A Robust Comparative Analysis of Graph Neural Networks on Dynamic Link Prediction

Graph neural networks (GNNs) are rapidly becoming the dominant way to learn on graph-structured data. Link prediction is a near-universal benchmark for new GNN models. Many advanced models such as Dynamic graph neural networks (DGNNs) specifically target dynamic graphs. However, these models, particularly DGNNs, are rarely compared to each other or existing heuristics. Different works evaluate their models in different ways, thus one cannot compare evaluation metrics and their results directly. Motivated by this, we perform a comprehensive comparison study. We compare link prediction heuristics, GNNs, discrete DGNNs, and continuous DGNNs on the dynamic link prediction task. In total we summarize the results of over 3200 experimental runs (
$\approx 1.5$
years of computation time). We find that simple link prediction heuristics perform better than GNNs and DGNNs, different sliding window sizes greatly affect performance, and of all examined graph neural networks, that DGNNs consistently outperform static GNNs. This work is a continuation of our previous work, a foundation of dynamic networks and theoretical review of DGNNs. In combination with our survey, we provide both a theoretical and empirical comparison of DGNNs.