Towards an Inductive Logic Programming Approach for Explaining Black-Box Preference Learning Systems

In this paper we advocate the use of Inductive Logic Programming as a device for explaining black-box models, e.g. Support Vector Machines (SVMs), when they are used to learn user preferences. We present a case study where we use the ILP system ILASP to explain the output of SVM classifiers trained on preference datasets. Explanations are produced in terms of weak constraints, which can be easily understood by humans. We use ILASP both as a global and a local approximator for SVMs, score its fidelity, and discuss how its output can prove useful e.g. for interactive learning tasks and for identifying unwanted biases when the original dataset is not available. Finally, we highlight directions for further work and discuss relevant application areas.