Causal Inference via Graph Neural Networks with Directional Edges

Abstract

This paper proposes a novel approach to causal inference leveraging Graph Neural Networks (GNNs) with explicitly modeled directional edges. Traditional GNNs primarily focus on learning representations based on pairwise relationships, often leading to spurious correlations. To address this, we introduce a modification to GNN architectures that incorporates directional edges, representing hypothesized causal influences between nodes. This allows the network to learn and propagate causal effects, significantly improving the accuracy of causal inference tasks. We demonstrate the efficacy of this approach through a theoretical framework and outline the key components for implementation. The core claim is that GNNs can be enhanced by explicitly modeling causal relationships through directed edges. The underlying mechanism involves modifying GNNs to incorporate directional edges representing causal influences. This shifts the focus from purely correlational relationships to explicitly modeling causal dependencies, offering a more robust and interpretable method for causal inference. ---