The article introduces CAGE, a causality-aware global explanation framework based on Shapley values. The key contributions are:
CAGE introduces a novel sampling procedure for out-coalition features that respects the causal relations of the input features, overcoming the independence assumption made by previous global explanation methods.
The authors show theoretically that CAGE satisfies desirable causal properties, indicating that it is designed from first principles.
Empirical analysis on both synthetic and real-world data demonstrates that explanations from CAGE are more faithful compared to causally agnostic global explanation methods like SAGE.
The article first provides background on causal models, interventions, and Shapley-based global explanations. It then presents the CAGE framework, proving its causal soundness. An approximation algorithm is also introduced to compute the CAGE values.
The experiments on synthetic data show that CAGE can better capture the true causal feature importance compared to SAGE, especially when there are causal dependencies among the features. On the real-world Alzheimer's disease dataset, while the differences are less pronounced, CAGE still exhibits the pattern of reducing the importance of features that are solely effects of other features.
The discussion highlights the challenges of CAGE, such as the requirement of a predefined causal structure, and suggests future research directions to overcome these limitations.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Nils Ole Bre... at arxiv.org 04-18-2024
https://arxiv.org/pdf/2404.11208.pdfDeeper Inquiries