Support Vector Machine (SVM)

Support Vector Machine finds the widest-margin hyperplane that separates classes, for classification and regression.

The closest points to that boundary are support vectors; they fix the plane's position. If a flat boundary cannot separate the data, the kernel trick maps inputs into a higher-dimensional space where a linear separator exists, then the decision maps back. SVM handles high-dimensional data with relatively few training examples.

In text classification each word can be a dimension, thousands of features, and SVM often avoids overfitting. In bioinformatics it separates disease-related gene expression from healthy patterns.

Because the model depends only on support vectors, it stays compact and fast at prediction. On smaller datasets, SVM can still beat deep nets that lack enough data to generalize. Finance and vehicle stacks still use it for fraud detection, image recognition, and sensor fusion.

Maximum margin means the plane sits as far as possible from both classes, with support vectors defining that gap. Kernels such as RBF map points so a linear separator in the feature space corresponds to a curved boundary in input space. Text bags with thousands of word dimensions are a classic fit: few documents, many features, SVM resists overfitting. Gene-expression matrices look similar.

Prediction uses only the support vectors, so the stored model is small. On modest datasets this still competes with deep nets. Fraud, vision, and sensor-fusion stacks still ship SVM where data is not internet-scale. SVMs (1995) find a maximum-margin split. They dominated many tabular and text tasks before deep nets scaled.