Adds support for step-wise interpolation to auc and average_precision_score

Author: ndingwall

sklearn/metrics/ranking.py

@@ -36,8 +36,10 @@
 from .base import _average_binary_score
 
 
-def auc(x, y, reorder=False):
-    """Compute Area Under the Curve (AUC) using the trapezoidal rule
+def auc(x, y, reorder=False, interpolation='linear',
+        interpolation_direction='right'):
+    """Estimate Area Under the Curve (AUC) using finitely many points and an
+    interpolation strategy.
 
     This is a general function, given points on a curve.  For computing the
     area under the ROC-curve, see :func:`roc_auc_score`.
@@ -54,6 +56,24 @@ def auc(x, y, reorder=False):
         If True, assume that the curve is ascending in the case of ties, as for
         an ROC curve. If the curve is non-ascending, the result will be wrong.
 
+    interpolation : string ['trapezoid' (default), 'step']
+        This determines the type of interpolation performed on the data.
+
+        ``'linear'``:
+            Use the trapezoidal rule (linearly interpolating between points).
+        ``'step'``:
+            Use a step function where we ascend/descend from each point to the
+            y-value of the subsequent point.
+
+    interpolation_direction : string ['right' (default), 'left']
+        This determines the direction to interpolate from. The value is ignored
+        unless interpolation is 'step'.
+
+        ``'right'``:
+            Intermediate points inherit their y-value from the subsequent point.
+        ``'left'``:
+            Intermediate points inherit their y-value from the previous point.
+
     Returns
     -------
     auc : float
@@ -76,13 +96,6 @@ def auc(x, y, reorder=False):
         Compute precision-recall pairs for different probability thresholds
 
     """
-    check_consistent_length(x, y)
-    x = column_or_1d(x)
-    y = column_or_1d(y)
-
-    if x.shape[0] < 2:
-        raise ValueError('At least 2 points are needed to compute'
-                         ' area under curve, but x.shape = %s' % x.shape)
 
     direction = 1
     if reorder:
@@ -99,20 +112,42 @@ def auc(x, y, reorder=False):
                 raise ValueError("Reordering is not turned on, and "
                                  "the x array is not increasing: %s" % x)
 
-    area = direction * np.trapz(y, x)
-    if isinstance(area, np.memmap):
-        # Reductions such as .sum used internally in np.trapz do not return a
-        # scalar by default for numpy.memmap instances contrary to
-        # regular numpy.ndarray instances.
-        area = area.dtype.type(area)
+    if interpolation == 'linear':
+
+        area = direction * np.trapz(y, x)
+
+    elif interpolation == 'step':
+
+        # we need the data to start in ascending order
+        if direction == -1:
+            x, y = list(reversed(x)), list(reversed(y))
+
+        if interpolation_direction == 'right':
+            # The left-most y-value is not used
+            area = sum(np.diff(x) * np.array(y)[1:])
+
+        elif interpolation_direction == 'left':
+            # The right-most y-value is not used
+            area = sum(np.diff(x) * np.array(y)[:-1])
+
+        else:
+            raise ValueError("interpolation_direction '{}' not recognised."
+                             " Should be one of ['right', 'left']".format(
+                                 interpolation_direction))
+    else:
+        raise ValueError("interpolation value '{}' not recognized. "
+                         "Should be one of ['linear', 'step']".format(
+                             interpolation))
+
     return area
 
 
 def average_precision_score(y_true, y_score, average="macro",
-                            sample_weight=None):
+                            sample_weight=None, interpolation="linear"):
     """Compute average precision (AP) from prediction scores
 
-    This score corresponds to the area under the precision-recall curve.
+    This score corresponds to the area under the precision-recall curve, where
+    points are joined using either linear or step-wise interpolation.
 
     Note: this implementation is restricted to the binary classification task
     or multilabel classification task.
@@ -126,8 +161,7 @@ def average_precision_score(y_true, y_score, average="macro",
 
     y_score : array, shape = [n_samples] or [n_samples, n_classes]
         Target scores, can either be probability estimates of the positive
-        class, confidence values, or non-thresholded measure of decisions
-        (as returned by "decision_function" on some classifiers).
+        class, confidence values, or binary decisions.
 
     average : string, [None, 'micro', 'macro' (default), 'samples', 'weighted']
         If ``None``, the scores for each class are returned. Otherwise,
@@ -148,14 +182,24 @@ def average_precision_score(y_true, y_score, average="macro",
     sample_weight : array-like of shape = [n_samples], optional
         Sample weights.
 
+    interpolation : string ['linear' (default), 'step']
+        Determines the kind of interpolation used when computed AUC. If there are
+        many repeated scores, 'step' is recommended to avoid under- or over-
+        estimating the AUC. See www.roamanalytics.com/etc for details.
+
+        ``'linear'``:
+            Linearly interpolates between operating points.
+        ``'step'``:
+            Uses a step function to interpolate between operating points.
+
     Returns
     -------
     average_precision : float
 
     References
     ----------
     .. [1] `Wikipedia entry for the Average precision
-           <https://en.wikipedia.org/wiki/Average_precision>`_
+           <http://en.wikipedia.org/wiki/Average_precision>`_
 
     See also
     --------
@@ -177,8 +221,20 @@ def average_precision_score(y_true, y_score, average="macro",
     def _binary_average_precision(y_true, y_score, sample_weight=None):
         precision, recall, thresholds = precision_recall_curve(
             y_true, y_score, sample_weight=sample_weight)
-        return auc(recall, precision)
-
+        return auc(recall, precision, interpolation=interpolation,
+                   interpolation_direction='right')
+
+    if interpolation == "linear":
+        # Check for number of unique predictions. If this is substantially less
+        # than the number of predictions, linear interpolation is likely to be
+        # biased.
+        n_discrete_predictions = len(np.unique(y_score))
+        if n_discrete_predictions < 0.75 * len(y_score):
+            warnings.warn("Number of unique scores is less than 75% of the "
+                          "number of scores provided. Linear interpolation "
+                          "is likely to be biased in this case. You may wish "
+                          "to use step interpolation instead. See docstring "
+                          "for details.")
     return _average_binary_score(_binary_average_precision, y_true, y_score,
                                  average, sample_weight=sample_weight)
 
@@ -252,7 +308,7 @@ def _binary_roc_auc_score(y_true, y_score, sample_weight=None):
 
         fpr, tpr, tresholds = roc_curve(y_true, y_score,
                                         sample_weight=sample_weight)
-        return auc(fpr, tpr, reorder=True)
+        return auc(fpr, tpr, reorder=True, interpolation='linear')
 
     return _average_binary_score(
         _binary_roc_auc_score, y_true, y_score, average,