Empirical comparison of forward and backward search strategies in L-GEM based feature selection with RBFNN

Feature selection is one of important steps in pattern classification. Without a set of good features, one can not construct an efficient pattern classifier. Therefore, in addition to a good selection criterion, a good search strategy is also important to efficient and effective feature selection. Searching strategies of feature selection methods could be divided into several types: exhaustive searches, heuristic searches, floating searches and random searches. In this work, we perform a comparative study between two different search strategies: sequential forward search and sequential backward search. The Localized Generalization Error (L-GEM) is adopted as the selection criterion. Radial Basis Function Neural Network (RBFNN) is adopted as the classifier and we will perform experiments on UCI datasets. Experimental results show that the number of features of the dataset influences the performance of feature selections. Overall, the sequential backward search performs better when the number of features is large enough.