Days that stand out
Unsupervised anomaly leaderboard
Live CenDTect-style DBSCAN over a rolling 45-day window of per-country OONI feature vectors. Higher anomaly score = farther from this country's recent normal. Useful as a second opinion on the supervised classifier — DBSCAN surfaces shape-anomalous days the labels never saw.
AUC vs labels
0.6575
Promote floor
0.65
Passed
YES
n scored
3921
| # | Country | Anomaly score | Block rate | Outage rate | n measurements | As of |
|---|---|---|---|---|---|---|
| 1 | United States (US) | 7.678 | 100.0% | 0.0% | 162 | 2026-11-18 |
| 2 | DE (DE) | 7.611 | 100.0% | 0.0% | 146 | 2026-11-18 |
| 3 | China (CN) | 7.371 | 33.3% | 66.7% | 12 | 2026-11-17 |
| 4 | CA (CA) | 7.329 | 100.0% | 0.0% | 97 | 2026-11-18 |
| 5 | Saudi Arabia (SA) | 7.287 | 23.5% | 76.5% | 17 | 2026-11-17 |
| 6 | NL (NL) | 7.266 | 100.0% | 0.0% | 49 | 2026-11-18 |
| 7 | South Korea (KR) | 7.126 | 100.0% | 0.0% | 67 | 2026-11-18 |
| 8 | Brazil (BR) | 6.978 | 100.0% | 0.0% | 112 | 2026-11-18 |
| 9 | Russia (RU) | 6.922 | 12.5% | 87.5% | 32 | 2026-11-17 |
| 10 | GB (GB) | 6.865 | 100.0% | 0.0% | 54 | 2026-11-18 |
| 11 | AU (AU) | 6.722 | 100.0% | 0.0% | 60 | 2026-11-18 |
| 12 | ZA (ZA) | 6.482 | 100.0% | 0.0% | 62 | 2026-11-18 |
| 13 | ES (ES) | 6.169 | 100.0% | 0.0% | 45 | 2026-11-18 |
| 14 | Cameroon (CM) | 5.899 | 50.0% | 50.0% | 8 | 2026-11-17 |
| 15 | JP (JP) | 5.671 | 100.0% | 0.0% | 54 | 2026-11-18 |
| 16 | Mexico (MX) | 5.668 | 100.0% | 0.0% | 52 | 2026-11-18 |
| 17 | SG (SG) | 5.639 | 100.0% | 0.0% | 91 | 2026-11-18 |
| 18 | FR (FR) | 5.405 | 100.0% | 0.0% | 8 | 2026-11-18 |
| 19 | Tanzania (TZ) | 5.234 | 80.0% | 20.0% | 5 | 2026-11-17 |
| 20 | Hungary (HU) | 5.041 | 21.1% | 79.0% | 19 | 2026-11-17 |
| 21 | Sweden (SE) | 4.769 | 100.0% | 0.0% | 7 | 2026-11-18 |
| 22 | Egypt (EG) | 4.448 | 80.0% | 20.0% | 5 | 2026-11-17 |
| 23 | India (IN) | 4.399 | 100.0% | 0.0% | 45 | 2026-11-18 |
| 24 | Israel (IL) | 4.258 | 40.0% | 10.0% | 10 | 2026-11-16 |
| 25 | Italy (IT) | 3.900 | 57.1% | 42.9% | 7 | 2026-11-17 |
| 26 | Slovenia (SI) | 3.672 | 80.0% | 20.0% | 5 | 2026-11-16 |
| 27 | Ukraine (UA) | 3.049 | 57.1% | 42.9% | 7 | 2026-11-17 |
| 28 | Zambia (ZM) | 2.910 | 100.0% | 0.0% | 13 | 2026-11-16 |
| 29 | Indonesia (ID) | 2.683 | 66.7% | 33.3% | 6 | 2026-11-17 |
| 30 | Kenya (KE) | 2.516 | 70.0% | 5.0% | 20 | 2026-11-16 |
Scroll to see every column.
GET /v1/anomaly/dbscan/leaderboard
Raw JSON of the same data, with pagination
DBSCAN finding
Methodology, eval methodology, honest scope, comparison vs supervised classifier
Why two models? The supervised v3.3 classifier learns from labeled censorship patterns but cannot recognize patterns it never saw labels for. The unsupervised DBSCAN surfaces “weird days” without training labels, so its agreement with labeled incidents answers a different question. Read the linked finding for the dated evaluation and its limits.
Generated: 2026-11-18T04:24:37.312894Z