Coming from river or PySAD¶

aberrant uses river's event-by-event contract, so most river code ports by changing imports. PySAD code needs one extra step: events become dict[str, float] instead of NumPy rows.

The loop¶

from aberrant.model.iforest import HalfSpaceTrees
from aberrant.transform import MinMaxScaler

detector = MinMaxScaler() | HalfSpaceTrees(seed=42)

for event in [{"x": 0.1, "y": 0.4}, {"x": 0.2, "y": 0.5}, {"x": 0.9, "y": 0.1}]:
    score = detector.score_one(event)
    detector.learn_one(event)

from river import anomaly, preprocessing

detector = preprocessing.MinMaxScaler() | anomaly.HalfSpaceTrees(seed=42)

for event in [{"x": 0.1, "y": 0.4}, {"x": 0.2, "y": 0.5}, {"x": 0.9, "y": 0.1}]:
    detector.learn_one(event)
    score = detector.score_one(event)

import numpy as np
from pysad.models import HalfSpaceTrees

X = np.array([[0.1, 0.4], [0.2, 0.5], [0.9, 0.1]])
detector = HalfSpaceTrees(feature_mins=X.min(axis=0), feature_maxes=X.max(axis=0))

for row in X:
    score = detector.fit_score_partial(row)

Higher scores mean more anomalous in all three libraries. aberrant keeps score_one and learn_one separate: for prequential (test-then-train) evaluation, call score_one first, which is what PySAD's fit_score_partial does internally. Scalers return neutral values before their first learned event, so this order works from the first event; see the transformer guide.

Rows to events¶

PySAD and most NumPy pipelines produce rows. Name the columns once and turn each row into a mapping:

import numpy as np

from aberrant.model.sketch import StreamingLODA

X = np.random.default_rng(0).normal(size=(500, 3))
names = ["temperature", "pressure", "vibration"]

detector = StreamingLODA(seed=0)
scores = []
for row in X:
    event = dict(zip(names, row.tolist(), strict=True))
    scores.append(detector.score_one(event))
    detector.learn_one(event)

For a labelled benchmark, PrequentialEvaluator replaces PySAD's streamers and metric wrappers; see Evaluation.

Equivalent components¶

The table maps components that implement the same published method. Defaults and score scales can still differ, so compare results rather than assuming identical numbers.

aberrant also covers families that neither library ships: graph-edge detectors (MIDAS, AnoEdgeL, ISCONNA), time-series discords (RollingMatrixProfile, XLagDAMP), online SVMs and autoencoders. The model guide lists them all.