Méthode comprimée et distribuée de factorisation pondérée en matrices non-négatives pour les matrices de grande dimension

Weighted Non-Negative Matrix Factorization (WNMF) allows factoring a data matrix while taking into account the confidence in each data point. When the weights are not binary, most WNMF methods are poorly suited to processing large data matrices. In this paper, we tackle this problem by proposing a compressed (by random projections) and distributed WNMF approach. Random projections are among the main techniques to process big data. They have been combined with WNMF in an EM formalism that requires some computations which might not fit in memory. In this paper, we propose a technique which does not require such computations and which can distribute the update rules of the factor matrices.

Recommended citation: Matthieu Puigt, Asmae El Hyani, Kinan Abbas, Gilles Roussel, and Guillaume Caron. (2025). "Méthode comprimée et distribuée de factorisation pondérée en matrices non-négatives pour les matrices de grande dimension." In Proceedings of XXXe Colloque Francophone de Traitement du Signal et des Images (GRETSI 2025), Aug 2025, Strasbourg, France
Download Paper