What is outlier detection?
Outlier detection, also called anomaly detection, finds the samples that stand apart from the bulk of a dataset. Those outliers can be problems to fix, like corrupted images or mislabeled examples, or valuable rarities, like an unusual scenario worth adding more of. Either way, they are the samples most worth a human's attention.
It is often performed in an embedding space, where distance captures how unlike the rest a sample is.
Key takeaways
- Outlier detection surfaces samples far from the dataset norm.
- Outliers can be errors to fix or rare cases worth keeping.
- It is commonly done using distances in an embedding space.
How it works
Samples are mapped into a feature or embedding space, and those that sit far from dense regions, or that a model reconstructs poorly, are flagged as outliers. Techniques range from distance and density methods to reconstruction-based approaches using autoencoders. The flagged set is then reviewed rather than removed automatically, since some outliers are exactly what you want to collect more of.
Why it matters
A handful of corrupt or mislabeled samples can distort training and evaluation out of proportion to their number, and outlier detection is an efficient way to find them without inspecting everything by hand. It is equally useful for discovering the rare, long-tail cases that models struggle with.
Frequently asked questions
Are outliers always bad?
No. Some are errors to remove, but others are rare, important cases worth collecting more of. That is why flagged outliers are reviewed, not deleted blindly.
How is outlier detection related to anomaly detection?
They refer to the same idea: identifying samples that deviate significantly from normal patterns in the data.
Related terms