PyOD 3.6.9
A maintenance release with fixes to HBOS, predict_proba and weighted PCA, plus a Core Team section in the README. Default PCA scores change on rank-deficient data (constant, duplicated or collinear columns) and on float32 input; see Score Changes and Behavior Notes. No new detector or dependency is introduced.
Score Changes
- PCA (weighted, the default): components whose explained variance is at most
16 * n_features * eps(float64) times the largest one are left out, with aUserWarningthat reports how many. A constant, duplicated or collinear column used to give such a component a near-infinite weight. Scores stay bit-identical on 13 of the 17 bundled benchmark datasets. mnist and optdigits go from non-finite scores to ROC AUC 0.85 and 0.51. On arrhythmia and cardio the largest raw score falls from about 3e36 to 3e9 and from about 2e18 to 4e4, with ROC AUC unchanged on arrhythmia and 0.950 to 0.951 on cardio. #772 by @AHMETHAKANBEZIR1 with maintainer follow-ups, fixes #770.
Bug Fixes
- HBOS:
decision_function,predictandpredict_probaraiseValueErrorwhenXhas a different number of features than infit. With fewer features HBOS returned scores against the first histograms without an error; with more it raisedIndexErroror an unrelatedValueErrorfrom the bin lookup. Scores for input of the training width are unchanged. #790 by @xiyanghu. - predict_proba:
BaseDetector.predict_probasizes its output from the decision scores, so list input works, as it already did fordecision_functionandpredict. #791 by @xiyanghu.
Documentation
- The README and the documentation landing page have a Core Team section: a team card with Yue Zhao, Xiyang Hu, Zain Nasrullah, Winston (Zheng) Li and the ten contributors outside the core team with the most commits.
- The About page lists Xiyang Hu as co-lead, dates Lorenzo Perini's work from 2021, and names the detectors Daniel Kulik added.
- The README, the documentation landing page and the Impact page read 65+ million downloads, and the Impact page reads 10K+ GitHub stars.
Behavior Notes
Weighted PCA converts float32 input to float64 in fit and decision_function. Scores can change substantially on nearly collinear float32 data, and the float64 copy takes twice the memory of the float32 input during the call. If every component is left out (all-constant training data), all training and prediction scores are 0.
Contributors
Thanks to @xiyanghu and @AHMETHAKANBEZIR1.
Full changelog: v3.6.8...v3.6.9