A framework for differentiable Supervised Graph Prediction

Paul KRZAKALA, Junjie Yang, Rémi Flamary, Florence d'Alché-Buc, Charlotte Laclau, Matthieu Labeau

Keywords: Graph Prediction, Graphs, Deep Learning, Optimal Transport, Differentiable

TL;DR: We introduce a general framework to train a deep neural network to output a graph from a variety of input modalities. The framework is built using a novel Optimal Transport loss.

Abstract: We introduce a general framework to train a deep neural network to output a graph from a variety of input modalities. The framework is built using a novel Optimal Transport loss that exhibits all necessary properties (permutation invariance and differentiability) and allows for handling graphs of any size. We showcase the versatility and state-of-the-art performances of the proposed approach on various real-world tasks and a novel challenging synthetic dataset.

Submission Number: 1