A fitted support vector machine can score a new observation by summing contributions from its support vectors and adding the bias. Each contribution combines a dot product with the support vector’s class label and fitted coefficient. section 8.5 section 8.5.2
The score sums over support vectors; αᵢ weights each vector, yᵢ gives its class sign, and b is the bias. section 8.5.2
A positive score predicts class +1; a negative score predicts class −1. For a linear SVM, this form gives the same score as using the explicit weight vector, while making the training observations’ influence visible. It can also avoid constructing that vector when the feature space is too large for an explicit representation. section 8.5 section 8.5.2
Two support vectors
With support vectors (1, 0) and (0, 1), labels +1 and −1, coefficients 1, and bias 0, the new observation (2, 1) receives contributions 2 and −1. Its score is 1, so it is assigned to class +1—the same result as the direct form. section 8.5.2
