This paper approaches the algorithm of selection of variables named MIFS-U and presents an alternative method for estimating entropy and mutual information, "measures" that constitute the base of this selection algorithm. This method has, for foundation, the Cauchy-Schwartz quadratic mutual information and the Rényi quadratic entropy, combined, in the case of continuous variables, with Parzen Window density estimation. Experiments were accomplished with public domain data, being such method compared with the original MIFS-U algorithm, broadly used, that adopts the Shannon entropy definition and makes use, in the case of continuous variables, of the histogram density estimator. The results show small variations between the two methods, what suggest a future investigation using a classifier, such as Neural Networks, to qualitatively evaluate these results, in the light of the final objective which is greater accuracy of classification.
variable selection; MIFS-U; entropy; mutual information; Shannon; Rényi; Parzen Window; Information-Theoretic Learning; ITL
