Feature selection via Boolean independent component analysis

We frame feature selection as a representation problem within a wider task of clustering feature vectors, and root its solution on a special procedure for extracting from them a set of boolean components that we expect to be independent. The overall clustering procedure is based on a divide et impera strategy: first give data a suitable representation then compute an assignment function. With the former we aim to find components of the feature vector minimizing – with the help of a special Schur-concave function – the mutual information between data and cluster features. We assess a subsymbolic tool to implement the optimization process and wisely use clustering algorithms to complete the procedure. We adopt the crucial problem of feature selection of DNA microarray data in cancer diagnosis as a benchmark to toss various aspects of the procedure.