Fixed all_points_core_distance for clusters with pairwise distances o…

…f zero Previously, all_points_core_distance would return NaN if all points within a cluster had a distance of zero to one another, i.e. were all the same. Now, it returns an array of zeros with length(distance_matrix).
scikit-learn-contrib · Aug 23, 2022 · f2e2af9 · f2e2af9
1 parent 523cd86
commit f2e2af9
Showing 1 changed file with 5 additions and 1 deletion.
diff --git a/hdbscan/validity.py b/hdbscan/validity.py
@@ -31,7 +31,11 @@ def all_points_core_distance(distance_matrix, d=2.0):
         distance_matrix != 0]) ** d
     result = distance_matrix.sum(axis=1)
     result /= distance_matrix.shape[0] - 1
-    result **= (-1.0 / d)
+
+    if result.sum() == 0:
+        result = np.zeros(len(distance_matrix))
+    else:
+        result **= (-1.0 / d)
 
     return result