ExplainabilityTransparencyAccountabilitycs.LG
Induction of Non-monotonic Logic Programs To Explain Statistical Learning Models
Farhad Shakerin
September 18, 2019
ICLP Technical Communications
Key Takeaways
-
Logic programs can make AI models more interpretable: The paper demonstrates that complex statistical learning models (like ensemble trees) can be translated into non-monotonic logic programs—human-readable rule-based systems that explain how the model makes decisions, addressing a key challenge in responsible AI.
-
SHAP-based feature importance improves explanation quality: By using TreeExplainer (a SHAP implementation) to identify which features matter most for individual predictions, the method produces more accurate and relevant explanations compared to previous approaches, suggesting that feature importance weighting is crucial for faithful model explanations.
-
Scalability matters for practical AI deployment: The algorithm shows "significant improvement in terms of...running time" compared to existing systems, meaning explainability techniques can now be applied to larger, real-world models without prohibitive computational costs—an important consideration for responsible AI in production systems.
-
Rule-based explanations enable better auditing: Converting statistical models into explicit logic programs creates transparent decision rules that stakeholders, regulators, and affected individuals can understand and scrutinize, supporting accountability and trust in AI systems.
Abstract
We present a fast and scalable algorithm to induce non-monotonic logic programs from statistical learning models. We reduce the problem of search for best clauses to instances of the High-Utility Itemset Mining (HUIM) problem. In the HUIM problem, feature values and their importance are treated as transactions and utilities respectively. We make use of TreeExplainer, a fast and scalable implementation of the Explainable AI tool SHAP, to extract locally important features and their weights from ensemble tree models. Our experiments with UCI standard benchmarks suggest a significant improvement in terms of classification evaluation metrics and running time of the training algorithm compared to ALEPH, a state-of-the-art Inductive Logic Programming (ILP) system.