Incorporating Expert Knowledge into Rule Learning via Reinforcement Learning

Abstract

Rule learning algorithms have great interpretability compared with other machine learning models. They also express strong power in discovering interactions between diverse variables. However, the performance of rule learning is greatly limited by the quality and volume of available training data. We conduct a literature review to show that the combination of humans with machine learning algorithms is a solution to these problems. Thus, in this thesis, we propose an integration method of expert knowledge and rule learning via reinforcement learning (ERRL) which automatically involves an expert in the rule generation step of RuleFit. We split a node of decision tree in each time step by using the framework of the Markov Decision Process. Then we incorporate human knowledge with reinforcement learning by shaping rewards based on the expert responses to the chosen action. In an empirical evaluation, we train the ERRL model on a simulated dataset with binary variables. We show that incorporating expert knowledge can improve the classification accuracy compared with ERRL model without humans involved. ERRL model provides rule learning with an effective way to generate rules by iteratively and automatically engaging humans in the learning process.