Enhancing Deep Learning with Statistical Inference for Node Classification in Temporal Graphs

Jan von Pichowski, Vincenzo Perri, Lisi Qarkaxhija, Ingo Scholtes

Keywords: graph neural networks, statistical inference, random graph ensembles, temporal graphs, higher-order networks

TL;DR: We develop SIT-GNN, a two-step temporal graph neural network that first identifies significant time-respecting sequences through statistical inference and then leverages these sequences to classify the nodes in a temporal graph.

Abstract: Modeling temporal and sequential patterns in temporal graphs is a critical research challenge in the development of time-aware GNNs. We propose a significance test based on a graph null model where timestamps are randomly shuffled to identify significant time-respecting paths, i.e., sequences of time-stamped edges that follow a temporal order. By combining this inference method with graph learning, we develop a two-step model, SIT-GNN, capable of capturing significant sequences in time-respecting paths. We demonstrate this novel capability with synthetic data and explain the enhanced classification performance in empirical data through an analysis with respect to the significance test. To the best of our knowledge, our work is the first to introduce statistically informed GNNs that leverage sequential patterns in terms of time-respecting paths. SIT-GNN represents a step towards bridging the gap between statistical graph inference and neural graph representation learning, with potential applications to static GNNs.

Format: Long paper, up to 8 pages.

Submission Number: 6