DiCausal: Exploiting Domain Knowledge for Interactive Causal Discovery

Abstract

We propose an interactive causal discovery system called DiCausal, which allows users to apply their simple domain knowledge of how variables are generated and interactively edit the graph during the causal discovery process without incurring too much burden. A novel form of domain knowledge representation and an adapted feature engineering method are introduced in DiCausal. Two existing causal discovery algorithms are adapted for verification. Experiment proves that such a way of incorporating domain knowledge into the discovery algorithms can achieve better results than pure data-driven methods.