Another use of SMOTE for interpretable data collaboration analysis

• DOI: 10.1016/j.eswa.2023.120385
• Publication Date: 2023-05-09T01:05:10Z
• Authors: Akira Imakura; Masateru Kihira; Yukihiko Okada; Tetsuya Sakurai
• Keywords: Representation

Recently, data collaboration (DC) analysis has been developed for privacy-preserving integrated across multiple institutions. DC centralizes individually constructed dimensionality-reduced intermediate representations and realizes via without sharing the original data. To construct representations, each institution generates shares a shareable anchor dataset its representation. Although, random functions well in general, using an whose distribution is close to that of raw expected improve recognition performance, particularly interpretable analysis. Based on extension synthetic minority over-sampling technique (SMOTE), this study proposes construction performance increasing risk leakage. Numerical results demonstrate efficiency proposed SMOTE-based method over existing constructions artificial real-world datasets. Specifically, achieves 6, 4, 36 percentage point improvements regarding NMI, ACC essential feature selection, respectively, methods income dataset. The provides another use SMOTE not imbalanced classifications but key technology