scikit-learn
diff --git a/‎sklearn/utils/stats.py
Copy file name to clipboard
+67-8Lines changed: 67 additions & 8 deletions b/‎sklearn/utils/stats.py
Copy file name to clipboard
+67-8Lines changed: 67 additions & 8 deletions
diff --git a/‎sklearn/utils/tests/test_stats.py
Copy file name to clipboard
+89Lines changed: 89 additions & 0 deletions b/‎sklearn/utils/tests/test_stats.py
Copy file name to clipboard
+89Lines changed: 89 additions & 0 deletions
@@ -1,18 +1,77 @@
-import numpy as np
+"""Statistical utilities including weighted percentile"""
 
+import numpy as np
 from sklearn.utils.extmath import stable_cumsum
 
 
 def _weighted_percentile(array, sample_weight, percentile=50):
+    """Compute the weighted ``percentile`` of ``array``
+    with ``sample_weight``.
+
+    This approach follows
+
+               N
+        S_N = sum w_k
+              k=1
+
+        p_n = 1 / S_N * (x_n - w_n / 2)
+
+        v = v_k + (v_{k + 1} - v_k) * (P - p_k) / (p_{k + 1} - p_k)
+
+    from
+    https://en.wikipedia.org/wiki/Percentile#The_weighted_percentile_method.
+
+
+    Parameters
+    ----------
+    array : array-like, shape = (n_samples,)
+        Array of data on which to calculate the weighted percentile
+
+    sample_weight : array-like, shape = (n_samples,)
+        Array of corresponding sample weights with which to calculate
+        the weighted percentile
+
+    percentile : int, optional (default: 50)
+        Integer value of Pth percentile to compute
+
+    Returns
+    -------
+    v : float
+        Linearly interpolated weighted percentile.
+
+    Examples
+    --------
+    >>> import numpy as np
+    >>> from sklearn.utils.stats import _weighted_percentile
+    >>> weight = np.array([1, 1])
+    >>> data = np.array([0, 1])
+    >>> _weighted_percentile(data, weight, percentile=0)
+    0.0
+    >>> _weighted_percentile(data, weight, percentile=50)
+    0.5
+    >>> _weighted_percentile(data, weight, percentile=90)
+    1.0
     """
-    Compute the weighted ``percentile`` of ``array`` with ``sample_weight``.
-    """
+    if not isinstance(array, np.ndarray):
+        array = np.array(array)
+
+    if not isinstance(sample_weight, np.ndarray):
+        sample_weight = np.array(sample_weight)
+
+    if (sample_weight < 0).any():
+        raise ValueError("sample_weight must contain positive or 0 weights")
+
+    if percentile < 0:
+        raise ValueError("percentile must be positive or 0")
+
     sorted_idx = np.argsort(array)
+    sorted_array = array[sorted_idx]
+
+    # if there are no weights, return the min of ``array``
+    if sample_weight.sum() == 0:
+        return sorted_array[0]
 
     # Find index of median prediction for each sample
     weight_cdf = stable_cumsum(sample_weight[sorted_idx])
-    percentile_idx = np.searchsorted(
-        weight_cdf, (percentile / 100.) * weight_cdf[-1])
-    # in rare cases, percentile_idx equals to len(sorted_idx)
-    percentile_idx = np.clip(percentile_idx, 0, len(sorted_idx)-1)
-    return array[sorted_idx[percentile_idx]]
+    p_n = 100. / weight_cdf[-1] * (weight_cdf - sample_weight / 2.)
+    return np.interp(percentile, p_n, sorted_array)
@@ -0,0 +1,89 @@
+import numpy as np
+from sklearn.utils.testing import assert_equal, assert_raises
+from sklearn.utils.stats import _weighted_percentile
+
+
+def test_weighted_percentile_negative_weights_raises():
+    weight = np.array([1, -1])
+    data = np.array([0, 1])
+    assert_raises(ValueError, _weighted_percentile, data, weight)
+
+
+def test_weighted_percentile_negative_percentile_raises():
+    weight = np.array([1, -1])
+    data = np.array([0, 1])
+    percentile = -50
+    assert_raises(ValueError, _weighted_percentile, data, weight,
+                  percentile=percentile)
+
+
+def test_weighted_percentile_median_interpolated_list():
+    weight = [1, 1]
+    data = [0, 1]
+    percentile = 50
+    expected = 0.5
+    actual = _weighted_percentile(data, weight, percentile=percentile)
+    assert_equal(expected, actual)
+
+
+def test_weighted_percentile_median_interpolated_tuple():
+    weight = (1, 1)
+    data = (0, 1)
+    percentile = 50
+    expected = 0.5
+    actual = _weighted_percentile(data, weight, percentile=percentile)
+    assert_equal(expected, actual)
+
+
+def test_weighted_percentile_median_interpolated():
+    weight = np.array([1, 1])
+    data = np.array([0, 1])
+    percentile = 50
+    expected = 0.5
+    actual = _weighted_percentile(data, weight, percentile=percentile)
+    assert_equal(expected, actual)
+
+
+def test_weighted_percentile_median_regular():
+    weight = np.array([1, 1, 1])
+    data = np.array([0, 1, 2])
+    percentile = 50
+    expected = 1.0
+    actual = _weighted_percentile(data, weight, percentile=percentile)
+    assert_equal(expected, actual)
+
+
+def test_weighted_percentile_0_regular():
+    weight = np.array([1, 1, 1])
+    data = np.array([0, 1, 2])
+    percentile = 0
+    expected = 0.0
+    actual = _weighted_percentile(data, weight, percentile=percentile)
+    assert_equal(expected, actual)
+
+
+def test_weighted_percentile_90_regular():
+    weight = np.array([1, 1, 1])
+    data = np.array([0, 1, 2])
+    percentile = 90
+    expected = 2.0
+    actual = _weighted_percentile(data, weight, percentile=percentile)
+    assert_equal(expected, actual)
+
+
+def test_weighted_percentile_70_interpolated():
+    weight = np.array([1, 1, 1, 1])
+    data = np.arange(0, 4, 1)
+    percentile = 70
+    expected = 2.3
+    actual = _weighted_percentile(data, weight, percentile=percentile)
+    assert_equal(expected, actual)
+
+
+def test_weighted_percentile_70_mixed_weights():
+    weight = np.array([1, 0, 1, 1])
+    data = np.arange(0, 4, 1)
+    percentile = 50
+    expected = 2.0
+    actual = _weighted_percentile(data, weight, percentile=percentile)
+    assert_equal(expected, actual)