Cluster-based pattern discrimination: A novel technique for feature selection

The study of feature selection methods has become an area of intensive research in pattern recognition. In this paper, a new feature selection approach, called cluster-based pattern discrimination (CPD), is introduced. Classes are independently partitioned into clusters to group together similar patterns: a different subspace is defined for each cluster by determining an optimal subset of features. The similarity between an unknown pattern x and a given cluster is computed through a classifier. To combine these similarities we use the “max rule” which simply assigns each pattern to the class that contains the cluster for which the pattern has the maximum similarity. Moreover, extensive experiments carried out on different databases prove the advantages of the proposed approach.