ProxySHAP: Efficient Approximation of Shapley and Banzhaf Interactions
A novel technique named ProxySHAP has been developed to effectively estimate Shapley and Banzhaf interaction values, which reflect intricate dynamics in machine learning models. Existing estimators for these advanced interactions struggle with a balance between speed and precision. ProxySHAP merges the sample efficiency of tree-based proxy models with a systematic approach to consistency through residual correction. This method features a polynomial-time extension of interventional TreeSHAP, enabling the calculation of exact interaction indices for tree ensembles while avoiding exponential dependencies related to tree depth. Moreover, the analysis of the residual adjustment strategy defines the conditions under which Maximum Sample Reuse (MSR) mitigates proxy bias without exponentially increasing variance with interaction size. Comprehensive benchmarking indicates that ProxySHAP establishes a new benchmark in the field.
Key facts
- ProxySHAP reconciles tree-based proxy models with residual correction for consistency.
- Polynomial-time generalization of interventional TreeSHAP for exact interaction indices.
- Bypasses exponential tree-depth dependencies in prior methods.
- Maximum Sample Reuse (MSR) corrects proxy bias without exponential variance scaling.
- Extensive benchmarking demonstrates state-of-the-art performance.
- Method addresses Shapley and Banzhaf interactions in machine learning.
- Current estimators trade off speed and accuracy.
- ProxySHAP achieves high sample efficiency.
Entities
—