Singular vector decomposition based hybrid pattern search — An efficient co-clustering method

Abstract

With the rapid development of machine-learning and data-mining techniques, biclustering (co-clustering) has become an important and widespread technique in multiple areas such as gene expression analysis, text mining and market segmentation. In this work, we proposed an efficient co-clustering method named SVD-based hybrid pattern search (SHPS). It is a score-function-based method, and specifically both the mean square-residue and correlation-based scores were tested in our studies. For a data matrix, SHPS first uses SVD layers to approximate it, and then searches the SVD subspaces for hybrid patterns (cliquish or linear) along the row or column direction. Groups along the two directions are combined, and those with a score smaller than a pre-defined threshold will be outputted. After testing our method on multiple types of matrices and comparing it with the traditional Cheng and Church method, SHPS showed a good performance with multiple co-clusters and better scores. Additionally, using more SVD layers may further improve the results. Overall, SHPS can be a good and efficient alternative in future co-clustering-related studies and applications.