Loda

class capymoa.anomaly.Loda

Bases: AnomalyDetector

Loda: Lightweight on-line detector of anomalies We implement a streaming version of Loda that updates the histograms after every window of instances.

>>> from capymoa.datasets import ElectricityTiny
>>> from capymoa.anomaly import Loda
>>> from capymoa.evaluation import AnomalyDetectionEvaluator
>>> stream = ElectricityTiny()
>>> schema = stream.get_schema()
>>> learner = Loda(schema, n_projections=10, window_size=100, random_state=42)
>>> evaluator = AnomalyDetectionEvaluator(schema)
>>> while stream.has_more_instances():
...     instance = stream.next_instance()
...     proba = learner.score_instance(instance)
...     evaluator.update(instance.y_index, proba)
...     learner.train(instance)
>>> auc = evaluator.auc()
>>> print(f"AUC: {auc:.2f}")
AUC: 0.65

Reference:

Pevný, T. (2016). Loda: Lightweight on-line detector of anomalies. Machine Learning, 102(2), 275-304.

__init__(

schema: Schema,

n_projections: int = 100,

window_size: int = 256,

max_bins: str | int = 'auto',

random_state: int = 42,

)

Initialize the Loda anomaly detector.

Parameters:

• schema – Schema of the data stream.

• n_projections – Number of random projection histograms in the ensemble.

• window_size – Number of recent instances used to fit each histogram.

• max_bins – Upper bound on bins per histogram. "auto" uses floor(n / log n) via the Birge-Rozenholc criterion; an integer caps the search at that value.

• random_state – Random seed for reproducibility.

predict(instance: Instance) → int

score_instance(instance: Instance) → float

Return the anomaly score for a single instance.

Computes the average negative log-likelihood of the instance under the ensemble of one-dimensional histograms. Higher scores indicate greater anomalousness. Returns 0.0 before the first full window of window_size instances has been seen.

Parameters:

instance – The instance to score.

Returns:

Anomaly score. Ranges from 0.0 (least anomalous) to infinity (most anomalous).

train(instance: Instance)

Train on a single instance.

Projects the instance onto each random projection vector and adds it to the circular window. Histograms are rebuilt once every window_size instances using the Birge-Rozenholc criterion to select the optimal number of bins.

Parameters:

instance – The instance to train on.