[Paper Review] Understanding Global Feature Contributions With Additive Importance Measures

Ian Covert, Scott Lundberg+1 more|arXiv (Cornell University)|Apr 1, 2020

Explainable Artificial Intelligence (XAI)42 references168 citations

TL;DR

The paper unifies global feature importance via an additive framework and introduces SAGE, a model-agnostic method using Shapley values to account for feature interactions and provide faster, reliable global explanations.

ABSTRACT

Understanding the inner workings of complex machine learning models is a long-standing problem and most recent research has focused on local interpretability. To assess the role of individual input features in a global sense, we explore the perspective of defining feature importance through the predictive power associated with each feature. We introduce two notions of predictive power (model-based and universal) and formalize this approach with a framework of additive importance measures, which unifies numerous methods in the literature. We then propose SAGE, a model-agnostic method that quantifies predictive power while accounting for feature interactions. Our experiments show that SAGE can be calculated efficiently and that it assigns more accurate importance values than other methods.

Motivation & Objective

• Define global feature importance in terms of predictive power and its interaction with feature subsets.
• Introduce a unifying framework of additive importance measures for various existing methods.
• Propose SAGE, a model-agnostic method using Shapley values to account for feature interactions.
• Provide an efficient sampling-based approximation for calculating SAGE values with uncertainty estimates.
• Demonstrate that SAGE yields more accurate feature attributions and can detect corrupted features.

Proposed method

• Formalize predictive power for subsets of features through model-based and universal metrics.
• Define additive importance measures where a proxy u(S) approximates v(S) via u(S)=phi0+sum_{i in S} phi_i.
• Introduce SAGE as Shapley values of v_f, the predictive power of the model from feature subsets.
• Propose a sampling-based algorithm to approximate SAGE values with convergence guarantees and uncertainty estimates.
• Show connections between SAGE and SHAP, and discuss invariance to invertible feature transformations.
• Provide an efficient computational approach that outperforms naive SHAP-based calculation for global explanations.

Experimental results

Research questions

• RQ1 How can global feature importance be defined in terms of predictive power across feature subsets?
• RQ2 How can additive importance measures unify existing global feature importance methods?
• RQ3 Can SAGE accurately quantify a model’s dependence on each feature while accounting for interactions?
• RQ4 Is there an efficient, model-agnostic way to compute SAGE values with uncertainty estimates?
• RQ5 Do SAGE values align with intuitive and theoretical properties (e.g., Shapley axioms, mutual information in optimal settings)?

Key findings

• SAGE assigns feature importances as Shapley values of the model-based predictive power, summing to the total predictive power.
• SAGE satisfies efficiency, symmetry, dummy, monotonicity, and linearity properties, making it a principled credit allocation.
• SAGE is invariant to invertible feature transformations and relates to SHAP, providing additive approximations to v_f across the feature domain.
• A sampling-based approximation for SAGE is significantly faster than naive local SHAP calculations and provides uncertainty estimates.
• Empirical comparisons show SAGE more accurately reflects predictive power than baselines like permutation tests, mean importance, and ablation in several datasets.
• SAGE can help identify corrupted features in suboptimal models and monitor model integrity over time.