Abstract: Discovering causality from the data set is one of the basic problems in the scientific field. However, there is still no effective algorithm to discover causalities from highdimensional data. We present an algorithm for large scale causality discovery, which is based on conditional independence test and mutual information. Firstly, the algorithm reduces the dimensionality of data set by conditional independence test and mutual information, and then we employ an approach based on nonlinear additive noise models to distinguish the directions between nodes. Experimental results show that it outperforms other methods.

Key words: causality inference; conditional independence test; mutual information