Extracting interpretable models from matrix factorization models
Ivan Sanchez Carmona, Sebastian Riedel · Neural Information Processing Systems · 2015
Matrix factorization models have been successfully used in many real-world tasks, such as knowledge base completion and recommendation systems. However, explaining the causes that elicit a particular prediction by a manual inspection of its latent representations is a difficult task. In this paper we try to overcome this problem by exploring descriptive model classes in their ability to faithfully approximate the behavior of a pre-trained matrix factorization model. Crucially, our choice of descriptive model will allow us to provide an interpretable structured proof for each prediction of the original model. We compare the descriptive models in these two scopes: Fidelity and interpretability. We find that Bayesian network trees, a class of models that has not been considered for this purpose before, capture the matrix factorization model faithfully while providing multistep explanations of predictions.