A Cooperation Index for Model Pruning

ICLR 2026 Conference Submission18266 Authors

Keywords: Parameter Importance, Model Pruning, SHAP value

TL;DR: We propose a new criterion, cooperation Index, for effective model pruning to improve both generalizability and interpretability of the model..

Abstract: In complex models, pruning redundant parameters reveals its core functional elements, improving both generalizability and interpretability. Effective pruning can be achieved by criteria to identify redundant parameters, and the SHAP Value (SV) has been considered such a criterion, which is interpreted as averaging the marginal contributions across all possible parameter accumulation paths. However, we find that its averaging process systematically overweights redundant parameters, failing as a decision-making agent. Instead, quantifying the speed of decay of the marginal contribution can serve as a more effective decision criterion for model pruning. We show that it is more effective to count the number of cooperative contributions of parameters for pruning parameters in backward elimination, leading to a more optimal set of remaining parameters.

Primary Area: interpretability and explainable AI

Submission Number: 18266