tadmetric 0.2.2

A lightweight evaluation toolkit for time-series anomaly detection

tadmetric

Lightweight evaluation utilities for time-series anomaly detection.

tadmetric is now centered on one small class, Tadmetric, for the workflow:

1. Load anomaly scores and binary labels
2. Evaluate a threshold with one of the built-in modes
3. Search for the best threshold when needed

Installation

pip install tadmetric

Python 3.9+ is supported.

Quick Start

from tadmetric import Tadmetric

score = [0.1, 0.2, 0.4, 0.9, 0.7, 0.1, 0.0]
label = [0, 0, 1, 1, 1, 0, 0]

tm = Tadmetric(score, label)

point = tm.evaluate(0.5, mode="point-wise")
adjusted = tm.evaluate(0.5, mode="point-adjusted", calc_latency=True)
composite = tm.evaluate(0.5, mode="composite")

print(point.asdict())
print(adjusted.latency)
print(composite.f1)

Modes

• point-wise: regular point-level precision, recall, and F1
• point-adjusted: if an anomaly event is detected once, the whole event is credited
• composite: point-wise precision with event-wise recall

The mode argument accepts a few aliases such as point, point_wise, point_adjusted, and comp.

Threshold Search

Use search() when you want the best threshold over a threshold grid:

from tadmetric import Tadmetric

tm = Tadmetric(score, label)
best = tm.search(mode="point-adjusted")

print(best.threshold)
print(best.f1)
print(best.result.asdict())

By default, search() scans steps=100 thresholds from score.min() to score.max(). If multiple thresholds tie, it prefers the higher threshold.

best = tm.search(start=0.0, end=1.0, steps=101, mode="point-wise", verbose=False)

Returned Objects

evaluate() returns a MetricResult with:

• f1, precision, recall
• tp, tn, fp, fn
• latency for point-adjusted evaluation when requested
• detected_events, total_events for event-aware modes

search() returns a SearchResult with the best threshold and the corresponding MetricResult.

Validation Rules

• score and label must be 1D and non-empty
• score must be finite numeric values
• label must be binary values 0 or 1
• score and label must have the same length

Development

pip install -e .[dev]
pytest -q