BUG: highly-degenerate roc curves

GaelVaroquaux · larsmans · commit 84eaed32864e · 2013-02-20T13:49:55.000+01:00
Fixes 1658

Make sure that roc_curves start at [0, 0] and end at [1, 1]
diff --git a/sklearn/metrics/metrics.py b/sklearn/metrics/metrics.py
@@ -482,11 +482,11 @@ def roc_curve(y_true, y_score, pos_label=None):
 
     if n_pos == 0:
         warnings.warn("No positive samples in y_true, "
-                      "true positve value should be meaningless")
+                      "true positive value should be meaningless")
         n_pos = np.nan
     if n_neg == 0:
         warnings.warn("No negative samples in y_true, "
-                      "false positve value should be meaningless")
+                      "false positive value should be meaningless")
         n_neg = np.nan
 
     thresholds = np.unique(y_score)
@@ -520,11 +520,28 @@ def roc_curve(y_true, y_score, pos_label=None):
         tpr[-1] = (sum_pos + current_pos_count) / n_pos
         fpr[-1] = (sum_neg + current_neg_count) / n_neg
 
-    # hard decisions, add (0,0)
-    if fpr.shape[0] == 2:
+    thresholds = thresholds[::-1]
+
+    if not (n_pos is np.nan or n_neg is np.nan):
+        # add (0,0) and (1, 1)
+        if not (fpr[0] == 0 and fpr[-1] == 1):
+            fpr = np.r_[0., fpr, 1.]
+            tpr = np.r_[0., tpr, 1.]
+            thresholds = np.r_[thresholds[0] + 1, thresholds,
+                               thresholds[-1] - 1]
+        elif not fpr[0] == 0:
+            fpr = np.r_[0., fpr]
+            tpr = np.r_[0., tpr]
+            thresholds = np.r_[thresholds[0] + 1, thresholds]
+        elif not fpr[-1] == 1:
+            fpr = np.r_[fpr, 1.]
+            tpr = np.r_[tpr, 1.]
+            thresholds = np.r_[thresholds, thresholds[-1] - 1]
+    elif fpr.shape[0] == 2:
+        # trivial decisions, add (0,0)
         fpr = np.array([0.0, fpr[0], fpr[1]])
         tpr = np.array([0.0, tpr[0], tpr[1]])
-    # trivial decisions, add (0,0) and (1,1)
+        # trivial decisions, add (0,0) and (1,1)
     elif fpr.shape[0] == 1:
         fpr = np.array([0.0, fpr[0], 1.0])
         tpr = np.array([0.0, tpr[0], 1.0])
@@ -535,7 +552,7 @@ def roc_curve(y_true, y_score, pos_label=None):
     if n_neg is np.nan:
         fpr[0] = np.nan
 
-    return fpr, tpr, thresholds[::-1]
+    return fpr, tpr, thresholds
 
 
 ###############################################################################
diff --git a/sklearn/metrics/tests/test_metrics.py b/sklearn/metrics/tests/test_metrics.py
@@ -93,6 +93,17 @@ def test_roc_curve():
     assert_almost_equal(roc_auc, auc_score(y_true, probas_pred))
 
 
+def test_roc_curve_end_points():
+    # Make sure that roc_curve returns a curve start at 0 and ending and
+    # 1 even in corner cases
+    rng = np.random.RandomState(0)
+    y_true = np.array([0] * 50 + [1] * 50)
+    y_pred = rng.randint(3, size=100)
+    fpr, tpr, thr = roc_curve(y_true, y_pred)
+    assert_equal(fpr[0], 0)
+    assert_equal(fpr[-1], 1)
+
+
 def test_roc_returns_consistency():
     """Test whether the returned threshold matches up with tpr"""
     # make small toy dataset
@@ -101,8 +112,8 @@ def test_roc_returns_consistency():
 
     # use the given thresholds to determine the tpr
     tpr_correct = []
-    for t in range(len(thresholds)):
-        tp = np.sum((probas_pred >= thresholds[t]) & y_true)
+    for t in thresholds:
+        tp = np.sum((probas_pred >= t) & y_true)
         p = np.sum(y_true)
         tpr_correct.append(1.0 * tp / p)
 
