An end-to-end framework for Supervised Graph Prediction.
By Paul Krzakala
Paul Krzakala will give a talk on an end-to-end framework for Supervised Graph Prediction.
Abstract
Our work is about Supervised Graph Prediction (SGP), that is any supervised learning task where the output to predict is a graph. Since the graphs are expected to be of arbitrary sizes the output space has a complex non-euclidean structure. As often in structured prediction problems, most of the existing work relies on surrogate representation of the output space which comes at the price of a computationally expensive decoding step. Our new approach, on the other hand, allows us to tackle those tasks in a fully end-to-end manner through a standard deep learning pipeline. To this end, we introduce a novel framework composed of three key components: a representation suited to graphs of arbitrary sizes, an associated neural network architecture and an optimal transport loss that is both differentiable, scalable and permutation invariant.
arxiv link: https://arxiv.org/pdf/2402.12269.pdf
Biography
PhD student with Florence d’Alche-Buc (LTCI, Télécom), Charlotte Laclau (LTCI, Télécom) and Rémi Flamary (CMAP, Polytechnique).