crospint: Coordinate Rotation for Spatial Interpolation

A scikit-learn-compatible pipeline for spatial and spatio-temporal interpolation using coordinate rotation and ensemble methods.

Overview

Tree-based models (LightGBM, Random Forest) split data along axis-aligned boundaries, which limits their ability to capture spatial patterns that don't align with the coordinate axes. crospint solves this by augmenting the feature set with rotated copies of the geographic coordinates, allowing the model to perform oblique spatial splits.

The package offers full compatibility with scikit-learn. It accepts Polars DataFrames as input, while maintaining full compatibility with ML libraries that require Pandas DataFrames as input. It also provides TwoStepsModel, a housing price estimator that handles log-transformation, price-per-square-meter conversion, retransformation bias correction, and iterative calibration.

Citation

If you use this package in research, I kindly request that you cite it.

@misc{Meslin2025,
  author  = {Meslin, Olivier},
  title   = {crospint: a Python scikit-learn-compatible pipeline for spatial and spatio-temporal interpolation using coordinate rotation and ensemble methods},
  year    = {2025},
  publisher = {GitHub},
  url     = {https://github.com/oliviermeslin/crospint},
}

## Installation

Install with `pip`:

```bash
pip install crospint

or with uv:

uv add crospint

Dependencies: polars, pandas, pyarrow, scipy, scikit-learn, lightgbm, matplotlib.

Quick start

import numpy as np
import polars as pl
from datetime import date, timedelta
from lightgbm import LGBMRegressor
from crospint import create_model_pipeline

# Generate synthetic data
np.random.seed(42)
n = 500
x = np.random.uniform(100_000, 900_000, n)
y = np.random.uniform(6_000_000, 7_100_000, n)
floor_area = np.random.uniform(20, 250, n).astype(int)
seashore_distance = np.random.uniform(0, 500_000, n)
start = date(2018, 1, 1)
days_range = (date(2023, 12, 31) - start).days
transaction_date = [start + timedelta(days=int(d)) for d in np.random.uniform(0, days_range, n)]
price_per_sqm = (2000 + 500 * np.sin(x / 200_000)
                 + 300 * np.cos(y / 300_000)
                 - 0.0005 * seashore_distance
                 + np.random.normal(0, 200, n))
transaction_amount = (floor_area * np.clip(price_per_sqm, 500, None)).round(0)

df = pl.DataFrame({
    "x": x, "y": y, "floor_area": floor_area,
    "transaction_date": transaction_date,
    "seashore_distance": seashore_distance,
    "transaction_amount": transaction_amount,
}).cast({"transaction_date": pl.Date})

# Create and configure the pipeline
pipe = create_model_pipeline(model=LGBMRegressor(n_estimators=100, verbose=-1))
pipe.set_params(
    coord_rotation__coordinates_names=("x", "y"),
    coord_rotation__number_axis=11,
    date_conversion__date_name="transaction_date",
)

# Train and predict
features = ["x", "y", "floor_area", "transaction_date", "seashore_distance"]
target = df["transaction_amount"].to_numpy()
pipe.fit(df[features], target)
predictions = pipe.predict(df[features])

Documentation

• User guide -- detailed walkthrough of all features: pipeline configuration, housing price modeling, calibration, outlier detection, and API reference.

Notebooks

Two notebooks illustrate the valuation method described in Meslin [2026]:

• How to apply the conditional tail removal procedure

• How to train price models at the national scale

License

MIT -- see LICENSE for details.