Towards Extracting Faithful and Descriptive Representations of Latent Variable Models.

Methods that use latent representations of data, such as matrix and tensor factorization or deep neural methods, are becoming increasingly popular for applications such as knowledge base population and recommendation systems. These approaches have been shown to be very robust and scalable but, in contrast to more symbolic approaches, lack interpretability. This makes debugging such models difficult, and might result in users not trusting the predictions of such systems. To overcome this issue we propose to extract an interpretable proxy model from a predictive latent variable model. We use a socalled pedagogical method, where we query our predictive model to obtain observations needed for learning a descriptive model. We describe two families of (presumably more) descriptive models, simple logic rules and Bayesian networks, and show how members of these families provide descriptive representations of matrix factorization models. Preliminary experiments on knowledge extraction from text indicate that even though Bayesian networks may be more faithful to a matrix factorization model than the logic rules, the latter are possibly more useful for interpretation and debugging.