NÓBREGA, C. S. B.; http://lattes.cnpq.br/8188640293164060; NÓBREGA, Caio Santos Bezerra.
Resumo:
The increase in sophistication and complexity of Machine Learning (ML) models has turned them into black boxes where the reasoning behind their predictions is hard to understand by humans. Even for low-risk tasks such as movie recommendation, show why a recommendation was made has become a key requirement since it can potentially improve users’ trust and help them to make better decisions. In this sense, there is a growing need to equip such models with interpretability in order to make them clear or easy to understand. A popular approach to achieve this goal is known as post-hoc interpretability, which consists of separating the process of generating explanations from the complex model, i.e, adding a layer of interpretability on top of it. In this thesis, we propose to investigate post-hoc interpretability methods for complex recommender systems. In particular, we propose an adaptation of LIME (Local Interpretable Model-agnostic Explanations), a popular post-hoc interpretability method, whose objective is to learn an interpretable model, under a space of interpretable features in the neighborhood of the instance being predicted. In addition, we investigate how different neighborhood generation strategies can impact the quality of the explanations. We conduct offline experiments and show that our proposed adaptation is a promising alternative since it is comparable in terms of fidelity, i.e., can locally mimic the behavior of a complex recommender, and has the additional advantage of enabling different styles of explanations.
Finally, we show that the user’s consumption history is the neighborhood strategy that best
suits our approach.