A temporal adjacency matrix combines the graphs from multiple timesteps into one larger graph. Each timestep becomes a layer: intra-timestep edges preserve connectivity within a frame, while inter-timestep edges connect corresponding nodes across adjacent frames. section 6.3.1 section 6.5.2
A practical construction follows this pipeline:
- Repeat the per-timestep edge list across the prediction window. section 6.3.1
- Mask nodes in the final timestep so a graph neural network can infer their unknown features from earlier, unmasked nodes. section 6.3.1
Pose prediction
A pose dataset combines location and velocity into node features, repeats the pose edges for every timestep, and connects corresponding nodes across adjacent layers. Masked nodes in the final layer become prediction targets for a graph attention network. section 6.3.1
The representation is useful for small graphs and short forecasting horizons, but it can become very large. Temporal embeddings offer an alternative, though they may require more complicated components and training. section 6.5.2
