Machine Learning Regressed Causal Inference for Discrete ANM

In this paper, Random Forest and Deep Neural Network regressed causal inference algorithms are proposed for discovering causal direction in discrete Additive Noise Model (ANM). Assuming ANM, a key step in determining causal direction is to find out the functional relation between cause and effect. This step is solved by deep-network- and random-forest-based regressions between cause and effect variables. The experiments show that Random Forest and Deep Neural Network have a good robustness and high accuracy rate in both simulated data and real data, especially Random Forest.