Inject safer test value for MeanShift.bandwidth in estimator checks (#21501)

jjerphan · glemaitre · jeremiedbb · glemaitre · commit 48ab3947accc · 2021-12-25T12:26:49.000+01:00
This modify the test configuration so that it makes sense for when a sole sample is provided for MeanShift. This test was passing previously for this configuration but was not supposed to. The new implementation strategy for kneighbors which uses PairwiseDistancesArgKmin (see #21462) is numerically stabler for this case, motivating this modication. Co-authored-by: Guillaume Lemaitre <g.lemaitre58@gmail.com> Co-authored-by: Jérémie du Boisberranger <jeremiedbb@users.noreply.github.com>
diff --git a/sklearn/utils/estimator_checks.py b/sklearn/utils/estimator_checks.py
@@ -632,6 +632,11 @@ def _set_checking_parameters(estimator):
     if "n_init" in params:
         # K-Means
         estimator.set_params(n_init=2)
+    if name == "MeanShift":
+        # In the case of check_fit2d_1sample, bandwidth is set to None and
+        # is thus estimated. De facto it is 0.0 as a single sample is provided
+        # and this makes the test fails. Hence we give it a placeholder value.
+        estimator.set_params(bandwidth=1.0)
 
     if name == "TruncatedSVD":
         # TruncatedSVD doesn't run with n_components = n_features
