An Approach to Reduce the Number of Conditional Independence Tests in the PC Algorithm

Abstract

The PC algorithm is one of the most prominent constraint-based methods for learning causal structures from observational data. The algorithm relies on conditional independence (CI) tests to infer the structure and its time consumption heavily depends on the number of performed CI tests. We present a modification, called ED-PC, such that – in the oracle model – both ED-PC and the original PC algorithm infer the same structure. However, by using a new idea allowing the detection of a v-structure without explicit knowledge of a separating set, our method reduces the number of needed CI tests significantly. This is made possible by detecting nonadjacencies considerably earlier.