DAPER joint learning from partially structured graph databases

Abstract

In this paper, we are interested in learning specific probabilistic relational models, named Directed Acyclic Probabilistic Entity Relationship (DAPER) models, from partially structured databases. Algorithms for such a learning task already exist for structured data coming from a relational database. They have been also extended to partially structured data stored in a graph database where the Entity Relationship (ER) schema is first identified from data, and then the DAPER dependency structure is learnt for this specific ER schema. We propose in this work a joint learning from partially structured graph databases where we want to learn at the same time the ER schema and the probabilistic dependencies. The Markov Logic Network (MLN) formalism is an efficient solution for this task. We show with an illustrative example that MLN structure learning can effectively learn both parts of the DAPER model in one single task, with a comparative precision, but with a very high complexity.