Using Inductive Logic Programming to globally approximate Neural Networks for preference learning: challenges and preliminary results

In this paper we explore the use of Answer Set Programming (ASP), and in particular the state-of-the-art Inductive Logic Programming (ILP) system ILASP, as a method to explain black-box models, e.g. Neural Networks (NN), when they are used to learn user preferences. To this aim, we created a dataset of users preferences over a set of recipes, trained a set of NNs on these data, and performed preliminary experiments that investigate how ILASP can globally approximate these NNs. Since computational time required for training ILASP on high dimensional feature spaces is very high, we focused on the problem of making global approximation more scalable. In particular we experimented with the use of Principal Component Analysis (PCA) to reduce the dimensionality of the dataset while trying to keep our explanations transparent.