selimfirat/pysad

IForestASD (iforest_asd.py) is a thin subclass of ReferenceWindowModel with sliding_size == window_size, so it retrains a fresh IForest on every window, unconditionally. The docstring says concept drift "is not implemented since it is a part of the simulation" and there is a # TODO: implement concept drift method on line 14. But the drift check is the part of the paper that makes iForestASD different from plain reference windowing. In Algorithm 2 of Ding and Fei (2013, ding2013anomaly in refs.bib), the current forest scores the instances of the newest window, the anomaly rate u of that window is compared against a threshold, and the forest is rebuilt from the new window only if u exceeds the threshold. Otherwise the existing forest is kept and the window is discarded. Proposal: Add drift_threshold: float | None = None to IForestASD.__init__. With None the model behaves exactly as today (retrain on every window), so nothing changes for existing users. With a value in [0, 1], on each fu

Following #72 and #106, pysad/utils is another good module for type hints. It's small (array_streamer.py, pandas_streamer.py, window.py, data.py and the helpers in __init__.py), and it's the part of the API that almost every user calls first. It would be useful to add type hints to the utils module so editors and type checkers can help users. Done when: Every public function and method in pysad/utils/*.py and pysad/core/base_streamer.py has argument and return annotations (e.g. get_minmax_array(X: np.ndarray) -> tuple[np.ndarray, np.ndarray], and Iterator[...] for the streamers' iter) PandasStreamer annotates its arguments as pd.DataFrame without importing pandas at runtime, since pandas is optional (use from __future__ import annotations with if TYPE_CHECKING:) Only built-in syntax supported on Python 3.10+ is used, with no new dependencies Docstrings stay as they are, and pytest tests/utils tests/test_utils.py passes Optional: mypy pysad/utils reports no errors in these files

tests/models/test_standard_absolute_deviation.py checks that StandardAbsoluteDeviation returns one score per instance and rejects bad arguments, but never checks the scores themselves. It would be useful to pin down what the model returns on a small stream, so a regression in the formula is caught. Done when: On a small fixed univariate stream (for example [3, 1, 4, 1, 5, 9, 2, 6] as a column), fit_score with the default substracted_statistic="mean" is checked against |x_t - mean(x_1..x_t)| / std(x_1..x_t), with the population standard deviation (np.std, ddof=0) The same is checked for substracted_statistic="median", subtracting the running median instead absolute=False is checked to keep the sign, so it's negative for values below the mean The first score (a single value, so zero variance) is asserted explicitly, whatever it is today pytest tests/models/test_standard_absolute_deviation.py passes The model is in pysad/models/standard_absolute_deviation.py. Tests only: please don

Windowed components accept a window_size of 0 or less without complaint and then fail later, far from the mistake: from pysad.statistics import AverageMeter, RunningStatistic from pysad.transform.postprocessing import RunningAveragePostprocessor from pysad.transform.probability_calibration import ConformalProbabilityCalibrator from pysad.utils import Window RunningStatistic(AverageMeter, window_size=0).update(1.0).get() # ZeroDivisionError: float division by zero RunningAveragePostprocessor(window_size=0).fit_transform_partial(1.0) # ZeroDivisionError: float division by zero ConformalProbabilityCalibrator(window_size=0).fit_transform_partial(1.0) # returns nan, with a RuntimeWarning window = Window(0) window.update(1.0) window.get() # [], because the window silently keeps nothing It would be useful to reject these values when the object is created, with a clear message. Done when: RunningStatistic, Window a

tests/transform/postprocessing/test_postprocessors.py runs all eight score postprocessors on random scores, but it only checks the output shape. A regression that returned the wrong statistic, or the wrong window, would still pass. It would be useful to check the actual values the postprocessors return, as #68 did for the ensemblers. Done when: On a small fixed list of scores (for example [3, 1, 4, 1, 5, 9, 2, 6]), AveragePostprocessor, MaxPostprocessor and MedianPostprocessor are checked against np.mean, np.max and np.median of all the scores seen so far, including the current one RunningAveragePostprocessor, RunningMaxPostprocessor and RunningMedianPostprocessor with a small window_size (for example 3) are checked against the same functions over the last window_size scores ZScorePostprocessor and RunningZScorePostprocessor are checked against (score - mean) / std with the population standard deviation (np.std, ddof=0). The first output is nan because a single score has zero varia