Restrict HDBSCAN metric options to L2 #5415

python/cuml/cluster/hdbscan/hdbscan.pyx

@@ -165,12 +165,8 @@ cdef extern from "cuml/cluster/hdbscan.hpp" namespace "ML::HDBSCAN::HELPER":
                                    float cluster_selection_epsilon) except +
 
 _metrics_mapping = {
-    'l1': DistanceType.L1,
-    'cityblock': DistanceType.L1,
-    'manhattan': DistanceType.L1,
     'l2': DistanceType.L2SqrtExpanded,
     'euclidean': DistanceType.L2SqrtExpanded,
-    'cosine': DistanceType.CosineExpanded
 }
 
 
@@ -838,7 +834,7 @@ class HDBSCAN(UniversalBase, ClusterMixin, CMajorInputTagMixin):
         if self.metric in _metrics_mapping:
             metric = _metrics_mapping[self.metric]
         else:
-            raise ValueError("'affinity' %s not supported." % self.affinity)
+            raise ValueError(f"metric '{self.metric}' not supported, only 'l2' and 'euclidean' are currently supported")
 
         cdef uintptr_t core_dists_ptr = self.core_dists.ptr
 

python/cuml/tests/test_hdbscan.py

@@ -475,6 +475,25 @@ def test_hdbscan_core_dists_bug_4054():
     assert adjusted_rand_score(cu_labels_, sk_labels_) > 0.99
 
 
+@pytest.mark.parametrize(
+    "metric, supported",
+    [("euclidean", True), ("l1", False), ("l2", True), ("abc", False)],
+)
+def test_hdbscan_metric_parameter_input(metric, supported):
+    """
+    tests how valid and invalid arguments to the metric
+    parameter are handled
+    """
+    X, y = make_blobs(n_samples=10000, n_features=15, random_state=12)
+
+    clf = HDBSCAN(metric=metric)
+    if supported:
+        clf.fit(X)
+    else:
+        with pytest.raises(ValueError):
+            clf.fit(X)
+
+
 def test_hdbscan_empty_cluster_tree():
 
     raw_tree = np.recarray(