Set boruvka leaf size as primary leaf size over 3 (heuristic).

Author: lmcinnes

hdbscan/_hdbscan_boruvka.c

@@ -159,10 +159,10 @@ cdef class KDTreeBoruvkaAlgorithm (object):
     cdef np.ndarray candidate_neighbor_arr
     cdef np.ndarray candidate_distance_arr
 
-    def __init__(self, tree, min_samples=5, metric='euclidean', **kwargs):
+    def __init__(self, tree, min_samples=5, metric='euclidean', leaf_size=20, **kwargs):
 
         self.core_dist_tree = tree
-        self.tree = KDTree(tree.data, metric=metric, leaf_size=10)
+        self.tree = KDTree(tree.data, metric=metric, leaf_size=leaf_size)
         self._data = np.array(self.tree.data)
         self._raw_data = self.tree.data
         self.node_bounds = self.tree.node_bounds
@@ -539,10 +539,10 @@ cdef class BallTreeBoruvkaAlgorithm (object):
     cdef np.ndarray candidate_neighbor_arr
     cdef np.ndarray candidate_distance_arr
 
-    def __init__(self, tree, min_samples=5, metric='euclidean', **kwargs):
+    def __init__(self, tree, min_samples=5, metric='euclidean', leaf_size=20, **kwargs):
 
         self.core_dist_tree = tree
-        self.tree = BallTree(tree.data, metric=metric, leaf_size=10)
+        self.tree = BallTree(tree.data, metric=metric, leaf_size=leaf_size)
         self._data = np.array(self.tree.data)
         self._raw_data = self.tree.data
         self.min_samples = min_samples

hdbscan/_hdbscan_boruvka.pyx

@@ -143,7 +143,7 @@ def _hdbscan_boruvka_kdtree(X, min_cluster_size=5, min_samples=None, alpha=1.0,
     min_samples = min(dim - 1, min_samples)
 
     tree = KDTree(X, metric=metric, leaf_size=leaf_size)
-    alg = KDTreeBoruvkaAlgorithm(tree, min_samples, metric=metric)
+    alg = KDTreeBoruvkaAlgorithm(tree, min_samples, metric=metric, leaf_size=leaf_size//3)
     min_spanning_tree = alg.spanning_tree()
 
     return _tree_to_labels(X, min_spanning_tree, min_cluster_size) + (min_spanning_tree,)
@@ -156,7 +156,7 @@ def _hdbscan_boruvka_balltree(X, min_cluster_size=5, min_samples=None, alpha=1.0
     min_samples = min(dim - 1, min_samples)
 
     tree = BallTree(X, metric=metric, leaf_size=leaf_size)
-    alg = BallTreeBoruvkaAlgorithm(tree, min_samples, metric=metric)
+    alg = BallTreeBoruvkaAlgorithm(tree, min_samples, metric=metric, leaf_size=leaf_size//3)
     min_spanning_tree = alg.spanning_tree()
 
     return _tree_to_labels(X, min_spanning_tree, min_cluster_size) + (min_spanning_tree,)