Yanchao Tang, Baoyao Yang, Yuhao Chen, Chenxiang Zhao · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840792
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).
Prototype-based federated learning (FL) provides a promising route for federated heterogeneity. By exchanging class prototypes rather than model parameters, it provides class-level semantic guidance under non-IID data and accommodates heterogeneous client backbones. Under severely skewed label distributions, however, two coupled limitations emerge. The broadcast global prototype is a cross-client average, and direct adoption can cause client semantic dilution. Meanwhile, local prototypes of rare classes are estimated from a few samples, leading to local margin erosion against well-supported neighbours. To address both, we propose ProRC, a prototype personalization framework grounded in shrinkage estimation. ProRC replaces naive alignment with a reliability-calibrated, margin-aware interpolation between global and local prototypes, and comprises two complementary client-side modules. The first module, Reliability-Calibrated Prototype Adaptation (RaPA), calibrates each global prototype by the reliability of its local counterpart, restoring client-specific semantics only to the extent the data warrant. The second module, Margin-Aware Prototype Refinement (MaPR), audits the candidate against the local true-class margin and attenuates margin-reducing steps. Together, the two modules are designed to mitigate client semantic dilution and local margin erosion, while preserving the prototype-exchange protocol and per-round payload of prior methods. Extensive experiments on label-skewed and real-world image benchmarks, together with ablation and reliability analyses, show that ProRC achieves the best overall average accuracy among the compared methods and improves the strongest prototype baseline in most heterogeneous settings.
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.