scikit-learn
diff --git a/‎doc/whats_new/upcoming_changes/sklearn.tree/30557.fix.rst‎
Lines changed: 2 additions & 0 deletions b/‎doc/whats_new/upcoming_changes/sklearn.tree/30557.fix.rst‎
Lines changed: 2 additions & 0 deletions
diff --git a/‎sklearn/tree/_partitioner.pyx‎
Lines changed: 9 additions & 4 deletions b/‎sklearn/tree/_partitioner.pyx‎
Lines changed: 9 additions & 4 deletions
diff --git a/‎sklearn/tree/tests/test_tree.py‎
Lines changed: 23 additions & 0 deletions b/‎sklearn/tree/tests/test_tree.py‎
Lines changed: 23 additions & 0 deletions
@@ -0,0 +1,2 @@
+- Use `log2` instead of `ln` for building trees to maintain behavior of previous
+  versions. By `Thomas Fan`_
@@ -11,7 +11,7 @@ and sparse data stored in a Compressed Sparse Column (CSC) format.
 # SPDX-License-Identifier: BSD-3-Clause
 
 from cython cimport final
-from libc.math cimport isnan, log
+from libc.math cimport isnan, log2
 from libc.stdlib cimport qsort
 from libc.string cimport memcpy
 
@@ -503,8 +503,8 @@ cdef class SparsePartitioner:
         # O(n_samples * log(n_indices)) is the running time of binary
         # search and O(n_indices) is the running time of index_to_samples
         # approach.
-        if ((1 - self.is_samples_sorted) * n_samples * log(n_samples) +
-                n_samples * log(n_indices) < EXTRACT_NNZ_SWITCH * n_indices):
+        if ((1 - self.is_samples_sorted) * n_samples * log2(n_samples) +
+                n_samples * log2(n_indices) < EXTRACT_NNZ_SWITCH * n_indices):
             extract_nnz_binary_search(X_indices, X_data,
                                       indptr_start, indptr_end,
                                       samples, self.start, self.end,
@@ -702,12 +702,17 @@ cdef inline void shift_missing_values_to_left_if_required(
         best.pos += best.n_missing
 
 
+def _py_sort(float32_t[::1] feature_values, intp_t[::1] samples, intp_t n):
+    """Used for testing sort."""
+    sort(&feature_values[0], &samples[0], n)
+
+
 # Sort n-element arrays pointed to by feature_values and samples, simultaneously,
 # by the values in feature_values. Algorithm: Introsort (Musser, SP&E, 1997).
 cdef inline void sort(float32_t* feature_values, intp_t* samples, intp_t n) noexcept nogil:
     if n == 0:
         return
-    cdef intp_t maxd = 2 * <intp_t>log(n)
+    cdef intp_t maxd = 2 * <intp_t>log2(n)
     introsort(feature_values, samples, n, maxd)
 
 
 
@@ -36,6 +36,7 @@
     DENSE_SPLITTERS,
     SPARSE_SPLITTERS,
 )
+from sklearn.tree._partitioner import _py_sort
 from sklearn.tree._tree import (
     NODE_DTYPE,
     TREE_LEAF,
@@ -2814,3 +2815,25 @@ def test_build_pruned_tree_infinite_loop():
         ValueError, match="Node has reached a leaf in the original tree"
     ):
         _build_pruned_tree_py(pruned_tree, tree.tree_, leave_in_subtree)
+
+
+def test_sort_log2_build():
+    """Non-regression test for gh-30554.
+    Using log2 and log in sort correctly sorts feature_values, but the tie breaking is
+    different which can results in placing samples in a different order.
+    """
+    rng = np.random.default_rng(75)
+    some = rng.normal(loc=0.0, scale=10.0, size=10).astype(np.float32)
+    feature_values = np.concatenate([some] * 5)
+    samples = np.arange(50)
+    _py_sort(feature_values, samples, 50)
+    # fmt: off
+    # no black reformatting for this specific array
+    expected_samples = [
+        0, 40, 30, 20, 10, 29, 39, 19, 49,  9, 45, 15, 35,  5, 25, 11, 31,
+        41,  1, 21, 22, 12,  2, 42, 32, 23, 13, 43,  3, 33,  6, 36, 46, 16,
+        26,  4, 14, 24, 34, 44, 27, 47,  7, 37, 17,  8, 38, 48, 28, 18
+    ]
+    # fmt: on
+    assert_array_equal(samples, expected_samples)
Original file line number	Diff line number	Diff line change
`@@ -0,0 +1,2 @@`
	`1`	+- Use `log2` instead of `ln` for building trees to maintain behavior of previous
	`2`	+ versions. By `Thomas Fan`_