Federated Learning (FL) presents a compelling distributed learning paradigm that enables resource-constrained clients to collaboratively train machine learning models while preserving data privacy. However, inter-client data heterogeneity poses fundamental challenges to federated optimization efficacy. Although Knowledge Distillation (KD) effectively addresses model performance alignment under heterogeneity, its inherent constraints restrict client parameter exploration capacity, thus, inducing confinement to suboptimal basins. To reconcile this trade-off, we introduce FedKDMR, a novel FL paradigm unifying KD constraints with exploration via model recombination. FedKDMR imposes global model consistency and robustness in training through dynamic KD while sufficiently harnessing model recombination-induced perturbations for diverse parameter exploration. We establish convergence guarantees for strongly convex and smooth objectives. Extensive experiments on FL benchmark datasets demonstrate that FedKDMR achieves a superior accuracy-robustness trade-off against state-of-the-art methods when tackling non-independent and identically distributed and heterogeneous data in FL environments.

Abstract In this paper, we investigate Thompson sampling-based sequential block elimination approaches for dynamic assignment problems in a pure-exploration Multi-Armed Bandit (MAB) setting with limited budget constraints. The problem can be considered as a bandit game-play between the environment and a decision-maker in a metric space. Many instances of problems in fields such as e-commerce, logistics, mobility management, data management and operations research can be framed as dynamic assignment problems with budget constraints. Given an l-dimensional action space representing l variants of an entity and a budget for exploring the action space, the optimal dynamic assignment problem refers to the task of identifying the values to be assigned to different variants of the entity that maximizes the total reward by utilizing at most the given budget of rounds of play. We contribute a class of block elimination-based MAB algorithms specifically designed for the dynamic assignment problem with limited budget. Our algorithms begin by discretizing the continuous action space into a finite set of discrete actions, then proceed with a recursive block elimination procedure to remove sub-optimal actions. The elimination is carried out by calculating confidence bounds over blocks of actions. We explore two different confidence bound estimation techniques. We perform comprehensive experiments on two problem instances from distributed data management and logistics. Our results showcase that our approach yields a lower misidentification probability (i.e., the probability of recommending a non-optimal action) compared to state-of-the-art elimination-based pure-exploration MAB algorithms.

Federated Learning (FL) is used to train machine learning models across distributed data and clients (devices). A key challenge in FL is addressing the complexities posed by heterogeneous clients with varying computational resources. While effective methods like knowledge distillation have been proposed to assist clients with limited resources, prevailing teacher-student models often focus solely on knowledge transfer, neglecting whether student clients can effectively and inclusively assimilate this knowledge. We introduce a leader-driven hierarchical distributed inclusive FL framework (LIFE) to address this challenge. LIFE engages hierarchical leaders across clients with various levels of computational capacities, aiming to optimize the performance of subordinate client models through a novel all-inclusive, leader-driven, level-by-level aggregation. Knowledge distillation is gradually achieved by fine-tuning subordinate models at all levels. Comprehensive experiments and comparisons with state-of-the-art methods across different tasks showcase that LIFE significantly enhances knowledge transfer across a variety of clients in an all-inclusive FL environment.

Distributed analytics involving classification tasks demand robust model training. Real-time arbitrary classification tasks on distributed clients pose challenges due to constraints in data sharing. Federated (Meta)-Learning (FML) has been introduced for global distributed (meta)-model training, which generalizes well over distributed data and classification tasks. Current FML approaches assume fixed labels over unskewed class proportions and data distributions along with uniform task distributions. However, global meta-models can only be used for tasks that do not require addressing arbitrary out-of- distribution label issues. In real-world cases, class imbalance and label shifting are common issues in clients' data. On-demand tasks arriving at clients involve unseen labels. Therefore, 'one (meta)-model-fits-all‘ is not the best option. To address these challenges, we introduce multiple cluster-based meta-models, each one tailored to specific label distribution. Our framework, coined Cluster-based & Label-aware Federated Meta-Learning (CL-FML), involves distributed client clustering based on label shifting and cluster-based FML identifying the most suitable clients to engage per task. CL-FML leverages lightweight data augmentation to deal with arbitrary class-imbalanced tasks. Our comprehensive experiments and comparative assessment against baselines showcase that CL-FML efficiently achieves high accuracy by fast convergence, significantly reducing training rounds and communication load.

