A new influence measure for dependent features in classification problems
In this work, we introduce a novel approach to quantifying the influence of features on classification outcomes while accounting for dependencies among features. By adapting properties from cooperative game theory, we provide an axiomatic characterization and demonstrate that our measure generalizes the Banzhaf-Owen value for games with a priori unions. We also propose a sampling method for estimating the measure in large-scale datasets. Finally, the practical relevance of our proposals is illustrated through diverse applications.
Keywords: Classification problems Influence measure of features Dependent features Axiomatic characterization Banzhaf-Owen value