emperor-mew/voidly-anomaly-dbscan-v1

Voidly Atlas Unsupervised Anomaly (CenDTect-style DBSCAN) v1

Version: v1 | Trained: 2026-05-21T04:19:40.634338Z | License: CC BY 4.0

Per-country rolling 45-day window, DBSCAN(eps=75th-pct kNN, min_samples=3) over 12 standardized OONI features. Promoted as a second-opinion signal — the supervised v3.3 classifier still dominates at AUC 0.99; DBSCAN surfaces shape-anomalous days that labels never saw.

Eval

Metric	Value
auc	0.6506
auc_pr	0.3639
precision_at_p90	0.4326
recall_at_p90	0.1662
n_scored	3922
n_positive	1023
n_negative	2899
promote_floor_auc	0.6500
window_days	45

Features

• block_rate
• log_n
• pct_dns_block
• pct_tcp_reset
• pct_blockpage
• pct_tls_reset
• pct_outage
• pct_interference
• pct_block
• asn_unique_log
• source_diversity
• asn_unique_normalized

Honest caveats

• AUC 0.65 — just above the 0.65 promote floor. Not a primary signal.
• Use as a second opinion alongside the supervised v3.3 classifier (AUC 0.99).
• Labels are imperfect ground truth — DBSCAN may surface real anomalies that don't have a label.

Reproducibility

Paper: Aceto & Pescape 2025 — CenDTect (https://arxiv.org/abs/...)

Citation

@misc{voidly_voidly_anomaly_dbscan_v1,
  title  = {Voidly Atlas: voidly-anomaly-dbscan-v1 (v1)},
  author = {Voidly},
  year   = {2026},
  url    = {https://huggingface.co/emperor-mew/voidly-anomaly-dbscan-v1},
  note   = {Open censorship-research ML stack. CC BY 4.0.}
}

Method foundation: CenDTect (Aceto & Pescape 2025)