Feature evaluation using quadratic mutual information

Methods of feature evaluation are developed and discussed based on information theoretical learning (ITL). Mutual information was shown in the literature to be more robust and precise to evaluate a feature set. We propose to use quadratic mutual information (QMI) for feature evaluation. The concept of information potential leads to a more clearly physical meaning of the evaluation functions. Moreover, evaluation for feature sets in high-dimensional space could also be implemented efficiently. Experimental results are compared to classifier performances