single-player setting, we design an Observe-Before-Play Upper Confidence Bound (OBP-UCB) algorithm for $K$ arms with Bernoulli rewards, and prove a $T$-round regret upper bound $O(K^2\log T)$. In the multi-player setting, collisions will occur when players select the same arm to play in the same round. We design a centralized algorithm, C-MP-OBP, and prove its $T$-round regret relative to an offline greedy strategy is upper bounded in $O(\frac{K^4}{M^2}\log T)$ for $K$ arms and $M$ players. We also propose distributed versions of the C-MP-OBP policy, called D-MP-OBP and D-MP-Adapt-OBP, achieving logarithmic regret with respect to collision-free target policies. Experiments on synthetic data and wireless channel traces show that C-MP-OBP and D-MP-OBP outperform random heuristics and offline optimal policies that do not allow pre-observations.
Modern vehicle fleets, e.g., for ridesharing platforms and taxi companies, can reduce passengers' waiting times by proactively dispatching vehicles to locations where pickup requests are anticipated in the future. Yet it is unclear how to best do this: optimal dispatching requires optimizing over several sources of uncertainty, including vehicles' travel times to their dispatched locations, as well as coordinating between vehicles so that they do not attempt to pick up the same passenger. While prior works have developed models for this uncertainty and used them to optimize dispatch policies, in this work we introduce a model-free approach. Specifically, we propose MOVI, a Deep Q-network (DQN)-based framework that directly learns the optimal vehicle dispatch policy. Since DQNs scale poorly with a large number of possible dispatches, we streamline our DQN training and suppose that each individual vehicle independently learns its own optimal policy, ensuring scalability at the cost of less coordination between vehicles. We then formulate a centralized receding-horizon control (RHC) policy to compare with our DQN policies. To compare these policies, we design and build MOVI as a large-scale realistic simulator based on 15 million taxi trip records that simulates policy-agnostic responses to dispatch decisions. We show that the DQN dispatch policy reduces the number of unserviced requests by 76% compared to without dispatch and 20% compared to the RHC approach, emphasizing the benefits of a model-free approach and suggesting that there is limited value to coordinating vehicle actions. This finding may help to explain the success of ridesharing platforms, for which drivers make individual decisions.
Motivated by ride-sharing platforms' efforts to reduce their riders' wait times for a vehicle, this paper introduces a novel problem of placing vehicles to fulfill real-time pickup requests in a spatially and temporally changing environment. The real-time nature of this problem makes it fundamentally different from other placement and scheduling problems, as it requires not only real-time placement decisions but also handling real-time request dynamics, which are influenced by human mobility patterns. We use a dataset of ten million ride requests from four major U.S. cities to show that the requests exhibit significant self-similarity. We then propose distributed online learning algorithms for the real-time vehicle placement problem and bound their expected performance under this observed self-similarity.