A New Graph Neural Network Beats Prior Dynamic Link Predictors by Up to 277%

When a network keeps changing, yesterday’s connections can miss today’s pattern. That matters for things like financial transactions, trust graphs, and social communication, where topology and node attributes evolve together. This paper introduces SiST-GNN, a dynamic graph neural network that handles time and structure in one message-passing step instead of forcing them into separate stages. The idea is to give each node a memory of its past, pair it with its current features, and let graph convolution update them together. In the paper’s tests, SiST-GNN sets a new state of the art on every public benchmark for link prediction. It outperforms the strongest prior method by 109–277% in the fixed-split setting and by 68–194% in the live-update setting. The authors also built three dynamic node-classification tasks from continuous-time event streams. On those, SiST-GNN beats the leading discrete-time baseline by 7–22% and matches continuous-time methods that read the raw events directly. The takeaway is simple: letting a graph model reason about a node’s history and its neighborhood at the same time can pay off a lot.

109–277% is a huge jump for a graph model. SiST-GNN reaches it on link prediction, the task of guessing the next tie in a network that keeps moving. That matters in financial transactions, peer-to-peer trust graphs, and social platforms, because the pattern is never just who is linked now. It is also how each node got here. When time and structure get handled in separate steps, the second step sees a squeezed summary, not the full story. SiST-GNN takes a different bet: let a node's memory and its current features meet inside the same update, so the model can react to both at once.

Why the usual two-step order breaks

Across nine public baselines and fourteen model-dataset combinations, SiST-GNN wins in both settings the field uses most: fixed-split and live-update. On link prediction, it sets a new high on every public benchmark, beating the strongest prior method by 109–277% in fixed-split tests and by 68–194% in live-update tests. The same design also reaches beyond link prediction. On three dynamic node-classification tasks built by discretising continuous-time event streams, SiST-GNN beats the leading discrete-time baseline by 7–22% and matches continuous-time methods that read raw events directly. That is the surprise: a snapshot-based model can keep up even when the rival starts from the fuller event log.

How SiST-GNN folds memory into the graph

SiST-GNN gives each node a small memory of its past at every snapshot. It then pairs that memory with the node's current feature vector and treats the pair as two linked versions of the same node, joined by a cross-time edge. A standard graph convolution runs on this augmented graph, so the update can mix history and neighbourhood before either signal gets flattened into a summary. That is the key move. The neighbour message no longer arrives blind, because the node's own past sits in the same update step and can shape what the model keeps.

Why that changes the game

Temporal-first models compress history first and only then look at the graph. Spatial-first models do the reverse. SiST-GNN keeps both views alive long enough to affect each other, which means the model does not have to choose between a node's trajectory and its neighbours. That matters when the past carries meaning that a frozen snapshot misses. In link prediction, a tie that looks random in the present can look patterned once you know where the node has been. The gains across both evaluation modes show that this is more than a neat idea; the joint update pays off in practice.

What to test next

The practical win is clear: discrete-time models no longer need to give up when raw event logs are not available. On the three discretised node-classification tasks, SiST-GNN matched continuous-time methods that read those events directly, which makes snapshot data much more useful than it used to be. That means a model built from snapshots can stay close to the best event-level systems on this class of tasks. The update rule itself is doing the heavy lifting, and that is the part worth carrying into the next changing graph.