GeoShapley

A game theory approach to measuring spatial effects from machine learning models. GeoShapley is built on Shapley value and Kernel SHAP estimator.

Installation:

GeoShapley can be installed from PyPI:

$ pip install geoshapley

To install the latest version from Github:

$ pip install git+https://github.com/ziqi-li/geoshapley.git

Example:

GeoShapley can explain any model that takes tabular data + spatial features (e.g., coordinates) as the input. Examples of natively supported models include:

1. XGBoost/CatBoost/LightGBM
2. Random Forest
3. MLP or other scikit-learn modules.
4. TabNet
5. Explainable Boosting Machine
6. Statistical models: OLS/Gaussian Process/GWR

Other models can be supported by defining a helper function model.predict() to wrap around their original models' prediction or inference functions.

Currently, spatial features (e.g., coordinates) need to be put as the last columns of your pandas.DataFrame(X_geo).

Below shows an example on how to explain a trained MLP model. More examples can be found at the notebooks folder.

from geoshapley import GeoShapleyExplainer
from sklearn.neural_network import MLPRegressor
from sklearn.model_selection import train_test_split

X_train, X_test, y_train, y_test = train_test_split(X_geo, y, random_state=1)

#Fit a NN model based on training data
mlp_model = MLPRegressor().fit(X_train, y_train)

#Specify a small background data
background = X_train.sample(100).values

#Initilize a GeoShapleyExplainer
mlp_explainer = GeoShapleyExplainer(mlp_model.predict, background)

#Explain the data
mlp_rslt = mlp_explainer.explain(X_geo)

#Make a shap-style summary plot
mlp_rslt.summary_plot()

#Make partial dependence plots of the primary (non-spatial) effects
mlp_rslt.partial_dependence_plots()

#Calculate spatially varying explanations
mlp_svc = mlp_rslt.get_svc()

Reference:

Li, Z. (2024). GeoShapley: A Game Theory Approach to Measuring Spatial Effects in Machine Learning Models. Annals of the American Association of Geographers. Open access at: https://www.tandfonline.com/doi/full/10.1080/24694452.2024.2350982