Metric Learning on Temporal Graphs via Few-Shot Examples

Dongqi Fu, Liri Fang, Ross Maciejewski, Vetle I Torvik, Jingrui He

Keywords: Metric Learning, Few-Shot Learning, Temporal Graph

Abstract: Graph metric learning methods aim to learn the distance metric over graphs such that similar graphs are closer and dissimilar graphs are farther apart. This is of critical importance in many graph classification applications such as drug discovery and epidemics categorization. In many real-world applications, the graphs are typically evolving over time; labeling graph data is usually expensive and also requires background knowledge. However, state-of-the-art graph metric learning techniques consider the input graph as static, and largely ignore the intrinsic dynamics of temporal graphs; Furthermore, most of these techniques require abundant labeled examples for training in the representation learning process. To address the two aforementioned problems, we wish to learn a distance metric only over fewer temporal graphs, which metric could not only help accurately categorize seen temporal graphs but also be adapted smoothly to unseen temporal graphs. In this paper, we first propose the streaming-snapshot model to describe temporal graphs on different time scales. Then we propose the MetaTag framework: 1) to learn the metric over a limited number of streaming-snapshot modeled temporal graphs, 2) and adapt the learned metric to unseen temporal graphs via a few examples. Finally, we demonstrate the performance of MetaTag in comparison with state-of-the-art algorithms for temporal graph classification problems.

One-sentence Summary: The first attempt to learn temporal graph representations, on the graph-level, covering the whole lifetime, and only consuming a few labeled samples.