Skip to main content
Open llms.txtCopy tools

CartoBoost PropagationDelayGraphForecaster

Use PropagationDelayGraphForecaster when directed graph edges carry known or estimated propagation delays. The model consumes node-time panels, directed edges, edge distances, known future covariates, and delay priors.

Python Example​

from cartoboost.deep import GraphTemporalFrame, PropagationDelayGraphForecaster

frame = GraphTemporalFrame(
    node_ids=["PULocationID:161", "PULocationID:236", "PULocationID:132"],
    timestamps=[0, 1, 2, 3, 4, 5],
    target=[[42, 35, 18], [44, 36, 19], [51, 40, 24], [58, 46, 31], [55, 45, 34], [49, 43, 30]],
    indptr=[0, 2, 3, 3],
    indices=[1, 2, 2],
    data=[0.7, 0.3, 1.0],
    edge_distances=[1.2, 8.4, 2.1],
    horizon=2,
    frequency="hourly",
)

model = PropagationDelayGraphForecaster(horizon=2, edge_delay_prior=[1, 2, 1])
model.fit(frame)
forecast = model.predict(2)

DelayAwareGraphTransformer and DynamicAdjacencyTransformer are aliases. SpatioTemporalGraphForecaster(backbone="delay_aware_graph_transformer") routes to the same implementation.

Browser WASM Example​

PropagationDelayGraphForecaster Wasm example

Runs the Rust-backed browser export on a substantial multi-route taxi panel with a held-out prediction window.

Ready to run in this page.

Use When​

Use this model when directed edges have meaningful propagation delays known before the forecast cutoff. Prefer a simpler graph model when delay priors are unavailable.

Validation​

Compare against non-graph temporal and static-adjacency graph baselines. Report edge-delay sensitivity and keep all graph inputs cutoff-safe.

Limitations​

  • Incorrect delays can underperform a static adjacency model.
  • Future-derived graph weights or delays leak holdout information.
  • Accelerator availability varies by build and must be reported.