A cluster-based sequential feature selection algorithm

Feature selection is an effective machine learning method for reducing dimensionality, removing irrelevant features, increasing learning accuracy, and improving result comprehensibility. However, many existing feature selection methods are incapable for high dimensional data because of their high time complexity, especially wrapper feature selection algorithms. In this work, a fast sequential feature selection algorithm (AP-SFS) is proposed based on affinity propagation clustering. AP-SFS divides the original feature space into several subspaces by a cluster algorithm, then applies sequential feature selection for each subspace, and collects all selected features together. Experimental results on several benchmark datasets indicate that AP-SFS can be implemented much faster than sequential feature selection but has comparable accuracies.