Feature Selection Scheme Based on Zero-Sum Two-Player Game

We propose a new filter methodology for feature selection using the concept of game theory whereby features are assimilated to players. In this game theoretical context, a strategy corresponds to a particular affinity between a group of features forming a cluster, and the payoff function is computed based on the weighted distance between a feature and a cluster. A zero-sum two-player game problem is solved through a global combination of pair wise features. Finally, each feature is represented by the value of the objective function, at the optimal solution, which indicates the contribution of each feature. The importance of features is then evaluated by their optimal values. To validate the effectiveness of the proposed methodology, we have conducted a classification task utilizing SVM on various UCI and stat log datasets. The experimental results show that the proposed scheme leads to improvement in classification performance, when compared to mRMR and Fisher score algorithms.