Optimal Feature Subset Selection Using Similarity-Dissimilarity Index and Genetic Algorithms

Optimal feature subset selection is an important pre-processing step for classification in many real life problems where number of dimensions of feature space is large and some features are may be irrelevant or redundant. One example of such a situation is genes expression profile data to classify among normal and cancerous samples. Contribution of this paper is five folds. Similarity-dissimilarity index (MSDI) is proposed which can estimate the class discrimination quality of the high dimensional feature space without using any kind of classifier. A framework to find out the best features subset from the n-dimensional feature space using genetic algorithm is proposed to select the minimum possible important features optimally using MSDI as fitness function to evolve the population. Similarity-dissimilarity plot is proposed to visualize the high dimensional data that can be used to extract important information about the class discrimination quality of the feature space. It is possible to predict the best classification accuracy using MSDI when an appropriate classifier is used. Another index called average differential of similarity and dissimilarity distances above similarity-dissimilarity line is proposed which gives information about how far each class instances or clusters are from other classes and the compactness of the classes in the feature space. Effectiveness of the methods is highlighted by using a large set of benchmark datasets in cancer classification and size of features subset and predicted classification accuracy is compared with the published results.