Causal Graph Generation Algorithms for Explainable Artificial Intelligence

Abstract

The increasing deployment of complex artificial intelligence (AI) models in critical applications necessitates methods for understanding and explaining their decision-making processes. Traditional model interpretability techniques often struggle to capture the underlying causal relationships driving model predictions, leading to a superficial understanding of how models arrive at their conclusions. This paper proposes a novel algorithm for automatically generating causal graphs from AI models, leveraging techniques of counterfactual reasoning and causal discovery. The core claim is that these generated graphs provide a transparent representation of the causal relationships within the model, significantly enhancing its explainability. The mechanism involves analyzing the relationships between model inputs and outputs using these techniques, resulting in a graph that accurately reflects the model's decision-making logic. We demonstrate a new approach to interpreting complex AI models, increasing their trustworthiness.