A1487
Title: Distributed and personalized PCA via matrix beta-mean aggregation
Authors: Zhi-Yu Jou - Academia Sinica (Taiwan) [presenting]
Abstract: Distributed PCA (DPCA) has become increasingly important for analyzing large-scale data stored across multiple machines. A fundamental challenge in DPCA lies in efficiently aggregating information from multiple machines while retaining the statistical properties of the original dataset. Existing methods that aggregate only local leading eigenvectors, while ignoring important eigenvalue information, may lead to suboptimal estimation. To address this issue, beta-DPCA is proposed, a distributed PCA method that incorporates eigenvalue information through matrix beta-mean aggregation. In many real-world applications, however, local datasets often exhibit heterogeneity, whereas many standard DPCA methods implicitly assume a common principal subspace across datasets. To address this limitation, a personalized extension of beta-DPCA is further proposed that decomposes each local covariance structure into shared and individualized components. The proposed framework enables robust extraction of common features across data sources while preserving source-specific patterns.
