Title: Temporal Motif Signatures for Temporal Graph Neural Networks

Abstract:

Vanilla temporal graph neural networks (TGNNs) frequently overlook the predictive power of short-horizon motif patterns—such as repetition, reciprocity, star diversity, and triadic flow—embedded within real temporal interaction streams. We demonstrate this deficiency using MOOC interaction prediction, revealing that a simple family of four features, specifically past-window star counts, captures the majority of performance gains over robust static GNNs. Our analysis of numerous real-world and synthetic temporal datasets reveals that motif activity consistently clusters along three scale-invariant dimensions: dyadic recency/reciprocity, star diversity, and triadic flow. Leveraging this empirical structure, we introduce a compact, 13-coordinate motif feature map, $h(u, v, t)$, which is candidate-local and free from data leakage. This map linearly integrates into any static or temporal encoder without requiring architectural modifications. Furthermore, a temporal Weisfeiler-Leman (WL) analysis situates this augmentation within the first level of an anchored temporal-WL hierarchy, highlighting specific candidate-anchored pairs where motif features provide discriminative power. Empirically, we confirm that this augmentation yields consistent performance improvements across diverse tasks, including TGB link-property prediction (outperforming all five baselines), edge classification on Bitcoin Alpha, OTC, and MOOC datasets, and graph-level classification of synthetic temporal generators.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC