AMIFS: adaptive feature selection by using mutual information

An adaptive feature selection method based on mutual information, called AMIFS, is presented. AMIFS is an enhancement over Battiti´s MIFS and MIFS-U methods. In AMIFS the tradeoff between eliminating irrelevance or redundancy is controlled adoptively, instead of using a fixed parameter. The mutual information is computed by discrete probabilities in the case of discrete features or by using an extended version of Fraser´s algorithm in the case of continuous features. The performance of AMIFS is compared with that of MIFS and MIFS-U on artificial and benchmark datasets. The simulation results show that AMIFS outperforms both MIFS and MIFS-U, specially for high-dimensional data with many irrelevant and/or redundant features.