Calculation Result
Formula and Steps
Calculation History
Formula Used
A linear SVM uses f(x) = wᵀx + b. The sign selects a class. The magnitude measures boundary separation.
A kernel SVM sums weighted support-vector similarities. The intercept shifts the final decision. Kernel scores are not probabilities.
Signed distance equals f(x) / ||w|| for linear models. Margin width equals 2 / ||w||. Zero norm prevents distance calculation.
How to Use
Select the correct model mode. Enter vectors using commas or spaces. Keep every vector dimension consistent.
Choose preprocessing matching model training. Review warnings before trusting results. Then inspect contributions and margins.
Use examples for quick testing. Export results after validation. Calibrated probabilities require a separate method.
Example Data
| Mode | Input | Model | Expected use |
|---|---|---|---|
| Linear | 2.4, 1.3 | w = 0.8, -0.5; b = -0.2 | Feature contributions and distance |
| RBF | 1.2, 0.7 | Three support vectors | Kernel similarity contributions |
| One-vs-Rest | 1.5, 0.8 | Three class models | Class score comparison |
Frequently Asked Questions
What does a positive SVM score mean?
It usually selects the configured positive class. Threshold settings can change that outcome. Class orientation must match training.
Is the decision score a probability?
No, it is an uncalibrated model score. Its magnitude reflects relative separation. Use calibration for probability estimates.
What does a score near zero mean?
The sample lies near the decision boundary. Small perturbations may change classification. Review preprocessing and model stability.
Why must vector lengths match?
Each weight corresponds to one feature. Different lengths break the dot product. Kernel vectors also require equal dimensions.
How does gamma scale work?
Scale uses feature count and observed variance. This calculator estimates variance from support vectors. Training data may differ.
What are support-vector contributions?
Each support vector adds a weighted similarity. Large magnitudes influence the score strongly. Signs determine direction.
When is signed distance available?
It is directly available for linear models. Linear kernels can reconstruct an equivalent weight vector. Other kernels lack one global hyperplane.
How are multiclass ties resolved?
One-vs-One first compares vote counts. Score totals can break remaining ties. Labels can provide deterministic fallback ordering.
Why does preprocessing matter?
SVM coefficients expect training-time feature scaling. Different scaling changes every decision score. Reuse identical preprocessing parameters.