Temporal motif-based representation learning on continuous-time dynamic graphs

Marouane Alilou

doi:10.1007/s10618-026-01190-2

Return

Temporal motif-based representation learning on continuous-time dynamic graphs

2026-03-06

PRE

Marouane Alilou

Bikram Pratim Bhuyan *

Rachida Fissoune

Amar Ramdane-Chérif

DOI：10.1007/s10618-026-01190-2

Original

Original request for help

Save

Abstract

En 中文

Temporal graphs provide a powerful framework for modeling time-dependent interactions in dynamic systems such as social, biological, and communication graphs. Continuous-Time Dynamic Graphs (CTDGs) are commonly used to represent such graphs; however, existing approaches often fail to capture higher-order temporal motifs, limiting their effectiveness in downstream tasks. We propose a novel framework that combines temporally-biased random walks and motif-based incidence matrices to extract and encode multi-scale higher-order interaction patterns. By integrating these structural features with Node2Vec embeddings, we construct expressive node representations that jointly capture temporal and topological dynamics. Extensive experiments across five benchmark datasets demonstrate that our method consistently outperforms strong representative baselines under a unified inductive evaluation protocol, achieving improvements of up to 9.1% in AUC. Our framework is scalable, robust, and broadly applicable to real-world temporal graphs. All code and trained models are publicly released at: https://github.com/marwane-alilou/TemporalMotif-CTDG to foster reproducibility.8097

Keywords:

Temporal graphs

Higher-order temporal motifs

Graph representation learning

Biased random walks

Link prediction

Node classification