One-Class SVM Anomaly Boundary

Dataset shape

One-Class SVM parameters

Gamma (kernel width) 0.35

Nu (expected outlier fraction) 10%

Live metrics

Flagged anomalies

–

Injected anomalies

–

True positives

–

False positives

–

Recall

–

Threshold score

–

How it works

A one-class SVM with an RBF kernel learns a boundary that tightly wraps the bulk of "normal" data in feature space; anything falling outside is flagged as an anomaly — no anomaly examples are needed to train it. Its decision function is a weighted sum of RBF kernels centred on the training points, which is exactly the kernel density estimate rendered here as a translucent voxel volume:

score(x) = (1/N) Σᵢ exp(−γ·‖x − xᵢ‖²)
threshold = the ν-th percentile of score(xᵢ) over training points
flag x as anomaly  if  score(x) < threshold

Gamma — RBF kernel width. Low gamma → a smooth, generous boundary that hugs the overall cloud shape; high gamma → a tight, wrinkly boundary that can carve around individual points and start rejecting normal ones.
Nu — the fraction of training points the model is told to expect as outliers. It sets the threshold: raising nu shrinks the boundary and flags more points; lowering it grows the boundary and flags fewer.
Dataset shape — swap the "normal" manifold (single blob, two clusters, ring) to see the boundary conform to whatever shape the density actually has, not just a sphere.
Orange points are injected ground-truth anomalies; a red halo ring means the model flagged that point. A green point with a red halo is a false positive; an orange point with no halo is a missed anomaly (false negative).

Normal (ground truth) Anomaly (ground truth) Flagged by SVM Boundary voxels

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.