ENH: allow custom Graph object

gwlearn/base.py

@@ -94,6 +94,7 @@ def __init__(
         ]
         | Callable = "bisquare",
         include_focal: bool = False,
+        graph: graph.Graph = None,
         n_jobs: int = -1,
         fit_global_model: bool = True,
         measure_performance: bool = True,
@@ -108,6 +109,7 @@ def __init__(
         self.bandwidth = bandwidth
         self.kernel = kernel
         self.include_focal = include_focal
+        self.graph = graph
         self.fixed = fixed
         self._model_kwargs = kwargs
         self.n_jobs = n_jobs
@@ -380,6 +382,10 @@ class BaseClassifier(_BaseModel, ClassifierMixin):
         futher spatial analysis of the model performance (and generalises to models
         that do not support OOB scoring). However, it leaves out the most representative
         sample. By default False
+    graph : Graph, optional
+        Custom libpysal.graph.Graph object encoding the spatial interaction between
+        observations. If given, it is used directly and `bandwidth`, `fixed`, `kernel`,
+        and `include_focal` keywords are ignored.
     n_jobs : int, optional
         The number of jobs to run in parallel. ``-1`` means using all processors
         by default ``-1``
@@ -471,6 +477,7 @@ def __init__(
         ]
         | Callable = "bisquare",
         include_focal: bool = False,
+        graph: graph.Graph = None,
         n_jobs: int = -1,
         fit_global_model: bool = True,
         measure_performance: bool = True,
@@ -490,6 +497,7 @@ def __init__(
             fixed=fixed,
             kernel=kernel,
             include_focal=include_focal,
+            graph=graph,
             n_jobs=n_jobs,
             fit_global_model=fit_global_model,
             measure_performance=measure_performance,
@@ -548,7 +556,11 @@ def _is_binary(series: pd.Series) -> bool:
 
         if self.verbose:
             print(f"{(time() - self._start):.2f}s: Building weights")
-        weights = self._build_weights(geometry)
+
+        if self.graph is not None:
+            weights = self.graph
+        else:
+            weights = self._build_weights(geometry)
         if self.verbose:
             print(f"{(time() - self._start):.2f}s: Weights ready")
         self._setup_model_storage()
@@ -877,6 +889,10 @@ class BaseRegressor(_BaseModel, RegressorMixin):
         futher spatial analysis of the model performance (and generalises to models
         that do not support OOB scoring). However, it leaves out the most representative
         sample. By default False
+    graph : Graph, optional
+        Custom libpysal.graph.Graph object encoding the spatial interaction between
+        observations. If given, it is used directly and `bandwidth`, `fixed`, `kernel`,
+        and `include_focal` keywords are ignored.
     n_jobs : int, optional
         The number of jobs to run in parallel. ``-1`` means using all processors
         by default ``-1``
@@ -925,7 +941,10 @@ def fit(
         """
         self._validate_geometry(geometry)
 
-        weights = self._build_weights(geometry)
+        if self.graph is not None:
+            weights = self.graph
+        else:
+            weights = self._build_weights(geometry)
         self._setup_model_storage()
 
         # fit the models

gwlearn/ensemble.py

@@ -6,6 +6,7 @@
 import geopandas as gpd
 import numpy as np
 import pandas as pd
+from libpysal import graph
 from sklearn import metrics
 from sklearn.ensemble import GradientBoostingClassifier, RandomForestClassifier
 
@@ -31,6 +32,10 @@ class GWRandomForestClassifier(BaseClassifier):
         analysis of the model performance (and generalises to models that do not support
         OOB scoring). However, it leaves out the most representative sample. By default
         False
+    graph : Graph, optional
+        Custom libpysal.graph.Graph object encoding the spatial interaction between
+        observations. If given, it is used directly and `bandwidth`, `fixed`, `kernel`,
+        and `include_focal` keywords are ignored.
     n_jobs : int, optional
         The number of jobs to run in parallel. ``-1`` means using all processors by
         default ``-1``
@@ -153,6 +158,7 @@ def __init__(
         ]
         | Callable = "bisquare",
         include_focal: bool = False,
+        graph: graph.Graph = None,
         n_jobs: int = -1,
         fit_global_model: bool = True,
         measure_performance: bool = True,
@@ -172,6 +178,7 @@ def __init__(
             fixed=fixed,
             kernel=kernel,
             include_focal=include_focal,
+            graph=graph,
             n_jobs=n_jobs,
             fit_global_model=fit_global_model,
             measure_performance=measure_performance,
@@ -303,6 +310,10 @@ class GWGradientBoostingClassifier(BaseClassifier):
         futher spatial analysis of the model performance (and generalises to models
         that do not support OOB scoring). However, it leaves out the most representative
         sample. By default False
+    graph : Graph, optional
+        Custom libpysal.graph.Graph object encoding the spatial interaction between
+        observations. If given, it is used directly and `bandwidth`, `fixed`, `kernel`,
+        and `include_focal` keywords are ignored.
     n_jobs : int, optional
         The number of jobs to run in parallel. ``-1`` means using all processors
         by default ``-1``
@@ -393,6 +404,7 @@ def __init__(
         ]
         | Callable = "bisquare",
         include_focal: bool = False,
+        graph: graph.Graph = None,
         n_jobs: int = -1,
         fit_global_model: bool = True,
         measure_performance: bool = True,
@@ -408,6 +420,7 @@ def __init__(
             fixed=fixed,
             kernel=kernel,
             include_focal=include_focal,
+            graph=graph,
             n_jobs=n_jobs,
             fit_global_model=fit_global_model,
             measure_performance=measure_performance,

