telemeval 0.3.1

Leakage-safe, event-wise and affiliation-based evaluation for spacecraft-telemetry anomaly detection

telemeval

Leakage-safe, event-wise and affiliation-based evaluation for spacecraft-telemetry anomaly detection.

Status: v0.x — early releases; the API may still change until v1.

Time-series anomaly-detection evaluation is notoriously easy to get wrong: point-adjusted F1 can rank random predictions above real detectors, and subtle protocol bugs (like scoring training-window events against test-window predictions) silently inflate or deflate results. The metrics the literature recommends instead — corrected event-wise F-beta and affiliation-based precision/recall — have not had a maintained, permissively-licensed, pip-installable home.

telemeval is that home:

• A validated ingestion contract that raises typed, actionable errors instead of producing a number from a leaky or malformed evaluation — including a first-class train/test-window leakage guard.
• Event-wise precision / recall / F-beta with unambiguous, documented overlap semantics (recall matches ESA-ADB's event-wise recall exactly; precision is telemeval's run-based definition — the differences from ESA-ADB's TNR-corrected EW_* are documented and pinned by divergence tests, and an ESAScores-parity metric is on the roadmap).
• Affiliation-based precision / recall — the canonical reference implementation (Huet et al., KDD 2022, MIT) vendored, wrapped, tested, and maintained here.
• ADTQC detection-timing quality (ESA-ADB) — scores when each event was first caught, not just whether; previously available only inside ESA-ADB's research fork.
• Channel- and subsystem-aware F-beta (ESA-ADB) — did you flag the right source? Verified against the reference test suite's exact expected values.
• Honest reports: deterministic JSON and Markdown output stamped with explicit scope caveats.
• Telemetry-aware inputs: channel-keyed predictions (interval labels, binary masks, or continuous scores + threshold), an ESA-ADB-format loader, and a TimeEval-format reader.
• scikit-learn-style metric wrappers so the metrics drop into existing scoring code.

Core dependencies: numpy and pandas. Nothing else.

Quick start

import pandas as pd
from telemeval import evaluate

labels = pd.DataFrame(
    {
        "ID": ["anomaly_1"],
        "Channel": ["channel_41"],
        "StartTime": ["2024-01-01T00:02:00"],
        "EndTime": ["2024-01-01T00:03:00"],
    }
)
timestamps = pd.date_range("2024-01-01", periods=6, freq="1min")
predictions = {
    "channel_41": pd.DataFrame({"Timestamp": timestamps, "Score": [0, 0, 1, 1, 0, 0]})
}

result = evaluate(labels, predictions, dataset="my-mission")
print(result.metrics["event_wise"]["event_wise_fbeta"])     # 1.0
print(result.metrics["affiliation"]["affiliation_fbeta"])   # 1.0
result.save(json_path="report.json", markdown_path="report.md")

See docs/usage.md for the leakage guard, continuous scores + threshold, ESA-ADB and TimeEval-format loaders, parquet input, sklearn-style wrappers, and the metric registry. See docs/related-work.md for an honest map of prior art and when to use which tool.

What telemeval is not

• Not a detector library (see PyOD, aeon, darts).
• Not a benchmark harness or dataset collection (see TimeEval, TSB-AD).
• Not a serving/monitoring/MLOps stack.
• Not affiliated with or endorsed by ESA; it does not redistribute the ESA Anomaly Dataset.

Prior art worth knowing: TSADmetrics (GPL-3.0, generic time-series AD metrics), TSB-AD (benchmark suite), Merlion (point-adjusted F1), aeon (range/VUS metrics). telemeval's niche is the permissive license, the telemetry-domain contract with leakage guards, and maintained affiliation metrics; where metrics overlap we aim to reproduce prior-art numbers.

License

Apache-2.0. Vendored affiliation reference implementation is MIT (retained; see NOTICE). Dataset licenses (e.g. ESA Anomaly Dataset, CC BY 3.0 IGO) are separate from this code and are never bundled.