Anomaly Detection: Finding Needles in Haystacks

Anomaly detection suffers from severe class imbalance by definition. Standard classifiers fail here.

Home / Guides / Anomaly Detection Models

Isolation Forests

The go-to algorithm for tabular anomaly detection. It works by isolating anomalies (which are few and different) by randomly selecting a feature and a split value. Anomalies are isolated closer to the root of the tree.

Autoencoders

For high-dimensional data (images, complex sequences), train a neural network to compress and reconstruct normal data. Anomalies will have high reconstruction error. Overkill for simple tabular data.

Internal Resources