Model Fingerprint: General and Efficient Interpretability from Fundamental Prediction Logic

We prove that the model fingerprint framework of Li al. (2020) is mathematically equivalent to the aggregate feature importance scores of Lundberg and Lee (2017)'s SHAP measure, but the model fingerprint method’s focus on logical components further enables an effective approximation approach with polynomial (as opposed to exponential) computation complexity. Overall, the model fingerprint approach provides several advantages: (1) it delivers explicit attribution of linear, nonlinear, pairwise interaction and higher-order interaction effects, (2) it allows for much faster computation of lower-order effects, and (3) it offers a direct and intuitive explanation of prediction logic.