The paper introduces a novel Symbolic XAI framework that computes the relevance of logical formulas (queries) composed of a functionally complete set of logical connectives (conjunction and negation) to explain the predictions of machine learning models.
The key highlights are:
The framework decomposes the model's prediction into multi-order terms that represent the relevance of different feature subsets. This decomposition can be obtained using either propagation-based or perturbation-based explanation methods.
The relevance of a logical formula (query) is computed by filtering the multi-order terms where the query holds true and summing their corresponding relevance values. This allows capturing the relevance of complex logical relationships between features.
The paper proposes a search algorithm to identify the queries that best describe the model's prediction strategy by maximizing the correlation between the query's filter vector and the multi-order terms.
The effectiveness of the Symbolic XAI framework is demonstrated across three domains: natural language processing, computer vision, and quantum chemistry. The results show that the framework can provide insights into the model's decision-making process that are more human-understandable compared to classic feature-wise explanations.
다른 언어로
소스 콘텐츠 기반
arxiv.org
더 깊은 질문