Sparse Biclustering of Transposable Data.

Area of Science:

• Data Mining
• Machine Learning
• Bioinformatics

Background:

• Simultaneous clustering of rows and columns (biclustering) is crucial for analyzing large, transposable data matrices.
• Existing methods like k-means clustering are often applied separately to rows or columns, missing potential bicluster structures.

Purpose of the Study:

• To develop a novel biclustering algorithm that simultaneously clusters rows and columns of a data matrix.
• To introduce sparsity and interpretability into biclustering through an L1 penalty on bicluster means.
• To provide a unified framework encompassing k-means and singular value decomposition as special cases.

Main Methods:

• The proposed method maximizes the log-likelihood under a bicluster-specific mean and common variance assumption for normally distributed matrix elements.
• An L1 penalty is applied to bicluster means to enforce sparsity and enhance interpretability.
• The framework is extended using the matrix-variate normal distribution.

Main Results:

• The proposed biclustering method is shown to be a sparse, symmetrized version of k-means clustering.
• Row and column k-means clustering are demonstrated as special cases of the proposed method.
• A relaxation of the proposed method yields the singular value decomposition (SVD).

Conclusions:

• The novel biclustering approach effectively identifies sparse and interpretable biclusters in large datasets.
• The method provides a generalized framework for biclustering, connecting existing techniques.
• Demonstrated performance on simulated data and a gene expression dataset highlights its practical utility.