The article presents LLpowershap, a novel feature selection method that uses loss-based Shapley values to identify informative features and reduce noise. The method also demonstrates higher or equal predictive performance compared to other Shapley-based methods. It enhances the recently introduced SHAP value-based feature selection method, powershap, for classification problems. LLpowershap distributes the mismatch between prediction and the truth calculated by the logistic loss function among the input features. The article further discusses the benefits and modifications of using LLpowershap compared to previous methods.
Publication date: 23 Jan 2024
Project Page: arXiv:2401.12683v1
Paper: https://arxiv.org/pdf/2401.12683
