cartoboost 0.1.29

Clean-room CartoBoost-inspired regression package.

CartoBoost

CartoBoost is a Python regression toolkit for temporal, spatial, geotemporal, and graph-derived prediction problems. It keeps the estimator workflow familiar to scikit-learn users while adding modeling primitives for place, time, sparse taxi-zone membership, pickup-dropoff directionality, and learned graph context.

Use CartoBoost when a standard tabular booster is a strong baseline, but your problem still requires hand-built features to represent:

• wraparound time such as hour-of-day, weekday, or seasonal cycles;
• 2D spatial boundaries, corridors, depots, hotspots, and service regions;
• list-valued memberships such as pickup zones, dropoff zones, or H3 cells;
• directed movement such as pickup-to-dropoff taxi flows;
• high-cardinality IDs that benefit from learned embeddings.

Core Capabilities

CartoBoost supports:

• L2 and quantile regression objectives.
• Constant and linear residual leaves.
• Axis, histogram-axis, diagonal 2D, Gaussian/radial 2D, periodic, sparse-set, and fuzzy split behavior.
• Dense numeric arrays plus list-valued sparse-set features.
• Feature schemas for numeric, periodic, sparse-set, and model-contract validation.
• JSON model artifacts and portable weights artifacts.
• Optional SHAP explanations, Optuna tuning, Polars input support, and ONNX export for the supported dense axis-tree subset.
• Standalone neural embedding regressors for high-cardinality IDs, plus optional neural feature-generation workflows.
• Standalone node2vec, GraphSAGE, heterogeneous GraphSAGE, and typed-schema HinSAGE graph regressors and link predictors, plus optional graph feature encoders.
• General Rust-backed utilities outside the forecasting API, including single-series forecast helpers, local-level/local-linear Kalman filters, Croston/SBA/TSB intermittent demand, and ordinary kriging.
• Rust-native forecasting APIs for geographic and temporal single-series or panel taxi demand, including rolling-origin backtests, naive/seasonal naive/theta/optimized-theta/ETS/AutoARIMA models, supervised CartoBoost lag forecasting, Rust-core weighted ensembles, CLI runs, and portable forecast artifacts.

Forecasting

from cartoboost.forecasting import ForecastFrame, ThetaForecaster

frame = ForecastFrame.from_pandas(
    taxi_lane_demand,
    timestamp_col="pickup_date",
    target_col="pickup_trips",
    series_id_col="pickup_dropoff_lane",
    freq="D",
)

model = ThetaForecaster(season_length=7)
model.fit(frame)
forecast = model.predict(horizon=14)

Forecast outputs use deterministic columns: series_id, timestamp, horizon, model, and mean. The Python forecasting surface is a wrapper over cartoboost._native; it does not compute fallback forecasts when a native binding is missing. Use the forecasting docs for geographic-temporal CLI usage, lag-feature forecasting, backtesting, artifacts, and examples built around pickup/dropoff lanes and taxi-zone demand.

Benchmarks

The benchmark reports focus on the data-science question behind each run: dataset, target, feature set, split design, comparison models, metrics, and the meaning of the result.

• NYC taxi benchmarks measure transformed trip duration, fare amount, and pickup-zone demand from pickup/dropoff zones, trip attributes, and hour/day features.
• Forecasting benchmarks measure daily pickup/dropoff lane demand with lagged demand, rolling summaries, calendar fields, zone IDs, airport-lane flags, and borough context. External forecasting comparisons name the exact libraries: StatsForecast and functime.
• Synthetic model-suite benchmarks isolate dense numeric signal, repeated-ID residual signal, and source-target graph signal against LightGBM and XGBoost.
• Taxi-zone acceptance benchmarks check whether lane membership, route geometry, and periodic hour behavior are recoverable before making broader quality claims.

Install

Install the released package from PyPI:

uv add cartoboost

Optional integrations:

uv add "cartoboost[explain]"  # SHAP support
uv add "cartoboost[h3]"       # H3 lat/lon encoder
uv add "cartoboost[s2]"       # S2 lat/lon encoder
uv add "cartoboost[duckdb]"   # DuckDB relation inputs
uv add "cartoboost[optuna]"   # Optuna tuning
uv add "cartoboost[polars]"   # Polars inputs
uv add "cartoboost[onnx]"     # ONNX export subset

Verify the install:

python -c "import cartoboost; print(cartoboost.__version__)"
cartoboost --help

Basic Regression

from cartoboost import CartoBoostRegressor

model = CartoBoostRegressor(
    n_estimators=100,
    learning_rate=0.05,
    max_depth=4,
    min_samples_leaf=20,
    splitters=["axis"],
)

model.fit(X_train, y_train)
predictions = model.predict(X_test)

The estimator supports sklearn-style get_params, set_params, clone, Pipeline, GridSearchCV, and NumPy-array predictions.

