A temporal graph neural network models graphs whose node features, relationships, or both change across time. It is useful when predicting a future state depends on a sequence of earlier graph snapshots, such as human motion, customer purchases, or weather patterns. section 6.2 section 6.6
Represent each entity as a node and its relationships as edges, then preserve the time dimension instead of flattening all observations into one static graph. A temporal graph attention network (TGAT) treats edges as dynamic and attaches a timestamp to each edge. section 6.3.1 section 6.6
A practical spatiotemporal pipeline commonly follows these steps:
- Construct one graph layer per timestep by repeating and shifting the edge index, then reshape the result for graph processing. section 6.6
Pose prediction
In motion capture, 41 sensors represent joints from the toes to the neck. Body connections form intra-timestep edges, while corresponding nodes across adjacent timesteps capture temporal dependencies. The model predicts masked future joint features from recent motion. section 6.2
A naive static GAT can compress node features and temporal data into one graph, weakening temporal structure and producing inaccurate poses. Temporal encodings, timestamped edges, or recurrent processing provide alternatives; RNNs and GRUs can supply memory across snapshots. section 6.3.1 section 6.6
