Graph Neural Networks for temporal graphs: State of the art, open challenges, and opportunities

TL;DR

This work tackles learning on temporal graphs by organizing and formalizing TGNN methods. It provides a coherent formalization of learning settings and tasks, and introduces a taxonomy that separates snapshot-based from event-based approaches and, within those, model evolution versus embedding evolution, including temporal embedding and mailbox-based mechanisms. The paper also discusses open challenges—benchmarking, explainability, and expressivity—and argues for avenues beyond conventional GNNs in domains like climate science and epidemiology. By establishing a systematic framework, it aims to guide rigorous evaluation and future research in temporal graph representation learning.

Abstract

Graph Neural Networks (GNNs) have become the leading paradigm for learning on (static) graph-structured data. However, many real-world systems are dynamic in nature, since the graph and node/edge attributes change over time. In recent years, GNN-based models for temporal graphs have emerged as a promising area of research to extend the capabilities of GNNs. In this work, we provide the first comprehensive overview of the current state-of-the-art of temporal GNN, introducing a rigorous formalization of learning settings and tasks and a novel taxonomy categorizing existing approaches in terms of how the temporal aspect is represented and processed. We conclude the survey with a discussion of the most relevant open challenges for the field, from both research and application perspectives.

Figures (2)

• Figure 1: Learning settings. Schematic representation of the learning settings on TGs formalized in Section \ref{['subsec:indTran']}. The temporal graphs are represented as sequences of snapshots, with training (red) and inference (green) nodes connected by edges (solid lines), and where a dotted line connects instances of the same node (with possibly different features and/or labels) in successive snapshots. The four categories are obtained from the different combinations of a temporal and a topological dimension. The temporal dimension distinguishes the future setting, where the training nodes are all observed before the inference nodes (first row), from the past setting where inference is performed also on nodes appearing before the observation of the last training node (second row). The topological dimension comprises a transductive setting, where each inference node is observed (unlabelled) also during training (left column), and an inductive setting, where inference is performed on nodes that are unknown at training time (right column).
• Figure 2: The proposed TGNN taxonomy and an analysis of the surveyed methods. The top panel shows the new categories introduced in this work with the corresponding model instances (Section \ref{['sec:tax']}), where the colored bullets additionally indicate the main technology that they employ. The bottom table maps these methods to the task (Section \ref{['sec:task']}) to which they have been applied in the respective original paper, with an additional indication of their use in the future (F), past (P), inductive (I), or transductive (T) settings (Section \ref{['subsec:indTran']}). Notice that no method has been applied yet to clustering and visualization, for neither graphs nor nodes. Moreover, only four out of ten models have been tested in the past mode (three in PT, one in PI).

Theorems & Definitions (15)

• Definition 1: Static Graph - SG
• Definition 2: Temporal Graph - TG
• Definition 3: Discrete Time Temporal Graph - DTTG
• Definition 4: Snapshot-based Temporal Graph - STG
• Definition 5: Event-based Temporal Graph - ETG
• Definition 6: Learning settings
• Definition 7: Temporal Node Classification
• Definition 8: Temporal Edge Classification
• Definition 9: Temporal Graph Classification
• Definition 10: Temporal Link Prediction