Bandit-based structure learning for bayesian network classifiers

In this work, we tackle the problem of structure learning for Bayesian network classifiers (BNC). Searching for an appropriate structure is a challenging task since the number of possible structures grows exponentially with the number of attributes. We formulate this search problem as a large Markov Decision Process (MDP). This allows us to tackle the problem using sequential decision making methods. Furthermore, we devise a Monte Carlo tree search algorithm to find a tractable solution for the MDP. The use of bandit-based action selection strategy enables us to have a systematic way of guiding the search, making the search in the large space of unrestricted structures tractable. The results of classification on different datasets show that the use of this method can significantly boost the performance of structure learning for BNCs.