Tianzi Zang, Jiayi Sun, Cheng Xu, Mingyuan Gao, Yicong Li, Juan Li · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840870
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Federated recommendation (FedRec) suffers from sparse and highly heterogeneous interaction data, which can drive client-side item representations toward inconsistent semantic directions and cause feature conflict during aggregation. Addressing feature conflict in FedRec is non-trivial, as implicit feedback reflects user-specific preferences rather than stable item semantics, making it difficult to support prototype-based alignment as in standard FL. To address this issue, we propose MoSA, a FedRec-specific feature alignment framework that introduces multi-modal semantic anchors via item clustering. MoSA uses the resulting cluster assignments as globally consistent pseudo-labels and performs dual-level alignment: (i) a cluster-aware item representation alignment module to guide client-side item embedding matrix under partial item coverage, and (ii) a recommender modulation module to improve cross-client parameter compatibility while preserving personalization. Extensive experiments on real-world datasets demonstrate that MoSA consistently improves recommendation performance across representative federated recommendation backbones, and ablation studies validate the effectiveness of each component. The codes are available here.
Neural Networks 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
No comments yet — start the discussion below.