FairFRL: Fairness-aware Federated Representation Learning for Cross-domain Sequential Recommendation

Le résumé fourni par la source

Cross-domain sequential recommendation is increasingly important in modern Web ecosystems, where user behaviors span multiple independently operated services that maintain strict data isolation for privacy and regulatory compliance. Federated learning offers a practical paradigm for such cross-domain collaboration, but user preferences evolve asynchronously across services, creating a substantial distribution shift. This drift leads to unstable and unequal domain contributions: behaviorally rich domains dominate global updates, while low-resource or volatile domains exert limited influence. Such an imbalance degrades recommendation accuracy and raises fundamental fairness concerns. To address these challenges, we propose FairFRL, a fairness-aware federated representation learning framework designed to mitigate contribution imbalance under dynamic cross-domain drift. FairFRL mitigates contribution imbalance under dynamic cross-domain drift by jointly regulating domain influence during federated aggregation and disentangling domain-shared and domain-exclusive semantics, while preserving data locality. Experiments on real-world Amazon multi-domain datasets show that FairFRL consistently outperforms strong federated and centralized baselines across multiple metrics and achieves more equitable cross-domain contributions. These results position FairFRL as a principled step toward responsible, fair, and socially aligned Web recommendation systems.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.