| --- |
| title: outliar |
| emoji: 🎲 |
| colorFrom: blue |
| colorTo: gray |
| sdk: static |
| pinned: false |
| license: mit |
| --- |
| |
| # outliar |
|
|
| An anomaly-detection benchmark on the Numenta Anomaly Benchmark, and an audit of |
| the metric the field reports. |
|
|
| Pick a detector — including `random`, which is uniform noise and never looks at |
| the data — and drag the threshold. Point-adjusted F1, the number reported almost |
| universally in time-series anomaly detection, will tell you the noise is |
| excellent. The composite score will not. |
|
|
| Everything runs in your browser: the detectors and the four scoring protocols are |
| a JavaScript port of the NumPy implementation, pinned to it to 1e-9 by a |
| Python↔Node parity test. NAB itself is fetched from the upstream GitHub raw |
| endpoint, not redistributed here. |
|
|
| `iforest` is missing from this demo because it needs scikit-learn's fitted trees, |
| and `random` uses a JavaScript PRNG rather than NumPy's PCG64 — it is still |
| uniform noise, so the argument is unaffected, but the digits will differ slightly |
| from the repo's benchmark. |
|
|
| Code, the full 47-series benchmark and the write-up: |
| **https://github.com/UsmarHaider/outliar** |
|
|
