Applying Genetic Programming to Improve Interpretability in Machine Learning Models

GPX offers symbolic explanations that humans can understand: Unlike black-box AI systems, the Genetic Programming Explainer generates tree-structured mathematical expressions that show why a model made a specific decision in plain, interpretable terms rather than opaque numerical outputs.

The method works across different AI systems: The research tested GPX on three major types of complex models (Random Forest, Deep Neural Networks, and Support Vector Machines) across 20 datasets, demonstrating that this approach to explainability is broadly applicable rather than limited to one AI technique.

Local explanations focus on individual decisions: GPX analyzes the specific neighborhood around each prediction, meaning it can explain why an AI system made a particular decision for one person or case without needing to explain the entire model—making it practical for real-world accountability.

GPX outperforms existing explainability methods: The paper reports that this approach produces "more accurate understanding of complex models than the state of the art," suggesting it's a meaningful improvement over current techniques for making AI decisions transparent to stakeholders and regulators.

Interpretability can be achieved without sacrificing model performance: By generating accurate local explanations of existing complex models rather than replacing them with simpler ones, GPX allows organizations to maintain AI system accuracy while still meeting transparency and accountability requirements.