Expand description
A linear SVM: L2-regularized, squared hinge loss, solved in the dual by
coordinate descent with shrinking (Hsieh, Chang, Lin, Keerthi &
Sundararajan, “A dual coordinate descent method for large-scale linear
SVM”, ICML 2008: liblinear’s solver for -s 1).
The primal it solves, with x̃ = (x, bias) and w̃ = (w, w_b):
min ½‖w̃‖² + Σ Cᵢ max(0, 1 − yᵢ w̃·x̃ᵢ)², where Cᵢ is C times the
weight of the message’s class (ham weighs more: calling ham spam costs
more). The bias is a feature of its own and is regularized, as liblinear’s
-B. Rows are dense f32 (the standardized message vector and header
features); every sum is f64.
Structs§
- Cancelled
- Learning was stopped from outside.
- Model
- A learned linear SVM:
decision(x) = w·x + b, positive for spam. - Params
- How the SVM learns.
- Problem
- The rows to learn from.