Description
Date depot: 11 février 2021
Titre: Interpretable Machine Learning for Music Recommendation
Directeur de thèse:
Vincent GUIGUE (ISIR (EDITE))
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Intelligence artificielle
Resumé: Interpretability (or explainable Artificial Intelligence (AI)) is a crucial topic in machine learning. More and more modern fields use machine learning techniques, especially deep neural networks, to predict, classify, recommend and generate content. But the growing number of their computations units and millions of parameters is often considered detrimental to the ability to explain how they operate (black-box models) compared to more traditional methods such as decision trees or linear regressions for which the computation is simple enough to be fully rationalised (white-box models). Yet, interpretability is absolutely needed in many cases, as in the medical field, and generally with any task that involves sensitive data or requires to check the fairness of the decision process [1]. Beyond them, interpretability is a currently trending subject in many more fields because of its highly informative value. For users of a recommender system, for instance, it is a way to prevent the usual misunderstanding of why an item was recommended to them (imagine generating a nameless music playlist versus a one that is "based on" their favourite artists, a mood, a dynamic interaction) and for engineers and researchers, it is all about the how, it is a way to identify bias, flaws or redundancies in their models, choose its best configuration and create more efficient ones. The goal of this PhD is to overthrow the preconceived idea that deep learning models are necessarily black-box, and to build deep complex models that are also explainable for music
Doctorant.e: Afchar Darius