Scaling Up Graph Homomorphism Features with Efficient Data Structures

Paul Beaujean, Florian Sikora, Florian Yger

Keywords: Graph embeddings, Approximations, Feature generation, Scaling and parallel processing

TL;DR: We use Bloom filters to scale up a sampling algorithm which computes an additive approximation of graph homomorphism densities relative to pattern graphs.

Abstract: Typical datasets used in graph classification tasks only contain a few thousand graphs which rarely exceed hundreds of nodes. Graph homomorphism densities are permutation-invariant features that can be directly computed from graph data, and their approximation scales naturally to large graphs. We propose the use of efficient data structures for approximate set membership in the context of a sampling algorithm for graph homomorphism density which enables the use of large-scale datasets containing larger graphs. To validate our findings, we compare this method with existing approaches used for graph homomorphism features in synthetic experiments.