Temporal-Spatial Modeling

Use dense columns for numeric location and time features, and sparse-set columns for memberships such as pickup zones, dropoff zones, or encoded H3 cells.

from cartoboost import CartoBoostRegressor

schema = {
    "dense": [
        {"name": "pickup_x", "kind": "numeric"},
        {"name": "pickup_y", "kind": "numeric"},
        {"name": "hour_of_day", "kind": "periodic", "period": 24},
        {"name": "trip_distance", "kind": "numeric"},
    ],
    "sparse_sets": [
        {"name": "taxi_zones", "kind": "sparse_set"},
    ],
}

model = CartoBoostRegressor(
    n_estimators=200,
    learning_rate=0.04,
    max_depth=5,
    min_samples_leaf=30,
    splitters=["axis", "diagonal_2d", "gaussian_2d", "periodic:24", "sparse_set"],
    fuzzy=True,
    fuzzy_bandwidth=0.05,
)

model.fit(
    X_train_dense,
    y_train,
    sparse_sets={"taxi_zones": taxi_zones_train},
    feature_schema=schema,
)

predictions = model.predict(
    X_test_dense,
    sparse_sets={"taxi_zones": taxi_zones_test},
)

Why this helps:

• periodic:24 treats midnight-adjacent hours as neighbors.
• diagonal_2d learns oblique spatial boundaries more directly than axis-only trees.
• gaussian_2d isolates radial neighborhoods around local hotspots.
• sparse_set splits on list-valued route or cell membership without a wide one-hot matrix.
• fuzzy=True reduces hard jumps near spatial or temporal boundaries.

Graph Models

Graph models run independently through Node2VecStandaloneRegressor, GraphSageStandaloneRegressor, HeteroGraphSageStandaloneRegressor, HinSageStandaloneRegressor, and the matching standalone link predictors. Supported graph families are node2vec, GraphSAGE, HeteroGraphSAGE, and HinSAGE. Direction is a first-class contract: A -> B and B -> A can be separate facts, features, and embeddings.

Graph encoders can also emit graph-derived columns for another estimator when you explicitly want a feature-generation workflow.

See Graph Models And Features for standalone graph regressors, standalone link predictors, encoder configs, directional features, OD-pair nodes, metapaths, artifacts, and benchmark guidance.

Neural Embedding Models

Use NeuralEmbeddingStandaloneRegressor for direct supervised ID modeling without a boosted wrapper.

from cartoboost import NeuralEmbeddingStandaloneRegressor

model = NeuralEmbeddingStandaloneRegressor(dim=16, random_state=7)
model.fit(pickup_zone_ids_train, y_train, dense=X_train)
predictions = model.predict(pickup_zone_ids_test, dense=X_test)

Use NeuralEmbeddingRegressor when high-cardinality IDs carry stable signal and you explicitly want learned dense embeddings appended to a tabular model input.

from cartoboost import NeuralEmbeddingRegressor

model = NeuralEmbeddingRegressor(
    dim=16,
    use_residual=True,
    base_model_kwargs={"n_estimators": 80, "splitters": ["axis"]},
    final_model_kwargs={"n_estimators": 120, "splitters": ["axis", "periodic:24"]},
)

model.fit(X_train, y_train, ids=ids_train)
predictions = model.predict(X_test, ids=ids_test)

For a quick head-to-head comparison on one split:

from cartoboost import benchmark_neural_vs_cartoboost

results = benchmark_neural_vs_cartoboost(X, y, ids, split_ratio=0.8)

Use this helper as an initial signal check, then validate with your real temporal, spatial, grouped, or out-of-time split.

See Neural Embedding Models And Features for the standalone neural API, artifacts, fallback behavior, and optional feature generation workflow.

Save, Load, And Explain

model.save("model.cartoboost.json")
loaded = CartoBoostRegressor.load("model.cartoboost.json")

explanation = loaded.explain_shap(
    X_test_dense,
    background=X_train_dense,
    sparse_sets={"taxi_zones": taxi_zones_test},
    background_sparse_sets={"taxi_zones": taxi_zones_train},
)

Model artifacts are versioned JSON and include optional metadata, feature schema, and training configuration fields. Graph and neural standalone artifacts are complete model artifacts. Feature-generation artifacts should be persisted with whichever downstream model consumes their generated columns.

CLI

The CLI supports dense numeric CSV train, predict, eval, and inspect workflows. Use the Python API for list-valued sparse taxi-zone features and graph-derived feature pipelines.

cartoboost train --data train.csv --config configs/regression.toml --model-out model.json
cartoboost predict --model model.json --input test.csv --predictions-out predictions.csv
cartoboost eval --model model.json --data test_with_target.csv

Documentation

• Documentation Home
• Installation
• Getting Started
• Python Estimator
• Parameters
• Spatial Modeling
• Graph Models And Features
• Neural Embedding Models And Features
• Evaluation Protocol
• Feature Schema
• Sparse Features
• Model Artifacts
• Python API Reference
• CLI Reference
• Benchmarks
• Limitations