FIX attribute error is BIRCH (#23395)

jeremiedbb · glemaitre · commit ecbe2d7a0522 · 2022-08-05T14:58:55.000+02:00
diff --git a/doc/whats_new/v1.1.rst b/doc/whats_new/v1.1.rst
@@ -2,6 +2,23 @@
 
 .. currentmodule:: sklearn
 
+.. _changes_1_1_2:
+
+Version 1.1.2
+=============
+
+**In Development**
+
+Changelog
+---------
+
+:mod:`sklearn.cluster`
+......................
+
+- |Fix| Fixed a bug in :class:`cluster.Birch` that could trigger an error when splitting
+  a node if there are duplicates in the dataset.
+  :pr:`23395` by :user:`Jérémie du Boisberranger <jeremiedbb>`.
+
 .. _changes_1_1_1:
 
 Version 1.1.1
diff --git a/sklearn/cluster/_birch.py b/sklearn/cluster/_birch.py
@@ -89,6 +89,11 @@ def _split_node(node, threshold, branching_factor):
     node1_dist, node2_dist = dist[(farthest_idx,)]
 
     node1_closer = node1_dist < node2_dist
+    # make sure node1 is closest to itself even if all distances are equal.
+    # This can only happen when all node.centroids_ are duplicates leading to all
+    # distances between centroids being zero.
+    node1_closer[farthest_idx[0]] = True
+
     for idx, subcluster in enumerate(node.subclusters_):
         if node1_closer[idx]:
             new_node1.append_subcluster(subcluster)
diff --git a/sklearn/cluster/tests/test_birch.py b/sklearn/cluster/tests/test_birch.py
@@ -228,3 +228,56 @@ def test_feature_names_out():
 
     names_out = brc.get_feature_names_out()
     assert_array_equal([f"birch{i}" for i in range(n_clusters)], names_out)
+
+
+def test_transform_match_across_dtypes():
+    X, _ = make_blobs(n_samples=80, n_features=4, random_state=0)
+    brc = Birch(n_clusters=4)
+    Y_64 = brc.fit_transform(X)
+    Y_32 = brc.fit_transform(X.astype(np.float32))
+
+    assert_allclose(Y_64, Y_32, atol=1e-6)
+
+
+def test_subcluster_dtype(global_dtype):
+    X = make_blobs(n_samples=80, n_features=4, random_state=0)[0].astype(
+        global_dtype, copy=False
+    )
+    brc = Birch(n_clusters=4)
+    assert brc.fit(X).subcluster_centers_.dtype == global_dtype
+
+
+def test_both_subclusters_updated():
+    """Check that both subclusters are updated when a node a split, even when there are
+    duplicated data points. Non-regression test for #23269.
+    """
+
+    X = np.array(
+        [
+            [-2.6192791, -1.5053215],
+            [-2.9993038, -1.6863596],
+            [-2.3724914, -1.3438171],
+            [-2.336792, -1.3417323],
+            [-2.4089134, -1.3290224],
+            [-2.3724914, -1.3438171],
+            [-3.364009, -1.8846745],
+            [-2.3724914, -1.3438171],
+            [-2.617677, -1.5003285],
+            [-2.2960556, -1.3260119],
+            [-2.3724914, -1.3438171],
+            [-2.5459878, -1.4533926],
+            [-2.25979, -1.3003055],
+            [-2.4089134, -1.3290224],
+            [-2.3724914, -1.3438171],
+            [-2.4089134, -1.3290224],
+            [-2.5459878, -1.4533926],
+            [-2.3724914, -1.3438171],
+            [-2.9720619, -1.7058647],
+            [-2.336792, -1.3417323],
+            [-2.3724914, -1.3438171],
+        ],
+        dtype=np.float32,
+    )
+
+    # no error
+    Birch(branching_factor=5, threshold=1e-5, n_clusters=None).fit(X)
