Generalized Branch and Bound Algorithm for Feature Subset Selection

Branch and bound algorithm is a good method for feature selection which finds the optimal subset of features of a given cardinality when the criterion function satisfies the monotonicity property. To find an optimal feature subset of a different cardinality the method needs to be applied from the beginning. Also the method cannot be used when one do not know the cardinality of the subset that is required. This paper presents a generalization over the branch and bound algorithm which first finds optimal subsets of features of varying cardinalities in a single run. Then a method is given to find the best subset of features. The proposed method is experimentally verified and is found to be a faster and a suitable one when one do not know the number of features in the best subset of features.