OmniST
A topology-invariant spatiotemporal forecasting model whose computation does not depend on the number of sensor nodes. Combined with inductive split conformal calibration, it provides prediction intervals with coverage guarantees for arbitrary locations, including unseen sensors.
Timeline (last 90 days)
OmniST's computation has zero dependence on the number of sensor nodes, and the model weight shapes do not change with the number of nodes N.
OmniST uses tokenization based on geographic coordinates to generate sensor representations, rather than fixed vocabulary indices.
OmniST dynamically constructs k-nearest-neighbor adjacency based on latitude and longitude during each forward pass, rather than relying on a graph structure fixed during training.
OmniST uses a bidirectional GRU to independently model the time series of each sensor.
Inductive split conformal calibration provides finite-sample coverage guarantees for OmniST's prediction intervals, and these guarantees still hold when sensors are added or removed.
On real air quality data from Delhi (37 stations), OmniST's PM2.5 prediction shows only a small performance drop after removing a large number of stations.
In inductive prediction on unseen stations, OmniST outperforms nearest-neighbor interpolation and also outperforms the learned kriging GNN (IGNNK), and only OmniST carries coverage guarantees.
OmniST's coverage guarantee remains at a high level under random node dropping across four real datasets and three domains (Delhi and Beijing air quality, NOAA GHCN weather, Caltrans PeMSD7-M traffic); it also holds at the full sensor scale of PeMS-BAY.
OmniST's coordinate tokenization and per-node GRU have parallel efficiency issues, and the k-nearest-neighbor attention incurs computational overhead at large node scales, which are issues that need to be weighed during engineering deployment.
The largest scale covered by the paper so far is the full sensor set of PeMS-BAY, and OmniST's performance at larger scales remains to be seen.
All Facts (10)
OmniST's computation has zero dependence on the number of sensor nodes, and the model weight shapes do not change with the number of nodes N.
50%UnverifiedOmniST uses tokenization based on geographic coordinates to generate sensor representations, rather than fixed vocabulary indices.
50%UnverifiedOmniST dynamically constructs k-nearest-neighbor adjacency based on latitude and longitude during each forward pass, rather than relying on a graph structure fixed during training.
50%UnverifiedOmniST uses a bidirectional GRU to independently model the time series of each sensor.
50%UnverifiedInductive split conformal calibration provides finite-sample coverage guarantees for OmniST's prediction intervals, and these guarantees still hold when sensors are added or removed.
50%UnverifiedOn real air quality data from Delhi (37 stations), OmniST's PM2.5 prediction shows only a small performance drop after removing a large number of stations.
50%UnverifiedIn inductive prediction on unseen stations, OmniST outperforms nearest-neighbor interpolation and also outperforms the learned kriging GNN (IGNNK), and only OmniST carries coverage guarantees.
50%UnverifiedOmniST's coverage guarantee remains at a high level under random node dropping across four real datasets and three domains (Delhi and Beijing air quality, NOAA GHCN weather, Caltrans PeMSD7-M traffic); it also holds at the full sensor scale of PeMS-BAY.
50%UnverifiedOmniST's coordinate tokenization and per-node GRU have parallel efficiency issues, and the k-nearest-neighbor attention incurs computational overhead at large node scales, which are issues that need to be weighed during engineering deployment.
50%UnverifiedThe largest scale covered by the paper so far is the full sensor set of PeMS-BAY, and OmniST's performance at larger scales remains to be seen.
50%