Subspaces alignment and graph convolutional networks performance

Tuesday, April 2, 2019

2.30 p.m.

ISI seminar room 1st floor

Dr. Paul Expert - Imperial College London

Graph neural networks refer to a fairly broad set of emerging techniques attempting to generalize deep neural models to graphs [Bronstein 2017], thus extending notions from deep learning techniques to graph structured data. Recently, the particular method of Graph Convolutional Networks (GCNs) has been shown to perform particularly well in node classification tasks by combining the information contained in the features and relational graphs of the objects to be classified [Kipf 2017].
Although GCNs have been shown to perform well based on benchmarks, it remains unclear when and why the information contained in the graph can enhance the performance of GCNs compared to other simpler architectures such as multilayer perceptron.
In this talk, I will show how a certain degree of alignment between the information contained in the graph, the features, and implicitly in the ground truth, is needed to obtain good performance of GCNs.
We propose a metric that establishes a direct geometric relationship between the performance of the GCN classification and the spectral alignment of the graph, features and ground truth, and enables us to quantify the synergy needed for GCN to perform well. The metric we propose has a much wider range of applications.