transform scores

huiwengoh · Feb 27, 2024 · 99268ce · 99268ce
1 parent 09245a9
commit 99268ce
Showing 1 changed file with 5 additions and 1 deletion.
diff --git a/cleanlab/regression/learn.py b/cleanlab/regression/learn.py
@@ -490,7 +490,11 @@ def find_label_issues(
                     "If uncertainty is passed in as an array, it must have the same length as y."
                 )
 
-        label_quality_scores = np.exp(-abs(residual) / (uncertainty + TINY_VALUE))
+        residual_adjusted = abs(residual / (uncertainty + TINY_VALUE))
+
+        # adjust lqs by the median (for more human-readable scores)
+        residual_median = max(np.median(residual_adjusted), TINY_VALUE)
+        label_quality_scores = np.exp(-residual_adjusted / residual_median)
 
         label_issues_mask = np.zeros(len(y), dtype=bool)
         num_issues = math.ceil(len(y) * self.k)