Functional Graph-to-Graph Transformation Layers

Author(s):

Berend Christopher Rinderspacher

Abstract:

We derive three types of layers for use in graph-to-graph machine learning models on labeled graphs. Three layers are introduced: a vertex preserving layer; a graph contraction layer, which reduces the number of vertices; and a graph expansion layer, which increases the number of vertices. All three layers are constructed using the theory of intersection graphs. The framework is tested against the Modified National Institute of Standards and Technology image classification task, for which a very small model of just over 2000 parameters achieves over 95% accuracy in both training and test sets, compared with the state-of-the-art performance of ProjectionNet with over 77,000 parameters or the parametric matrix model with just under 5000 parameters.