Merge pull request #26 from selimfirat/develop

Release v0.3.4

Author: selimfirat

CHANGES.txt

@@ -3,4 +3,5 @@ v<0.2.0> <06/10/2023> -- Upgraded some libraries like numpy,pandas and small ref
 v<0.3.0>, <05/23/2025> -- Updated dependencies to latest versions, improved documentation, added quick start example in README. Added Inqmad model. Special thanks to @TechyNilesh @Joaggi @onixlas.
 v<0.3.1>, <06/04/2025> -- Enhanced README with comprehensive community engagement section, educational content, third-party integrations, and developer community highlights. Fixed sublist formatting in reStructuredText. Improved error handling for Inqmad model imports.
 v<0.3.2>, <06/14/2025> -- Updated PyOD version and added Python 3.13 support. Fixed Flaticon link in documentation. Enhanced error handling for optional JAX dependencies. Updated rrcf version to 0.4.4 and improved quantization in RelativeEntropy model. Improved test coverage and configuration. Added deprecation warning tests and enhanced model methods array handling. Special thanks to @onixlas for multiple contributions.
-v<0.3.3>, <06/19/2025> -- Fixed critical bugs in ReferenceWindowModel: resolved data mutation issues by ensuring reference window data is properly copied, fixed window_length parameter naming to window_size, and improved reference window label management. Enhanced test coverage and fixed minor typo in bibtex reference. Special thanks to @onixlas for bug reports and fixes.
+v<0.3.3>, <06/19/2025> -- Fixed critical bugs in ReferenceWindowModel: resolved data mutation issues by ensuring reference window data is properly copied, fixed window_length parameter naming to window_size, and improved reference window label management. Enhanced test coverage and fixed minor typo in bibtex reference. Special thanks to @onixlas for bug reports and fixes.
+v<0.3.4>, <06/24/2025> -- Fixed critical reference window reset bug in ReferenceWindowModel (#25), preventing incorrect window reinitialization during sliding window operations. Added validation for single class presence in AUROCMetric evaluation to prevent evaluation errors. Fixed hash seed type casting issue in StreamhashProjector's _hash_string method. Enhanced CI infrastructure with Python 3.13 support and updated dependency management. Special thanks to @onixlas for discovering and reporting the reference window bug.

README.rst

@@ -191,13 +191,10 @@ Quick Links
 Contributors
 ============
 
-.. raw:: html
-
-   <div align="center">
-     <a href="https://github.com/selimfirat/pysad/graphs/contributors">
-       <img src="https://contrib.rocks/image?repo=selimfirat/pysad" alt="Contributors" />
-     </a>
-   </div>
+.. image:: https://contrib.rocks/image?repo=selimfirat/pysad
+    :target: https://github.com/selimfirat/pysad/graphs/contributors
+    :alt: Contributors
+    :align: center
 
 We thank all our contributors for their valuable input and efforts to make PySAD better!
 

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "pysad"
-version = "0.3.3"
+version = "0.3.4"
 dependencies = [
   "numpy",
   "scikit-learn",
@@ -20,7 +20,7 @@ maintainers = [
   {name = "Selim Firat Yilmaz", email = "yilmazselimfirat@gmail.com"}
 ]
 description = "PySAD is an open-source python framework for anomaly detection on streaming multivariate data."
-readme = "README.rst"
+readme = {file = "README.rst", content-type = "text/x-rst"}
 license = {file = "LICENSE"}
 keywords = ["outlier detection", "anomaly detection", "machine learning"]
 classifiers = [

pysad/evaluation/metrics.py

@@ -63,6 +63,9 @@ class AUROCMetric(BaseSKLearnMetric):
     """
 
     def _evaluate(self, y_true, y_pred):
+        # Check if only one class is present
+        if len(set(y_true)) <= 1:
+            raise ValueError("Only one class present in y_true. ROC AUC score is not defined in that case.")
         return roc_auc_score(y_true, y_pred)
 
 

pysad/models/integrations/reference_window_model.py

@@ -64,7 +64,8 @@ def fit_partial(self, X, y=None):
             self.cur_window_y.append(y)
 
         if not self.initial_ref_window and len(
-                self.cur_window_X) < self.window_size:
+                self.cur_window_X) < self.window_size and (
+                self.reference_window_X is None or len(self.reference_window_X) < self.window_size):
             self.reference_window_X = self.cur_window_X.copy()
             self.reference_window_y = self.cur_window_y.copy() if y is not None else None
             self._fit_model()

pysad/transform/projection/streamhash_projector.py

@@ -53,7 +53,7 @@ def transform_partial(self, X):
 
     def _hash_string(self, k, s):
         import mmh3
-        hash_value = int(mmh3.hash(s, signed=False, seed=k)) / (2.0 ** 32 - 1)
+        hash_value = int(mmh3.hash(s, signed=False, seed=int(k))) / (2.0 ** 32 - 1)
         s = self.density
         if hash_value <= s / 2.0:
             return -1 * self.constant

pysad/version.py

@@ -21,4 +21,4 @@
 # 'X.Y.dev0' is the canonical version of 'X.Y.dev'
 #
 
-__version__ = '0.3.3'
+__version__ = '0.3.4'

tests/models/test_pyod_integrations.py

@@ -271,3 +271,55 @@ def test_reference_window_model_issue_23_fix():
     expected = [1.0, 2.0]
     assert ref_values == expected, \
         f"Expected {expected}, but got {ref_values}. Duplication detected!"
+
+
+def test_reference_window_model_issue_25_fix():
+    """Test for the fix of issue #25: Reference window reset bug.
+    
+    The bug: When cur_window_X length is less than window_size after sliding,
+    the reference_window_X is incorrectly reset to just the current window
+    instead of maintaining the properly sized reference window.
+    
+    Test scenario: window_size=4, sliding_size=2, streaming data [1,2,3,4,5,6,7,8,...]
+    The critical test is after sliding occurs and cur_window_X becomes small again,
+    the reference_window_X should NOT be reset to just the current window.
+    """
+    from pysad.models.integrations.reference_window_model import ReferenceWindowModel
+    from pyod.models.iforest import IForest
+    import numpy as np
+    
+    window_size = 4
+    sliding_size = 2
+    
+    model = ReferenceWindowModel(
+        model_cls=IForest,
+        window_size=window_size,
+        sliding_size=sliding_size,
+        initial_window_X=None
+    )
+    
+    # Build up the initial window
+    for i in range(1, 9):
+        model.fit_partial(np.array([float(i)]))
+    
+    # At this point, after several sliding operations, the reference window
+    # should be properly sized (window_size=4) and should NOT be reset to
+    # just the current window when cur_window_X is small
+    
+    # Verify reference window is properly sized
+    assert len(model.reference_window_X) == window_size, \
+        f"Reference window should be size {window_size}, got {len(model.reference_window_X)}"
+    
+    # Verify reference window is not just the current window
+    ref_values = [x[0] for x in model.reference_window_X]
+    cur_values = [x[0] for x in model.cur_window_X] if model.cur_window_X else []
+    
+    # The reference window should not be identical to the current window
+    # (this would indicate the bug where reference_window_X gets reset)
+    assert ref_values != cur_values, \
+        f"Reference window should not be reset to current window. Got ref={ref_values}, cur={cur_values}"
+    
+    # Reference window should contain the most recent window_size elements
+    # from the sliding window, not just the current partial window
+    assert len(ref_values) == window_size, \
+        f"Reference window should maintain size {window_size}, got {len(ref_values)}"