End-to-End Learning of Probabilistic Hierarchies on Graphs

Contributions 4 1. We propose a probabilistic model over hierarchies via continuous relaxation of a tree's parent assignment matrices. 2. We theoretically analyze the model by drawing connections to absorbing Markov chains, which 3. allows efficient and exact computation of lowest-common-ancestor (LCA) probabilities, which enables us to 4. learn hierarchies on graphs by end-to-end optimization of relaxed versions of quality metrics such as Dasgupta cost and TSD score.