DAG-GCN: Directed Acyclic Causal Graph Discovery from Real World Data using Graph Convolutional Networks

Abstract

Causal discovery has been challenging since the search space of directed acyclic graphs super-exponentially grows with respect to the number of nodes. Previously constraint-based and score-based methods have been used. In recent studies, a continuous optimization method has reached a high score, but the problem is still harsh in real-world observational data. Motivated by the success of recent GNN models, we extended previous methods to be suitable to actual world data. Our model is based on the DAG-GNN model, uses GCN, and tries to learn an adjacency matrix set as a model parameter. To solve the vanishing adjacency matrix problem, we use the He-Initialization method with Leaky ReLU and the batch normalization technique. We demonstrate our model on real-world data sets. Compared to the state-of-the art results, our proposed method reaches acceptable results.