Ongoing project about Geometric Matrix Completion with Recurrent Multi-Graph Neural Networks for Jean Feydy's Geometric Data Analysis course, in a group of 3 with Antonin Barbe and Zoé Herson. Matrix Completion models are one of the most common ways of formulating recommendation algorithms. The matrix considered contains users as columns, items as rows and the scores assigned by users to each item as coefficients. The authors propose to use prior graph knowledge between users and/or items. They use an architecture based on graphical spectral convolutions, as well as a recurrent neural network to complete this matrix. At the time of the paper's release in 2017, they assert to beat the state of the art of matrix completion algorithms on the considered datasets.

(Ongoing) This repository contains a re-implementation of the paper's original code from 2017 (in the mgcnn folder and coded in Tensorflow) in PyTorch, as well as methods for evaluating architecture performance. Finally, a proposal for a new architecture is formulated.