GraphPFM: Pyramidal Feature Matching for Graph Classification

We present a novel method to learn graph representations that can be used for graph-level classification tasks. Our method first computes a node embedding based on its topological context. Three different methods are then proposed to obtain a graph embedding using a multi-level aggregation of node embeddings based on:(a) feature resolution;(b) spatial context of nodes; and (c) the natural hierarchy within the graph. Each of these multi-level aggregation methods generate pyramids of histograms where the histograms are levelwise ‘bag of nodes’ representation of the graph. We use pyramidal matching on these multi-level features to perform classification tasks, and outperform existing methods including some deep learning methods on benchmark graph classification datasets, despite being a purely topological method and not using node attributes or node labels.