We introduce the notion of classification game, a cooperative game with players as features and a characteristic function that is based on popular (SVM) hinge loss surrogate function. Shapley value, an important solution concept of a cooperative game, apportions the grand coalition cost among players as per their averaged marginal contributions. We investigate the role of Shapley values of these games for classification task. Shapley value based error apportioning, SVEA, of a feature is its contribution to the training loss. Thresholding SVEA values uniformly at 0 identifies key features for any dataset, yielding a clean procedure for feature subset selection. Among other things, we also estimate the true hinge loss risk of a feature. Overall, our approach is interpretable and explainable. We plan to close with some open problems.
N. Hemachandra is a Professor in Industrial Engineering and Operations Research, IITB. His current academic interests include, learning theory and algorithms, reinforcement learning, game theory, control, scheduling and pricing of multi-class queues and applications of these to resource allocation problems in various Engg and Tech settings.