Unsupervised feature selection using Bayesian Tucker decomposition

Researchers Y-h. Taguchi and Yoh-ichi Mototake have introduced a novel unsupervised feature selection method called Bayesian Tucker decomposition (BTuD). Published in a 24-page arXiv paper with 10 supporting figures, the method treats residuals as Gaussian-distributed similar to linear regression frameworks. The authors demonstrate that conventional higher-order orthogonal iteration can generate Tucker decompositions consistent with their Bayesian implementation, providing a bridge between established tensor decomposition methods and probabilistic approaches.

BTuD was successfully tested across multiple complex datasets including synthetic data, global coupled maps with randomized coupling strength, and real-world gene expression profiles. The method's unsupervised nature means it can identify meaningful features without requiring labeled training data—a significant advantage for domains like genomics where labeled data is scarce. The researchers note that BTuD-based unsupervised feature extraction is expected to align with previously proposed tensor decomposition methods that have proven effective across diverse problem domains.

The paper represents a methodological advancement in automated feature engineering, particularly for high-dimensional data where manual feature selection is impractical. By framing Tucker decomposition within a Bayesian framework, the researchers provide both theoretical grounding and practical algorithms for unsupervised dimensionality reduction. This approach could accelerate analysis pipelines in fields ranging from bioinformatics to complex systems modeling where identifying relevant features from raw data remains a fundamental challenge.