scikit-learn · qinhanmin2014 · Nov 30, 2018 · Nov 28, 2018 · Nov 28, 2018 · Nov 28, 2018
diff --git a/doc/whats_new/v0.20.rst b/doc/whats_new/v0.20.rst
@@ -12,6 +12,16 @@ Version 0.20.2
 This is a bug-fix release with some minor documentation improvements and
    enhancements to features released in 0.20.0.
 
+Changed models
+--------------
+
+The following estimators and functions, when fit with the same data and
+parameters, may produce different models from the previous version. This often
+occurs due to changes in the modelling logic (bug fixes or enhancements), or in
+random sampling procedures.
+
+- :mod:`sklearn.neighbors` when ``metric=='jaccard'`` (bug fix)
+
 Changelog
 ---------
 
@@ -22,6 +32,12 @@ Changelog
   parameter may not have been updated correctly when a step is set to ``None``
   or ``'passthrough'``. :user:`Thomas Fan <thomasjpfan>`.
 
+:mod:`sklearn.neighbors`
+........................
+
+- |Fix| Fixed :class:`sklearn.neighbors.DistanceMetric` jaccard distance
+  function to return 0 when two all-zero vectors are compared.
+  :issue:`12685` by :user:`Thomas Fan <thomasjpfan>`.
 
 .. _changes_0_20_1:
 

diff --git a/sklearn/neighbors/dist_metrics.pyx b/sklearn/neighbors/dist_metrics.pyx
@@ -788,6 +788,11 @@ cdef class JaccardDistance(DistanceMetric):
             tf2 = x2[j] != 0
             nnz += (tf1 or tf2)
             n_eq += (tf1 and tf2)
+        # Based on https://github.com/scipy/scipy/pull/7373
+        # When comparing two all-zero vectors, scipy>=1.2.0 jaccard metric
+        # was changed to return 0, instead of nan.
+        if nnz == 0:
+            return 0
         return (nnz - n_eq) * 1.0 / nnz
 
 

diff --git a/sklearn/neighbors/tests/test_dist_metrics.py b/sklearn/neighbors/tests/test_dist_metrics.py
@@ -6,6 +6,8 @@
 
 import pytest
 
+from distutils.version import LooseVersion
+from scipy import __version__ as scipy_version
 from scipy.spatial.distance import cdist
 from sklearn.neighbors.dist_metrics import DistanceMetric
 from sklearn.neighbors import BallTree
@@ -101,6 +103,11 @@ def check_pdist(metric, kwargs, D_true):
 def check_pdist_bool(metric, D_true):
     dm = DistanceMetric.get_metric(metric)
     D12 = dm.pairwise(X1_bool)
+    # Based on https://github.com/scipy/scipy/pull/7373
+    # When comparing two all-zero vectors, scipy>=1.2.0 jaccard metric
+    # was changed to return 0, instead of nan.
+    if metric == 'jaccard' and LooseVersion(scipy_version) < '1.2.0':
+        D_true[np.isnan(D_true)] = 0
     assert_array_almost_equal(D12, D_true)