Towards Structurally Robust Cross-Device Federated Recommendation
Federated recommendation (FedRec) allows training recommender systems without collecting user data centrally[3]. However, in these settings, particularly in cross-device scenarios where each client corresponds to a single user, only local interaction histories are available. Consequently, the learning process relies on partial, fragmented, and sparse signals. This, in turn, makes collaboration between users indirect, with only a subset of users participating in each training round. In practice, this means the model is trained on incomplete and potentially biased data. Standard Federated Learning (FL) methods assume that simply averaging updates is enough, but in this setting, that is not sufficient to recover meaningful collaborative patterns. This thesis studies the following question: "How do federated constraints influence the propagation of collaborative signals in recommendation systems, and how can key mechanisms be redesigned to improve performance, fairness, and cross-domain transfer under realistic cross-device conditions? " To address this question, the thesis is structured around four interconnected research directions: Similarity-aware aggregation for personalized representations: Standard federated aggregation usually combines client updates uniformly or by data size, even when user representations reflect very different behavioral contexts. This can dilute useful personalization signals and amplify mismatched updates. This direction proposes distance-based aggregation, where user-specific parameters are weighted by representation similarity, with the goal of improving the performance of graph FedRec systems[2]. Controlled User Expansion and eligibility-aware client sampling : User expansion is a common method used in graph FedRec systems to leverage high order interactions and counteract the small local interaction graph each client has access to [4]. However, existing methods often rely on trusted third-party servers and introduce uncontrolled replication. This direction proposes a nearest neighbor based user expansion, and pairs it with a client selection strategy that account for both direct participation and indirect influence. The goal is to reduce over-representation of frequently selected clients and improve participation fairness. Bias and Fairness Analysis in FedRec : Fairness in centralized recommendation has been studied extensively, but it is less clear how FL settings change bias dynamics. Intermittent participation, non-IID data, and limited access to sensitive attributes may alter how disparities emerge and evolve. This direction compares centralized and FedRecs under aligned architectures and evaluation protocols to isolate the effect of FL on bias in recommendation. It also adapts fairness-aware training (based on [1]) so that group-level statistics can be estimated securely, without exposing sensitive information centrally. Structural bridging for cross-domain FedRec : Cross-domain recommendation in FL settings is especially difficult when overlap between domains is sparse [5]. In such cases, natural transfer paths are weak and knowledge sharing depends on very limited shared structure. In this direction we plan to explore synthetic bridge construction from semantic item representations to create additional cross-domain connections while preserving privacy. By selectively activating useful bridges and pruning harmful ones, the work aims to improve transfer in low-overlap regimes while keeping negative transfer low.