svm_model

Builtin index · Statistics & ML

result = svm_model(X, y, [kernel=], [C=1.0], [gamma=], [degree=3], [coef0=1])

Input: X, y, kernel, C, gamma, degree, coef0 · Output: result

What it is

Binary-classification support-vector machine by simplified SMO on an N-row, D-column X and a length-N target y that must take exactly two distinct values (no one-vs-rest wrapper). kernel is a string, one of "linear", "rbf", "poly"; C (number, default 1.0) is the regularization strength; gamma (number) scales the "rbf"/"poly" kernels; degree (integer, default 3) and coef0 (number, default 1) apply to "poly".

Overloads

3 forms. Each carries its own inputs, outputs and worked example — pick the tab for the one you are calling.

Xx = [0.0, 0.0; 0.0, 1.0; 1.0, 0.0; 1.0, 1.0; 0.1, 0.1; 0.1, 0.9; 0.9, 0.1; 0.9, 0.9]
Yx = [0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0]
svl = svm_model(Xx, Yx, kernel = "linear", C = 1.0)
print("linear kernel accuracy on XOR: {round(svl.score(Xx, Yx), 3)}")

linear kernel accuracy on XOR: 0.5

Xx = [0.0, 0.0; 0.0, 1.0; 1.0, 0.0; 1.0, 1.0; 0.1, 0.1; 0.1, 0.9; 0.9, 0.1; 0.9, 0.9]
Yx = [0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0]
svr = svm_model(Xx, Yx, kernel = "rbf", gamma = 2.0, C = 1.0)
print("rbf kernel accuracy on XOR:    {round(svr.score(Xx, Yx), 3)}")

rbf kernel accuracy on XOR:    1

Xx = [0.0, 0.0; 0.0, 1.0; 1.0, 0.0; 1.0, 1.0; 0.1, 0.1; 0.1, 0.9; 0.9, 0.1; 0.9, 0.9]
Yx = [0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0]
svp = svm_model(Xx, Yx, kernel = "poly", degree = 3, coef0 = 1.0, C = 1.0)
print("poly kernel accuracy on XOR:   {round(svp.score(Xx, Yx), 3)}")

poly kernel accuracy on XOR:   0.75

How it works

svm_model(X, y, kernel=, C=, gamma=) — binary-classification support vector machine, fit via svm_smo_fit's simplified SMO (Andrew Ng's CS229 "Simplified SMO" notes — a textbook dual-coordinate-ascent solver, chosen over full Platt SMO's heuristic working-set selection because it's straightforward to verify against a written reference and needs no general QP dependency). kernel="linear" (default) or "rbf" (gamma= controls width, default 1/n_features — the standard heuristic scikit-learn's own default used before "scale"); "poly" (degree=/coef0=, defaults 3/1) is a stretch-goal third kernel, exercised far less than the required linear/rbf pair. y must have exactly two distinct values (mapped to -1/+1 internally, smaller value -> -1); multi-class (one-vs-rest) wasn't built this pass — see svm_fit_model's own doc comment. The training kernel (Gram) matrix's X @ X^T cross term is computed via svm_cross_term, GPU-accelerated through the same gpu_matmul/qu_gpu::GpuContext path when this crate is built with --features gpu (plain CPU matmul otherwise) — see that function's own doc comment.

From the note above this builtin in the interpreter, where the reasoning was written down as it was built.