Non-parametric Learning of Embeddings for Relational Data Using Gaifman Locality Theorem

We consider the problem of full model learning from relational data. To this effect, we construct embeddings using symbolic trees learned in a non-parametric manner. The trees are treated as a decisionlist of first order rules that are then partially grounded and counted over local neighborhoods of a Gaifman graph to obtain the feature representations. We propose the first method for learning these relational features using a Gaifman graph by using relational tree distances. Our empirical evaluation on real data sets demonstrates the superiority of our approach over handcrafted rules, classical rule-learning approaches, the state-of-the-art relational learning methods and embedding methods.