Merge branch 'master' into periodic

pygam/terms.py

@@ -795,7 +795,7 @@ def build_columns(self, X, verbose=False):
 
 
 class FactorTerm(SplineTerm):
-    _encodings = ['one-hot']
+    _encodings = ['one-hot', 'dummy']
     def __init__(self, feature, lam=0.6, penalties='auto', coding='one-hot', verbose=False):
         """creates an instance of a FactorTerm
 
@@ -900,6 +900,32 @@ def compile(self, X, verbose=False):
                                           verbose=verbose)
         return self
 
+    def build_columns(self, X, verbose=False):
+        """construct the model matrix columns for the term
+
+        Parameters
+        ----------
+        X : array-like
+            Input dataset with n rows
+
+        verbose : bool
+            whether to show warnings
+
+        Returns
+        -------
+        scipy sparse array with n rows
+        """
+        columns = super(FactorTerm, self).build_columns(X, verbose=verbose)
+        if self.coding == 'dummy':
+            columns = columns[:, 1:]
+
+        return columns
+
+    @property
+    def n_coefs(self):
+        """Number of coefficients contributed by the term to the model
+        """
+        return self.n_splines - 1 * (self.coding in ['dummy'])
 
 class MetaTermMixin(object):
     _plural = [

pygam/tests/test_terms.py

@@ -191,6 +191,18 @@ def test_correct_smoothing_in_tensors(toy_interaction_X_y):
     gam = LinearGAM(te(0, 1, lam=[10000, 0.6])).fit(X, y)
     assert gam.statistics_['pseudo_r2']['explained_deviance'] < 0.1
 
+def test_dummy_encoding(wage_X_y, wage_gam):
+    """check that dummy encoding produces fewer coefficients than one-hot"""
+    X, y = wage_X_y
+
+    gam = LinearGAM(s(0) + s(1) + f(2, coding='dummy')).fit(X, y)
+
+    assert gam._modelmat(X=X, term=2).shape[1] == 4
+    assert gam.terms[2].n_coefs == 4
+
+    assert wage_gam._modelmat(X=X, term=2).shape[1] == 5
+    assert wage_gam.terms[2].n_coefs == 5
+
 def test_build_cyclic_p_spline(hepatitis_X_y):
     """check the cyclic p spline builds
 
@@ -247,7 +259,6 @@ def test_cyclic_p_spline_custom_period():
     assert np.allclose(gam.predict(X), 0.5)
     assert np.allclose(gam.edge_knots_[0], [0, 0.5])
 
-
 def test_tensor_terms_have_constraints(toy_interaction_X_y):
     """test that we can fit a gam with constrained tensor terms,
     even if those constraints are 'none'