Plugins

External packages can add detectors without being vendored into AnomalyBench.

Registering a detector

A plugin module calls register_detector at import time, passing a dotted module:Class path rather than the class itself — the registry loads lazily, so a plugin that depends on a heavy framework costs nothing until its detector is actually selected.

# plugins/my_module.py
from anomalybench.analytics.detectors import register_detector

register_detector("my_custom_detector", "my_package.detectors:MyDetector")

The detector class subclasses BaseDetector, implements fit and score, and declares its score_orientation:

from anomalybench.analytics.base import BaseDetector


class MyDetector(BaseDetector):
    score_orientation = "higher_is_more_anomalous"

    def fit(self, data, **params):
        ...
        return self

    def score(self, data):
        ...

Leaving score_orientation at its estimator_defined default means benchmark evaluation will refuse to score the detector. Declare it.

Loading plugins

benchmark-cli --plugins plugins.my_module --detectors my_custom_detector

--plugins imports the named modules before resolving detectors, which is when their register_detector calls run.

Collision protection

The built-in registry is frozen once the shipped detectors are registered, and re-registering an existing key raises ConfigurationError naming the path already bound to it. Shadowing a built-in has to be deliberate:

register_detector("knn", "my_package:MyKNN", allow_override=True)

This exists because a plugin that silently replaced isolation_forest would produce a benchmark whose leaderboard rows are indistinguishable from ones produced by the real thing.