Merge pull request #83 from MattScicluna/use_distance_in_graph

Use distance in graph

Author: bjoaofelipe

graphtools/graphs.py

@@ -1205,11 +1205,11 @@ def build_landmark_op(self):
                     self.data if not hasattr(self, "data_nu") else self.data_nu
                 )  # because of the scaling to review
                 if (
-                    n_samples > 5000
+                    n_samples > 5000 and self.distance == "euclidean"
                 ):  # sklearn.euclidean_distances is faster than cdist for big dataset
                     distances = euclidean_distances(data, data[landmark_indices])
                 else:
-                    distances = cdist(data, data[landmark_indices], metric="euclidean")
+                    distances = cdist(data, data[landmark_indices], metric=self.distance)
                 self._clusters = np.argmin(distances, axis=1)
 
             else:

test/test_landmark.py

@@ -284,29 +284,34 @@ def test_landmark_with_non_euclidean_distances():
     #     "but all distance metrics gave identical results"
     # )
 
-    # Test that the landmark operators are different shapes/values when different distances
-    # are used (this is a more sensitive test than just cluster assignments)
-    euclidean_landmark_sum = np.sum(euclidean_G.landmark_op)
-    manhattan_landmark_sum = np.sum(G_manhattan.landmark_op)
-    cosine_landmark_sum = np.sum(G_cosine.landmark_op)
+    # Compare landmark operators using Frobenius norm
+    euclidean_landmark_op = euclidean_G.landmark_op
+    manhattan_landmark_op = G_manhattan.landmark_op
+    cosine_landmark_op = G_cosine.landmark_op
+
+    diff_euclidean_manhattan = np.linalg.norm(
+        euclidean_landmark_op - manhattan_landmark_op, "fro"
+    )
+    diff_euclidean_cosine = np.linalg.norm(
+        euclidean_landmark_op - cosine_landmark_op, "fro"
+    )
 
     print(
-        f"Landmark operator sums: euclidean={euclidean_landmark_sum:.6f}, "
-        f"manhattan={manhattan_landmark_sum:.6f}, cosine={cosine_landmark_sum:.6f}"
+        f"Landmark operator differences: "
+        f"euclidean vs manhattan={diff_euclidean_manhattan:.6f}, "
+        f"euclidean vs cosine={diff_euclidean_cosine:.6f}"
     )
 
-    # The landmark operators should be different when using different distance metrics
     operators_different = (
-        abs(euclidean_landmark_sum - manhattan_landmark_sum) > 1e-10
-        or abs(euclidean_landmark_sum - cosine_landmark_sum) > 1e-10
+        diff_euclidean_manhattan > 1e-6 or diff_euclidean_cosine > 1e-6
     )
 
     if not operators_different:
         import warnings
 
         warnings.warn(
             "Landmark operators are identical across different distance metrics. "
-            "This strongly suggests the distance parameter is being ignored in build_landmark_op.",
+            "This suggests the distance parameter is being ignored in build_landmark_op.",
             UserWarning,
         )