scikit-learn · IvanLauLinTiong · Sep 3, 2022 · Sep 3, 2022 · Sep 3, 2022 · Sep 3, 2022
diff --git a/doc/whats_new/v1.2.rst b/doc/whats_new/v1.2.rst
@@ -196,6 +196,9 @@ Changelog
   :class:`cluster.AgglomerativeClustering` and will be renamed to `metric` in v1.4.
   :pr:`23470` by :user:`Meekail Zain <micky774>`.
 
+- |Enhancement| :class:`cluster.FeatureAgglomeration` preserves dtype for
+  `numpy.float32`. :pr:`24346` by :user:`LinTiong Lau <IvanLauLinTiong>`.
+
 - |Fix| :class:`cluster.KMeans` now supports readonly attributes when predicting.
   :pr:`24258` by `Thomas Fan`_
 

diff --git a/sklearn/cluster/_agglomerative.py b/sklearn/cluster/_agglomerative.py
@@ -1322,7 +1322,9 @@ def fit(self, X, y=None):
             Returns the transformer.
         """
         self._validate_params()
-        X = self._validate_data(X, ensure_min_features=2)
+        X = self._validate_data(
+            X, ensure_min_features=2, dtype=[np.float64, np.float32]
+        )
         super()._fit(X.T)
         self._n_features_out = self.n_clusters_
         return self
@@ -1331,3 +1333,6 @@ def fit(self, X, y=None):
     def fit_predict(self):
         """Fit and return the result of each sample's clustering assignment."""
         raise AttributeError
+
+    def _more_tags(self):
+        return {"preserves_dtype": [np.float64, np.float32]}
diff --git a/sklearn/cluster/_feature_agglomeration.py b/sklearn/cluster/_feature_agglomeration.py
@@ -38,20 +38,21 @@ def transform(self, X):
         """
         check_is_fitted(self)
 
-        X = self._validate_data(X, reset=False)
+        X = self._validate_data(X, reset=False, dtype=[np.float64, np.float32])
         if self.pooling_func == np.mean and not issparse(X):
             size = np.bincount(self.labels_)
             n_samples = X.shape[0]
             # a fast way to compute the mean of grouped features
             nX = np.array(
-                [np.bincount(self.labels_, X[i, :]) / size for i in range(n_samples)]
+                [np.bincount(self.labels_, X[i, :]) / size for i in range(n_samples)],
+                dtype=X.dtype,
             )
         else:
             nX = [
                 self.pooling_func(X[:, self.labels_ == l], axis=1)
                 for l in np.unique(self.labels_)
             ]
-            nX = np.array(nX).T
+            nX = np.array(nX, dtype=X.dtype).T
         return nX
 
     def inverse_transform(self, Xred):

diff --git a/sklearn/cluster/tests/test_feature_agglomeration.py b/sklearn/cluster/tests/test_feature_agglomeration.py
@@ -6,7 +6,7 @@
 
 from numpy.testing import assert_array_equal
 from sklearn.cluster import FeatureAgglomeration
-from sklearn.utils._testing import assert_array_almost_equal
+from sklearn.utils._testing import assert_allclose, assert_array_almost_equal
 from sklearn.datasets import make_blobs
 
 
@@ -53,3 +53,18 @@ def test_feature_agglomeration_feature_names_out():
     assert_array_equal(
         [f"featureagglomeration{i}" for i in range(n_clusters)], names_out
     )
+
+
+def test_feature_agglomeration_numerical_consistency(global_random_seed):
+    """Ensure numerical consistency among np.float32 and np.float64"""
+    rng = np.random.RandomState(global_random_seed)
+    X_64, _ = make_blobs(n_features=12, random_state=rng)
+    X_32 = X_64.astype(np.float32)
+
+    agglo_32 = FeatureAgglomeration(n_clusters=3)
+    agglo_64 = FeatureAgglomeration(n_clusters=3)
+
+    X_trans_64 = agglo_64.fit_transform(X_64)
+    X_trans_32 = agglo_32.fit_transform(X_32)
+
+    assert_allclose(X_trans_32, X_trans_64)