CMIM-2: An Enhanced Conditional Mutual Information Maximization Criterion for Feature Selection

A new greedy feature selection criterion is proposed as an enhancement of the conditional mutual information maximization criterion (CMIM). The new criterion, called CMIM-2, allows detecting relevant features that are complementary in the class prediction better than the original criterion. In addition, we present a methodology to approximate the conditional mutual information to spaces of three variables, avoiding its estimation in high-dimensional spaces. Experimental results for artificial and UCI benchmark datasets show that the proposed criterion outperforms the original CMIM criterion.