Towards Interpretable Probabilistic Classification Models for Knowledge Graphs

Nicola Fanizzi, Claudia d'Amato

Abstract: Tackling the problem of learning probabilistic classifiers that can be used the context of knowledge graphs, we describe an inductive approach based on learning networks of Bernoulli variables. Namely, we consider the application of multivariate Bernoulli models, a simple one and a two-levels mixture model. In addition, we also consider a hierarchical model combining the multivariate Bernoulli model with a restricted Boltzmann machine as the first level. We show how such models can be converted into probabilistic rule bases ensuring more understandability. A preliminary empirical evaluation is presented to test the effectiveness of these models on a number of classification problems with different knowledge graphs.