PreCLAS: An evolutionary tool for unsupervised feature selection

Abstract

Several research areas are being faced with data matrices that are not suitable to be managed with traditional clustering, regression, or classification strategies. For example, biological so-called omic problems present models with thousands or millions of rows and less than a hundred columns. This matrix structure hinders the successful progress of traditional data analysis methods and thus needs some means for reducing the number of rows. This article presents an unsupervised approach called PreCLAS for preprocessing matrices with dimension problems to obtain data that are apt for clustering and classification strategies. The PreCLAS was implemented as an unsupervised strategy that aims at finding a submatrix with a drastically reduced number of rows, preferring those rows that together present some group structure. Experimentation was carried out in two stages. First, to assess its functionality, a benchmark dataset was studied in a clustering context. Then, a microarray dataset with genomic information was analyzed, and the PreCLAS was used to select informative genes in the context of classification strategies. Experimentation showed that the new method performs successfully at drastically reducing the number of rows of a matrix, smartly performing unsupervised feature selection for both classification and clustering problems.