One Class SVM Score — Machine Learning/Anomaly Detection

One Class SVM Score

Machine Learning / Anomaly Detection

Scores a sample against a fitted one-class SVM – a boundary drawn around what the training data looked like – and says whether it falls outside:

score = intercept + Σ_i coef_i·K(sv_i, u), and outlier = 1 when score < 0, 0 otherwise.

This is the same kernel machine as SVM Predictor, read differently: the sign is the decision and the magnitude is how far inside or outside the boundary the sample sits. It answers "does this look like the data the model was trained on?", which is the question to ask before trusting any other model's output on the same sample.

Ports

• u – the feature vector, a column [d,1], where d must equal the number of columns of Support Vectors.
• score – a scalar [1,1]: signed distance to the boundary in kernel units. Positive is inside (normal), negative is outside (novel), and the magnitude is the margin.
• outlier – a scalar [1,1], 1 or 0: the decision on its own, for a monitor that wants to count, average or latch it. Emitted as a number rather than a boolean, so no backend needs a type the others lack.

Parameters

Identical to SVM Predictor's, and read the same way – the two blocks share one implementation, so a model pasted into either means the same thing.

• Support Vectors – an [N,d] matrix, one support vector per row: scikit-learn's support_vectors_, so it pastes in without transposing.
• Dual Coefficients – N values, scikit-learn's dual_coef_, which already carry their sign.
• Intercept – ⚠ a scalar, and for a one-class model this is −offset_, not offset_. scikit-learn reports an offset it subtracts; this block adds an intercept, so paste the negated value and the sign convention then matches decision_function exactly – positive inside, negative outside.
• Kernel – RBF (Gaussian) (the default, and what almost every fitted one-class SVM uses), Linear, or Polynomial. See SVM Predictor for the three formulas.
• Gamma, Coef0, Degree – the kernel's own parameters, as on that block.
• Sampling Time (s) – zero or less inherits the solver's rate; a positive value runs the block at that period.

Code export

All ten targets: Python, MATLAB, Java, Rust, C, C++, VHDL, Verilog, SystemVerilog and PLC Structured Text. The whole model is baked in at export time; there is no tunable parameter.

Linear and Polynomial export to genuine Q16.16 fixed point on the three HDL targets – they are multiply-accumulate only, and the linear kernel additionally folds to a single weight vector. RBF has an exponential, so its three HDL bodies are simulation-only real arithmetic that quantizes at the port boundary. Since RBF is the default here and by far the common case for a one-class model, expect this block's HDL rows to be the simulation-only ones in practice.

Simulink bridge

None. One-class SVM prediction belongs to the Statistics and Machine Learning Toolbox, which is not installed on this machine, and its predict blocks take a fitted model object that no parameter mapping could carry. The bridge reports the block rather than dropping it silently, and it has no parity testbench, which is the documented consequence of Support::None. Code export verification still covers it across all ten languages, in every kernel.

Notes

• Algebraic and stateless, and deliberately carries no state space: the outlier flag is a discontinuous function of the input, which no A/B/C/D describes.
• The flag is discontinuous at the boundary, so the three HDL targets can disagree with the simulation on a sample whose score sits within one Q16.16 quantum (about 1.5e-5) of zero. Threshold the score downstream in floating point if that matters – and note that a sample sitting exactly on a novelty boundary is, by construction, one the model has no opinion about.
• Pairs with the cheaper monitors rather than replacing them. Input Domain Monitor asks whether each feature is in the range it was fitted on, and Mahalanobis Distance asks how far the sample is from the mean under one covariance. This block draws an arbitrary, possibly non-convex boundary, which is what it is for – and it costs N kernel evaluations per sample to do it.