Reproduce “Unsupervised feature selection using Bayesian Tucker decomposition” on OpenResearch

About this paper

Y-h. Taguchi, Yoh-ichi Mototake

Bayesian Tucker decomposition frames unsupervised feature selection by modeling reconstruction residuals as Gaussian, rather than treating components as a Gaussian null for outlier testing. Tucker factors and a core tensor identify feature candidates without feature labels. Across 100 synthetic ensembles, the method selected an average of 9.91 of 10 designated signals with zero false positives, a result limited to that benchmark.

Key insights

• Shifting the Gaussian assumption from components to residuals in BTuD makes it compatible with outlier-based unsupervised feature selection, addressing a gap in prior Bayesian TD work.
• The conventional HOOI algorithm can effectively serve as a practical substitute for obtaining BTuD solutions when the regularization parameter is zero, streamlining computational processes.
• The proposed BTuD-based feature selection method showed strong empirical equivalence to the authors' previously developed TD-based unsupervised FE, providing a statistical justification for the earlier approach.

Summary by alphaXiv.