Merge pull request #15 from justinkay/bugfix/sorted_highlow

Remove sorted_highlow from form_discrete_distribution call in ppi_distribution_label_shift_ci.

Author: aangelopoulos

ppi_py/ppi.py

@@ -1759,7 +1759,7 @@ def ppi_distribution_label_shift_ci(
 
     # Invert Ahat
     Ahatinv = np.linalg.inv(Ahat)
-    qfhat = form_discrete_distribution(Yhat_unlabeled, sorted_highlow=True)
+    qfhat = form_discrete_distribution(Yhat_unlabeled)
 
     # Calculate the bound
     point_estimate = nu @ Ahatinv @ qfhat

tests/test_labelshift.py

@@ -0,0 +1,41 @@
+import numpy as np
+from ppi_py import *
+
+def test_ppi_distribution_label_shift_ci_sorted():
+    """Test ppi_distribution_label_shift_ci with various sorting scenarios."""
+    n = 1000
+    N = 10000
+    num_classes = 10
+
+    # Generate multi-class predictions (integer labels)
+    Y = np.random.randint(0, num_classes, n)
+    Yhat = np.random.randint(0, num_classes, n)
+    Yhat_unlabeled = np.random.randint(0, num_classes, N)
+
+    # Determine K and nu
+    unique_Yhat = np.unique(Yhat)
+    K = len(unique_Yhat)  # Should be equal to num_classes if all classes are present in Yhat
+
+    # Set nu (example: vector of ones)
+    nu = np.ones(K)
+
+    # Test 1: Unsorted Yhat_unlabeled (implicit in the current implementation)
+    ci_unsorted = ppi_distribution_label_shift_ci(Y, Yhat, Yhat_unlabeled, K, nu, alpha=0.1)
+
+    # Test 2: Explicitly sorted Yhat_unlabeled (to demonstrate no dependence on order)
+    uq, uq_counts = np.unique(Yhat_unlabeled, return_counts=True)
+    sort_indices = np.argsort(uq)
+    Yhat_unlabeled_sorted = Yhat_unlabeled[np.argsort(np.argsort(Yhat_unlabeled))] # Stable sort to mimic previous behavior.
+    ci_sorted = ppi_distribution_label_shift_ci(Y, Yhat, Yhat_unlabeled_sorted, K, nu, alpha=0.1)
+
+
+    # Test 3: Reverse sorted Yhat_unlabeled
+    uq, uq_counts = np.unique(Yhat_unlabeled, return_counts=True)
+    sort_indices = np.argsort(uq)[::-1]
+    Yhat_unlabeled_reverse_sorted = Yhat_unlabeled[np.argsort(np.argsort(Yhat_unlabeled))[::-1]]
+    ci_reverse_sorted = ppi_distribution_label_shift_ci(Y, Yhat, Yhat_unlabeled_reverse_sorted, K, nu, alpha=0.1)
+
+
+    # The confidence intervals should be identical regardless of the sorting of Yhat_unlabeled
+    np.testing.assert_allclose(ci_unsorted, ci_sorted)
+    np.testing.assert_allclose(ci_unsorted, ci_reverse_sorted)