Section: New Results

An oracle inequality for Quasi-Bayesian Non-Negative Matrix Factorisation

Participant : Benjamin Guedj.

We have extended the quasi-Bayesian perspective to the popular setting of non-negative matrix factorisation. This is a pivotal problem in machine learning (image segmentation, recommendation systems, audio source separation, ...) and we were able to propose an original estimator of the unobserved matrix. An oracle inequality is derived, along with several possible implementations. This work is now submitted to an international journal [38].

Joint work with Pierre Alquier.