TLDR: This paper introduces AxFi, two novel rigorous feature importance scores based on Shapley value and Banzhaf index. Unlike prior methods, AxFi considers non-WAXp sets and quantifies a feature’s effectiveness in excluding adversarial examples, offering more detailed insights. It provides strong guarantees, is computationally efficient for decision trees, and addresses limitations of existing methods like SHAP, making it valuable for high-stakes AI applications.
In the rapidly evolving world of Artificial Intelligence (AI), understanding how complex machine learning (ML) models make decisions is becoming increasingly crucial. This need for transparency and accountability has led to the rise of eXplainable Artificial Intelligence (XAI). One of the most common approaches in XAI involves feature attribution methods, which aim to identify how much each input feature contributes to a model’s prediction.
Traditional game theory-based methods, like SHAP scores, are widely used for this purpose. However, recent research has highlighted some limitations. Specifically, many rigorous feature attribution methods often rely on ‘weak abductive explanations’ (WAXp) as a basis for assigning importance. A key drawback of this approach is that it tends to overlook the contributions of ‘non-WAXp’ sets of features. These non-WAXp sets can still hold important information, especially concerning their relationship with ‘adversarial examples’ – subtly altered inputs designed to trick an ML model.
A new research paper, titled Rigorous Feature Importance Scores based on Shapley Value and Banzhaf Index, introduces a novel solution to this challenge. Authored by Xuanxiang Huang, Olivier Létoffé, and Joao Marques-Silva, the paper proposes two new feature importance scores, collectively called AxFi (Adversarial examples-based Feature importance). These scores are built upon the well-established game theory concepts of Shapley value and Banzhaf index, but with a crucial difference: they explicitly consider non-WAXp sets.
The core innovation of AxFi lies in its ability to quantify how effective each feature is at excluding adversarial examples. This means the scores provide a more fine-grained and informative perspective on feature importance in formal XAI. To achieve this, the authors introduce a new characteristic function called ‘contrastive explanation forest’ (CXp-Forest), which serves as the foundation for computing these detailed importance scores.
The paper highlights several key advantages of AxFi. Firstly, in high-stakes and safety-critical applications, such as medical diagnosis, AxFi offers rigorous guarantees that SHAP scores might not. This is particularly important because SHAP scores have been shown to sometimes assign low or zero contributions to features that are logically relevant. AxFi, on the other hand, satisfies a property called ‘Consistency with relevancy,’ ensuring that relevant features always receive a non-zero score.
Secondly, for decision tree models, AxFi scores can be computed in polynomial time, making them more efficient than some existing rigorous feature importance measures like Formal Feature Attribution (FFA) and Weighted Formal Feature Attribution (WFFA). While SHAP scores can be faster in general, especially for complex models like boosted trees and logistic regression, AxFi provides a valuable alternative where rigorous guarantees are paramount.
The research also delves into the properties satisfied by AxFi, including Efficiency, Symmetry, Additivity, Null player, and a newly defined ‘CXp-minimal monotonicity.’ This ensures that if a feature’s set of contrastive explanations is included in another’s, its score will be no greater.
Also Read:
- Unpacking AI Explanations: Why Simplicity Matters for True Insight
- Holistic Explainable AI: A New Framework for End-to-End AI Transparency
In experimental comparisons, AxFi scores often show high similarity with each other (Shapley-like vs. Banzhaf-like AxFi), but can differ significantly from SHAP and WFFA in terms of feature ranking. This suggests that AxFi offers a distinct and valuable perspective on feature importance, complementing existing methods by providing a more robust and informative measure, especially when dealing with the critical task of identifying and excluding adversarial examples.
