Cluster-Dependent Feature Selection for the RBF Networks

摘要

The paper addresses the problem of a radial basis function network initialization with feature section. Main idea of the proposed approach is to achieve the reduction of data dimensionality through feature selection carried-out independently in each hidden unit of the RBFN. To select features we use the so called cluster-dependent feature selection technique. In this paper three different algorithms for determining unique subset of features for each hidden unit are considered. These are RELIEF, Random Forest and Random-based Ensembles. The processes of feature selection and learning are carried-out by program agents working within a specially designed framework which is also described in the paper. The approach is validated experimentally. Classification results of the RBFN with cluster-dependent feature selection are compared with results obtained using RBFNs implementations with some other types of feature selection methods, over several UCI datasets.