Davoud Gholamiangonabadi, Muhammad Umair Danish, Zhenghua Chen, Katarina Grolinger · Machine Learning with Applications 2026 · 2026
DOI: 10.1016/j.mlwa.2026.101046
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).
Sentiment analysis increasingly faces challenges with traditional centralized machine learning as data are often decentralized and transferring them to a central location raises privacy and security risks. Federated learning (FL) mitigated these risks by training models locally without data sharing, yet its performance and convergence degrade under non-Independent and Identically Distributed (non-IID) data. To improve performance under non-IID data, Clustered FL (CFL) groups clients into clusters and trains a specialized model per group. However, traditional CFL methods face two critical limitations: (1) reliance on test client labels for cluster evaluation, which violates ML principles by overestimating performance, and (2) difficulty in identifying clusters without sharing local data. This paper proposes an unsupervised Autoencoder (AE)-based clustering method for FL that groups clients by leveraging latent representations of their data distributions, eliminating the need for test labels or data sharing. Additionally, a proximity-aware weighting mechanism prioritizes clients closer to cluster centers during aggregation, countering the bias of the conventional FedAvg method that favors larger datasets while disregarding smaller but valuable clients. Experiments across various non-IID sentiment analysis scenarios confirm that the proposed AE-based clustering combined with weighted aggregation successfully forms clusters without label dependency and improves FL performance. The proposed CFL approach improves accuracy significantly–from 68.92% to 78.36% for clients with balanced label distributions, and by up to 50% for extremely non-IID clients with only a single label.
No comments yet — start the discussion below.