Mahran Jazi, Sharva Gogawale, Irad Ben‐Gal · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3839960
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 learning (FL) enables collaborative model training across distributed clients without centralizing raw data, but its performance often degrades under non-independent and non-identically distributed (non-IID) data. This degradation is especially severe when clients observe only a limited subset of the global label space, causing biased local updates and client drift. Standard aggregation methods such as FedAvg and FedProx mainly address this problem at the optimization level, but they do not directly enrich the incomplete local data distributions that cause heterogeneity. This paper proposes Correlation-Guided Latent Sharing (CGLS), a federated learning framework that selectively augments client training using latent information received from statistically aligned peers. CGLS identifies compatible clients through gradient-direction correlation and mixes local training data with synthetic latent information from correlated neighbors. We formulate this process as a mixed-distribution federated objective and explain how targeted mixing can reduce gradient dispersion under severe label skew. Experiments on MNIST, FashionMNIST and CIFAR-10 under Dirichlet and balanced one-class non-IID partitions show that CGLS improves accuracy and macro-F1 over FedAvg and FedProx, with the largest gains appearing in the strongest heterogeneity settings.
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
No comments yet — start the discussion below.