Probabilistic sentential decision diagrams (PSDDs) were recently introduced as a tractable and interpretable representation of discrete probability distributions. PSDDs are tractable because they support a wide range of queries efficiently. They are interpretable because each parameter in the PSDD represents a conditional probability, as in Bayesian networks. This paper summarizes ongoing research that aims to answer two questions that are important to employ PSDDs as an explainable AI model. First, as a tractable and interpretable model, can PSDDs compete with more general machine learning models for density estimation? We answer this question positively, reporting state-of-the-art results on standard benchmarks. Second, can we effectively reduce the number of parameters in a learned PSDD to simplify its interpretation, without harming the quality of the learned model? For this task, we present an algorithm that merges PSDD substructures that are similar in KL-divergence, which we show can be done efficiently on PSDDs.

@inproceedings {LiangXAI17,
  author={Liang, Yitao and Van den Broeck, Guy},
  title = {Towards Compact Interpretable Models: Shrinking of Learned Probabilistic Sentential Decision Diagrams},
  booktitle = {IJCAI 2017 Workshop on Explainable Artificial Intelligence (XAI)},
  month = Aug,
  year = {2017},
  url = "https://starai.cs.ucla.edu/papers/LiangXAI17.pdf",
  keywords={workshop}
}