One-Class SVM
Learns a tight boundary around normal data only, then flags anything that falls outside it as an anomaly.
Also OCSVM · support vector data description · novelty detection SVM
Schölkopf et al., 2001
You only ever see healthy apples. You learn the shape of 'normal apple' so well that the first rotten one stands out, even though you never saw a rotten apple before.
Where it fits
+All relationships as a list
Is a kind of / part of
Uses / built on
Applies to
Techniques that apply here
How it works
A normal support vector machine separates two classes. In anomaly detection you often have only one: plenty of normal data and almost no examples of what can go wrong. A one-class SVM works with just that.
It draws the tightest reasonable boundary around the normal data. Thanks to the kernel trick, that boundary can be any blob-like shape, not just a circle. New points inside are normal; points outside are flagged.
The setting (nu) says how strict to be: lets roughly 5% of the training points fall outside, which helps when the “normal” data is a little dirty.
On large or high-dimensional datasets, an isolation forest is usually faster and easier to tune, so it is a common first choice today.
Everyday use cases
- Case 01
Machine health monitoring
Learns what normal vibration readings look like and raises an alarm when a motor starts behaving differently.
- Case 02
Intrusion detection
Flags network sessions that look unlike the normal traffic it was trained on.
Strengths, weaknesses and when to use it
Strengths
- Needs only normal examples, no labeled anomalies.
- Flexible curved boundaries through the kernel.
Weaknesses
- Sensitive to the kernel width and ν; hard to tune without some labeled anomalies.
- Scales poorly to very large datasets.
When to use itNovelty detection on small to medium datasets with a clean sample of normal behavior.
Compare with
Check yourself
- Q01
What does a one-class SVM need for training?
Show answer
B: Examples of normal data only
- Q02
What does the ν (nu) parameter control?
Show answer
B: Roughly the fraction of training points allowed to fall outside the boundary