Fix an error in ProbWeightRegression when cvxpy fails to converge (#768)

* fix: raise clear error in `ProbWeightRegression` when cvxpy fails to solve

When input features have very large values, cvxpy may fail to find a
solution, leaving coefficients as None. This caused a confusing error
during predict (matrix multiplication with None). Now a ValueError is
raised at fit time with the solver status and a suggestion to apply
feature scaling.

* Update sklego/linear_model.py

Co-authored-by: Francesco Bruzzesi <42817048+FBruzzesi@users.noreply.github.com>

---------

Co-authored-by: Francesco Bruzzesi <42817048+FBruzzesi@users.noreply.github.com>

Author: RobKuebler

sklego/linear_model.py

@@ -245,6 +245,13 @@ def fit(self, X, y):
         # Solve the problem.
         prob = cp.Problem(objective, constraints)
         prob.solve()
+
+        if prob.status != "optimal":
+            raise ValueError(
+                f"cvxpy could not find a solution (status: {prob.status}).\n"
+                "Consider feature scaling (e.g. StandardScaler) before fitting the model."
+            )
+
         self.coef_ = betas.value
         self.n_features_in_ = X.shape[1]
 

tests/test_estimators/test_probweight_regression.py

@@ -17,3 +17,14 @@ def test_shape_trained_model(random_xy_dataset_regr):
     mod_no_intercept = ProbWeightRegression()
     assert mod_no_intercept.fit(X, y).coef_.shape == (X.shape[1],)
     np.testing.assert_approx_equal(mod_no_intercept.fit(X, y).coef_.sum(), 1.0, significant=4)
+
+
+def test_raises_on_unsolvable_problem():
+    """Test that a clear error is raised when cvxpy cannot find a solution."""
+    np.random.seed(42)
+    X = np.random.randn(10, 5) * 1e15
+    y = X @ np.array([2, -1, 3, 0.5, -2]) + np.random.randn(10) * 100
+
+    model = ProbWeightRegression()
+    with pytest.raises(ValueError, match="cvxpy could not find a solution"):
+        model.fit(X, y)