gwlearn/linear_model.py

@@ -5,6 +5,7 @@
 import geopandas as gpd
 import numpy as np
 import pandas as pd
+from libpysal import graph
 from sklearn import metrics
 from sklearn.linear_model import LinearRegression, LogisticRegression
 
@@ -30,6 +31,10 @@ class GWLogisticRegression(BaseClassifier):
         futher spatial analysis of the model performance (and generalises to models
         that do not support OOB scoring). However, it leaves out the most representative
         sample. By default True
+    graph : Graph, optional
+        Custom libpysal.graph.Graph object encoding the spatial interaction between
+        observations. If given, it is used directly and `bandwidth`, `fixed`, `kernel`,
+        and `include_focal` keywords are ignored.
     n_jobs : int, optional
         The number of jobs to run in parallel. ``-1`` means using all processors
         by default ``-1``
@@ -158,6 +163,7 @@ def __init__(
         ]
         | Callable = "bisquare",
         include_focal: bool = True,
+        graph: graph.Graph = None,
         n_jobs: int = -1,
         fit_global_model: bool = True,
         measure_performance: bool = True,
@@ -174,6 +180,7 @@ def __init__(
             fixed=fixed,
             kernel=kernel,
             include_focal=include_focal,
+            graph=graph,
             n_jobs=n_jobs,
             fit_global_model=fit_global_model,
             measure_performance=measure_performance,
@@ -306,6 +313,10 @@ class GWLinearRegression(BaseRegressor):
         futher spatial analysis of the model performance (and generalises to models
         that do not support OOB scoring). However, it leaves out the most representative
         sample. By default True
+    graph : Graph, optional
+        Custom libpysal.graph.Graph object encoding the spatial interaction between
+        observations. If given, it is used directly and `bandwidth`, `fixed`, `kernel`,
+        and `include_focal` keywords are ignored.
     n_jobs : int, optional
         The number of jobs to run in parallel. ``-1`` means using all processors
         by default ``-1``
@@ -353,6 +364,7 @@ def __init__(
         ]
         | Callable = "bisquare",
         include_focal: bool = True,
+        graph: graph.Graph = None,
         n_jobs: int = -1,
         fit_global_model: bool = True,
         measure_performance: bool = True,
@@ -367,6 +379,7 @@ def __init__(
             fixed=fixed,
             kernel=kernel,
             include_focal=include_focal,
+            graph=graph,
             n_jobs=n_jobs,
             fit_global_model=fit_global_model,
             measure_performance=measure_performance,

gwlearn/tests/test_base.py

@@ -8,6 +8,7 @@
 import pandas as pd
 import pytest
 from geodatasets import get_path
+from libpysal.graph import Graph
 from sklearn.ensemble import RandomForestClassifier, RandomForestRegressor
 from sklearn.linear_model import LinearRegression, LogisticRegression
 
@@ -1440,3 +1441,49 @@ def test_regressor_fit_focal_inclusion(sample_regression_data):
 
     # RF should 'remember' focal point when included
     assert (y - no_focal.pred_).mean() > (y - focal.pred_).mean()
+
+
+def test_custom_graph_baseregressor(sample_regression_data):
+    """Test BaseRegressor with a custom graph object."""
+    X, y, geometry = sample_regression_data
+
+    # Create a fixed distance weights graph
+    g = Graph.build_distance_band(geometry, threshold=150000, binary=False)
+
+    # Create regressor with custom graph
+    reg = BaseRegressor(
+        LinearRegression,
+        bandwidth=100,  # This should be ignored when custom graph is provided
+        fixed=False,  # This should be ignored when custom graph is provided
+        graph=g,
+    )
+
+    # Fit the model
+    reg.fit(X, y, geometry)
+
+    # Check that the model was fit successfully
+    assert hasattr(reg, "pred_")
+    assert hasattr(reg, "local_r2_")
+
+
+def test_custom_graph_baseclassifier(sample_data):
+    """Test BaseClassifier with a custom graph object."""
+    X, y, geometry = sample_data
+
+    # Create a fixed distance weights graph
+    g = Graph.build_distance_band(geometry, threshold=150000, binary=False)
+    # Create classifier with custom graph
+    clf = BaseClassifier(
+        LogisticRegression,
+        bandwidth=100,  # This should be ignored when custom graph is provided
+        fixed=True,  # This should be ignored when custom graph is provided
+        graph=g,
+        max_iter=500,
+    )
+
+    # Fit the model
+    clf.fit(X, y, geometry)
+
+    # Check that the model was fit successfully
+    assert hasattr(clf, "proba_")
+    assert hasattr(clf, "score_")