cartoboost 0.1.19

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 route membership, source-target 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 route cells, zones, markets, or H3 cells;
• source-target movement such as origin-to-destination 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.
• Neural embedding features for high-cardinality IDs.
• node2vec, GraphSAGE, heterogeneous GraphSAGE, and typed-schema HinSAGE graph feature encoders.

Install

Install the released package from PyPI:

uv add cartoboost

Optional integrations:

uv add "cartoboost[explain]"  # SHAP support
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 route cells, zones, markets, 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": "route_cells", "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={"route_cells": route_cells_train},
    feature_schema=schema,
)

predictions = model.predict(
    X_test_dense,
    sparse_sets={"route_cells": route_cells_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 Features

CartoBoost can precompute graph-derived columns before booster training. Supported encoder 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.

See Graph Features for encoder configs, directional features, OD-pair nodes, metapaths, artifacts, and benchmark guidance.

Neural Embedding Hybrid

Use NeuralEmbeddingRegressor when high-cardinality IDs carry stable signal and you want learned dense embeddings appended to the tree 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.

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={"route_cells": route_cells_test},
    background_sparse_sets={"route_cells": route_cells_train},
)

Model artifacts are versioned JSON and include optional metadata, feature schema, and training configuration fields. Graph and neural feature artifacts should be persisted alongside the booster when features are precomputed offline.

CLI

The CLI supports dense numeric CSV train, predict, eval, and inspect workflows. Use the Python API for list-valued sparse route-cell 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 Features
• Neural Features
• Evaluation Protocol
• Feature Schema
• Sparse Features
• Model Artifacts
• Python API Reference
• CLI Reference
• Benchmarks
• Limitations