Generating rationale for molecular prediction using reinforcement learning

Abstract

This thesis studies generation of rationale for neural prediction problems using reinforcement learning. In particular, we focus on neural predictions in chemical property prediction tasks. We design a reinforcement learning agent that learns to incrementally extract the important regions of molecular graphs, and construct a predictor trained on only the selected regions. The ability for the model to predict a property based only on the partial graph exemplifies the importance of these substructures and therefore can be interpreted as rationales for the prediction task. We test our reinforcement learning model on several chemical datasets and show that our model can generate meaningful rationales while maintaining good predictive performances.