Temporal Motif Signatures for Temporal Graph Neural Networks

2026-05-31

AI Summary

Existing methods struggle to effectively model predictive local interaction motifs—such as reciprocation, repetition, star patterns, and triadic flows—in temporal graphs. This work proposes a compact, leakage-free 13-dimensional temporal motif feature mapping that can be linearly embedded into any static or temporal graph neural network without modifying its original architecture, thereby enhancing its sensitivity to local temporal patterns. Leveraging the observation that motif activity organizes along three stable axes across scales, we construct a universal set of candidate features and characterize their expressive power within the temporal Weisfeiler–Leman framework. Experiments demonstrate consistent performance gains across diverse tasks—including link property prediction, edge classification, and graph-level classification—when integrating our motif representation into five baseline models.

Abstract

Real temporal interaction streams carry predictive structure in short-horizon motif patterns -- repetition, reciprocity, star diversity, triadic flow -- that vanilla temporal graph neural networks (TGNNs) often fail to expose to their edge scorers. We show this concretely on MOOC interaction prediction, where a small four-feature family of past-window star counts already delivers most of the lift over a strong static GNN. Across a wide set of real and synthetic temporal datasets we find that motif activity organizes consistently along three scale-stable axes (dyadic recency/reciprocity, star diversity, triadic flow), and we use this empirical structure to design a compact 13-coordinate, leakage-safe, candidate-local motif feature map h(u, v, t) that linearly embeds into any static or temporal encoder without architectural changes. A temporal Weisfeiler-Leman (WL) analysis places the augmentation relative to the first level of an anchored temporal-WL hierarchy and exhibits a candidate-anchored pair on which motif features distinguish. We demonstrate empirically that the same augmentation consistently lifts performance across heterogeneous tasks: TGB link-property prediction across all five baselines, edge classification on Bitcoin Alpha/OTC and MOOC, and graph-level classification of synthetic temporal generators.

Problem

• temporal graph neural networks
• temporal motifs
• predictive structure
• interaction streams
• edge prediction

Innovation

• temporal motifs
• temporal graph neural networks
• motif signatures
• temporal Weisfeiler-Leman
• feature augmentation