Feature selection using tabu search with long-term memories and probabilistic neural networks

Feature selection is a dimensionality reduction problem in order to reduce measurement costs, shorten computational time, relieve the curse of dimensionality, and improve classification accuracy. In this paper, a hybrid approach using tabu search and probabilistic neural networks is proposed and applied to feature selection problems. The proposed tabu search algorithm differs from previous research by using a long-term memory instead of a short-term memory to avoid the necessity of the delicate tuning of the memory length and to decrease the risk of generating a cycle that traps the search in local optimal solutions. The probabilistic neural networks integrated in the proposed hybrid approach are an outgrowth of Bayesian classifiers that outperform backpropagation-based neural networks in their global convergence and rapid training. Extensive experiments on real-world data sets are performed and the comparison with previous research indicates that the proposed hybrid approach can select an equal or smaller number of features while improving classification accuracy.