Molecular descriptor selection combining genetic algorithms and fuzzy logic: application to database mining procedures

A new algorithm, devoted to molecular descriptor selection in the context of Data Mining problems, has been developed. This algorithm is based on the concepts of genetic algorithms (GA) for descriptor hyperspace exploration and combined with a stepwise approach to get local convergence. Its selection power was evaluated by a fitness function derived from a fuzzy clustering method. Different training and test sets were randomly generated at each GA generation. The fitness score was derived by combining the scores of the training and test sets. ility of the proposed algorithm to select relevant subsets of descriptors was tested on two data sets. The first one, an academic example, corresponded to the artificial problem of Bullseye, the second was a real data set including 114 olfactory compounds divided into three odor categories. In both cases, the proposed method allowed to improve the separation between the different data set classes.