Comparative Study of TGN and TGAT for Dynamic Graph Learning : An Analysis of Performance and Efficiency in Dynamic Graph Learning Models

Sammanfattning

Dynamic graphs are increasingly vital for modeling evolving relationships in domains such as social networks and recommendation systems. Among leading models for continuous-time dynamic graph learning are Temporal Graph Networks (TGN) and Temporal Graph Attention Networks (TGAT), which leverage memory modules and attention mechanisms, respectively. This thesis presents a comprehensive empirical comparison of TGN and TGAT using the Temporal Graph Benchmark (TGB), focusing on tasks like link prediction and node affinity prediction across datasets of varying scales. We assess the models not only on predictive accuracy eusing metrics such as Average Precision, F1-Score, and NDCG@10, but also on computational efficiency, including GPU memory usage, power consumption, and training duration. Our findings indicate that TGN generally outperforms TGAT in scalability and resource efficiency, particularly on medium and large datasets where TGAT frequently encounters out-of-memory failures. While TGAT demonstrates competitive accuracy on smaller datasets and marginal gains in node affinity prediction under specific conditions, its practical deployment is hindered by high computational demands. Overall, the results suggest that TGN provides a more robust and scalable solution for dynamic graph learning in real-world applications. TGAT’s strengths in modeling temporal dependencies are evident but call for optimization to make the architecture more viable at scale. This study underscores the importance of holistic evaluation, balancing performance with feasibility when selecting models for dynamic graph representation learning.