Skip to main content
Open llms.txtCopy tools

CartoBoost InvertedTemporalTransformer

Use InvertedTemporalTransformer for wide synchronized panels where entities are the attention tokens. This avoids treating every time step as an attention token and reports cross-entity ablations so the graph-free entity interaction claim can be checked.

Python Example​

from cartoboost.deep import EntityPanelFrame, InvertedTemporalTransformer

frame = EntityPanelFrame(
    entity_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],
    ],
    horizon=2,
    frequency="hourly",
)

model = InvertedTemporalTransformer(lookback=4, horizon=2)
model.fit(frame)
forecast = model.predict(2)

InvertedEntityTransformer is an alias. The same implementation is reachable through TemporalEntityTransformer(architecture="inverted_transformer").

Browser WASM Example​

InvertedTemporalTransformer 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 for synchronized panels where cross-entity dependence matters and entities are more useful as attention tokens than individual timestamps.

Validation​

Report horizon-wise error and an ablation that removes cross-entity attention. Use this only when synchronized entities are the modeling unit.

Limitations​

  • Requires aligned histories and enough repeated time windows.
  • Cross-entity attention can overfit stable identity effects.
  • Report seed sensitivity and cold-entity behavior.