Explainable AI via Graph-Based Causal Inference for Model Decisions

Abstract

The increasing deployment of deep learning models in critical applications necessitates a means to understand and trust their decisions. However, deep learning models are often 'black boxes,' offering limited insight into the reasoning behind their predictions. This paper proposes a novel approach to explain model decisions by leveraging graph-based causal inference. We represent the model's decision-making process as a graph, where nodes represent input features and the model's output, and edges represent causal relationships inferred from the model's behavior. By analyzing the structure of this graph, we can identify the key factors driving a specific prediction, providing a clear and interpretable explanation. The core of this approach is to move beyond simply identifying correlations between features and the output and instead to explicitly model the causal influences. This offers a more robust and reliable explanation than traditional methods. We demonstrate the effectiveness of this technique through a theoretical framework and discuss its potential applications in various domains. The resulting visual representation of the causal graph provides a significantly improved understanding of model decision-making compared to purely correlational explanations.