Causal Discovery from Observational Data using Graph Neural Networks

Abstract

This paper explores the application of Graph Neural Networks (GNNs) for causal discovery directly from observational data. Traditional causal inference methods often rely on strong assumptions regarding underlying causal mechanisms, which can be difficult to validate. We propose a novel framework where a GNN learns causal relationships by analyzing the correlations within the data. The core idea is to train the GNN on observational data to generate node embeddings that encode both correlations and causal dependencies. We utilize techniques like mutual information maximization and intervention prediction to further refine the learned graph structure, enhancing the accuracy of the discovered causal relationships. This data-driven approach offers a flexible and potentially more robust method for causal discovery compared to traditional approaches, particularly when limited prior knowledge about the system is available. The methodology presented here leverages the representational power of GNNs to effectively extract causal signals from complex observational datasets.