BUG fix and test fobidden zero node split for Poisson

Christian Lorentzen · Christian Lorentzen · commit 6833457108f0 · 2020-05-29T18:43:30.000+02:00
diff --git a/sklearn/tree/_criterion.pyx b/sklearn/tree/_criterion.pyx
@@ -1399,7 +1399,7 @@ cdef class Poisson(RegressionCriterion):
                 # Poisson loss does not allow non-positive predictions. We
                 # therefore forbid splits that have child nodes withs
                 # sum(y_i) <= 0.
-                return INFINITY
+                return -INFINITY
             else:
                 proxy_impurity_left += y_mean_left * log(y_mean_left)
                 proxy_impurity_right += y_mean_right * log(y_mean_right)
diff --git a/sklearn/tree/tests/test_tree.py b/sklearn/tree/tests/test_tree.py
@@ -1957,3 +1957,13 @@ def test_balance_property(criterion, tree):
     reg = tree(criterion=criterion)
     reg.fit(X, y)
     assert np.sum(reg.predict(X)) == pytest.approx(np.sum(y))
+
+
+def test_poisson_zero_nodes():
+    """Test that sum(y)=0 and therefore y_pred=0 never happens on nodes."""
+    X = [[0, 0], [0, 1], [0, 2], [0, 3],
+         [1, 0], [1, 2], [1, 2], [1, 3]]
+    y = [0, 0, 0, 0, 1, 2, 3, 4]
+    reg = DecisionTreeRegressor(criterion="poisson", random_state=1)
+    reg.fit(X, y)
+    assert np.all(reg.predict(X) > 0)
