Ranking Categorical Feature Levels in Supervised Classification via Shapley and Owen Value
This paper proposes a novel framework to measure the influence of individual categories within categorical variables in supervised classification problems. Instead of evaluating features as a whole, the approach focuses on category-level contributions by modeling them as players in a cooperative game. The value of each coalition is defined through predictive performance, allowing the use of Shapley and Owen values to quantify the marginal contribution of each category. Additionally, alternative formulations are introduced to address issues such as data sparsity and to capture the effect of categories on fixed predictive models.
Keywords: multi-class classification problem categorical feature levels ranking internal structure game theory