qhjqhj00/research-skills-pool/anomaly-detection-eval

anomaly-detection-eval

Sintel: A Machine Learning Framework to Extract Insights from Signals — Alnegheimish et al. (2022) (arXiv:2204.09108, 2022)

What this evaluates

Evaluates unsupervised and semi-supervised time series anomaly detection pipelines across multiple real-world and benchmark datasets. It measures how well different models identify known anomalous segments in telemetry, production traffic, and synthetic signals.

Datasets

• NAB — total 45; splits: test (45); repo https://github.com/numenta/NAB
• NASA — total 80; splits: test (80); repo https://github.com/khundman/telemanom
• YAHOO — total 367; splits: test (367); repo https://webscope.sandbox.yahoo.com/catalog.php?datatype=s&did=70

Metrics

• F1 score (primary) — range: [0, 1]
• Calculated using an overlapping segment approach between predicted and ground-truth anomaly segments. Standard precision, recall, and F1 are reported per pipeline and dataset.

Input / output format

Input: Univariate or multivariate time series signals (e.g., spacecraft telemetry, server traffic metrics).

Output: Binary anomaly labels or detected anomaly segments for each time step/segment.

Scoring recipe

def compute_f1(pred_segments, gold_segments):
    tp = sum(1 for p in pred_segments if any(overlap(p, g) for g in gold_segments))
    fp = len(pred_segments) - tp
    fn = len(gold_segments) - tp
    precision = tp / (tp + fp) if (tp + fp) > 0 else 0
    recall = tp / (tp + fn) if (tp + fn) > 0 else 0
    f1 = 2 * precision * recall / (precision + recall) if (precision + recall) > 0 else 0
    return f1, precision, recall

Common pitfalls

• No single pipeline dominates across all datasets; performance is highly dataset-dependent.
• High false positive rates can overwhelm human reviewers, making precision as critical as recall.
• Semi-supervised pipelines initially underperform unsupervised baselines until sufficient user annotations are accumulated.

Evidence (verbatim from paper)

Table [3]. Unsupervised anomaly detection results (F1 score, precision, and recall) per pipeline on each dataset.

Citation

@misc{alnegheimish2022sintel,
  title={Sintel: A Machine Learning Framework to Extract Insights from Signals},
  author={Alnegheimish et al. (2022)},
  year={2022},
  note={arXiv:2204.09108}
}

• arXiv: 2204.09108