CPD Tree Learning Using Contexts as Background Knowledge

Context specific independence (CSI) is an efficient means to capture independencies that hold only in certain contexts. Inference algorithms based on CSI are capable to learn the Conditional Probability Distribution (CPD) tree relative to a target variable. We model motifs as specific contexts that are recurrently observed in data. These motifs can thus constitute a domain knowledge which can be incorporated into a learning procedure. We show that the integration of this prior knowledge provides better learning performances and facilitates the interpretation of local structure.