Interpretability and Explainability in Machine Learning and their Application to Nonlinear Dimensionality Reduction

Abstract

Machine learning (ML) techniques are more and more frequently used today because of their high performance in many contexts. However, the rise in performance comes at the cost of a lack of control over the model that is learned. Indeed, while modelling was mainly done by experts in the past, the surge of data makes it possible to automatically derive models. Unfortunately, this automatization can result in the production of non-understandable models. This concept of model understandability is referred to as interpretability in the literature. Furthermore, when models are not interpretable, it is their ability to be explained (their explainability) that is exploited.

This thesis explores interpretability and explainability in ML. Several aspects of these concepts are studied. First, the problem of defining interpretability and explainability, as well as the vocabulary used in the literature, is presented. Second, the requirements of the law for these concept are studied. Then, the way interpretability and explainability involve users in their evaluation is discussed and guidelines from the human-computer interaction community are presented.

This thesis also applies the concepts of interpretability and explainability to the problem of nonlinear dimensionality reduction (NLDR). While the subjects of interpretability and explainability in NLDR have barely been touched in the literature, this thesis provides a conceptualization of interpretability and explainability in the context of NLDR, as well as new techniques to deal with them. In particular, two questions are central in this thesis ``how can interpretability can be measured in NLDR?'' and ``how can non-interpretable NLDR mappings be explained?''.

For measuring interpretability in NLDR, we analyze how existing metrics from different communities can be combined to predict user understanding of NLDR embeddings. In particular, ML quality metrics are used to assess how low-dimensional (LD) embeddings are faithful to the high-dimensional (HD) data, and information visualization quality metrics are used to assess how understandable visualizations are. In the context of NLDR mappings that are considered to be non-interpretable, IXVC was developed to explain the mapping between visual clusters in a NLDR embedding and HD data through an interactive pipeline. Another approach for explaining NLDR mappings through the embedding dimensions was developed in our two techniques BIR and BIOT. Even though previous work has tried to develop more explicit, parametric, mappings, to the best of our knowledge, our works in this thesis are the first to elaborate on the term ``interpretability'' in the field of NLDR.

Date of Award	16 Nov 2020
Original language	English
Awarding Institution	University of Namur
Sponsors	University of Namur
Supervisor	Benoît Frénay (Supervisor), Wim Vanhoof (President), Anthony Cleve (Jury), Bruno Dumas (Jury), John Aldo Lee (Jury) & Luis A. Galarraga (Jury)