Abstract We study the problem of predicting the next query to be recommended in interactive data exploratory analysis to guide users to correct content. Current query prediction approaches are based on sequence-to-sequence learning, exploiting past interaction data. However, due to the resource-hungry training process, such approaches fail to adapt to immediate user feedback. Immediate feedback is essential and considered as a signal of the user’s intent. We contribute with a novel query prediction ensemble mechanism, which adapts to immediate feedback relying on multi-armed bandits framework. Our mechanism, an extension to the popular Exp3 algorithm, augments Transformer-based language models for query predictions by combining predictions from experts, thus dynamically building a candidate set during exploration. Immediate feedback is leveraged to choose the appropriate prediction in a probabilistic fashion. We provide comprehensive large-scale experimental and comparative assessment using a popular online literature discovery service, which showcases that our mechanism (i) improves the per-round regret substantially against state-of-the-art Transformer-based models and (ii) shows the superiority of causal language modelling over masked language modelling for query recommendations.

Federated Learning (FL) has been successfully adopted for distributed training and inference of large-scale Deep Neural Networks (DNNs). However, DNNs are characterized by an extremely large number of parameters, thus, yielding significant challenges in exchanging these parameters among distributed nodes and managing the memory. Although recent DNN compression methods (e.g., sparsification, pruning) tackle such challenges, they do not holistically consider an adaptively controlled reduction of parameter exchange while maintaining high accuracy levels. We, therefore, contribute with a novel FL framework (coined FedDIP), which combines (i) dynamic model pruning with error feedback to eliminate redundant information exchange, which contributes to significant performance improvement, with (ii) incremental regularization that can achieve extreme sparsity of models. We provide convergence analysis of FedDIP and report on a comprehensive performance and comparative assessment against state-of-the-art methods using benchmark data sets and DNN models. Our results showcase that FedDIP not only controls the model sparsity but efficiently achieves similar or better performance compared to other model pruning methods adopting incremental regularization during distributed model training. The code is available at: https://github.com/EricLoong/feddip.

Abstract Recommender systems are widely used in online platforms for easy exploration of personalized content. The best available recommendation algorithms are based on using the observed preference information among collaborating entities. A significant challenge in recommender system continues to be item cold-start recommendation: how to effectively recommend items with no observed or past preference information. Here we propose a two-stage algorithm based on soft clustering to provide an efficient solution to this problem. The crux of our approach lies in representing the items as soft-cluster embeddings in the space spanned by the side-information associated with the items. Though many item embedding approaches have been proposed for item cold-start recommendations in the past—and simple as they might appear—to the best of our knowledge, the approach based on soft-cluster embeddings has not been proposed in the research literature. Our experimental results on four benchmark datasets conclusively demonstrate that the proposed algorithm makes accurate recommendations in item cold-start settings compared to the state-of-the-art algorithms according to commonly used ranking metrics like Normalized Discounted Cumulative Gain (NDCG) and Mean Average Precision (MAP). The performance of our proposed algorithm on the MovieLens 20M dataset clearly demonstrates the scalability aspect of our algorithm compared to other popular algorithms. We also propose the metricCold Items Precision(CIP) to quantify the ability of a system to recommend cold-start items. CIP can be used in conjunction with relevance ranking metrics like NDCG and MAP to measure the effectiveness of the cold-start recommendation algorithm.

We consider the problem of generating diverse, personalized recommendations such that a small set of recommended items covers a broad range of the user's interests. We represent items in a similarity graph, and we formulate the relevance/diversity trade-off as finding a small set of unrated items that best covers a subset of items positively rated by the user. In contrast to previous approaches, our method does not rely on an explicit trade-off between a relevance objective and a diversity objective, as the estimations of relevance and diversity are implicit in the coverage criterion. We show on several benchmark datasets that our approach compares favorably to the state-of-the-art diversification methods according to various relevance and diversity measures.