Fig. 4

In this toy illustration, the hyperlines (dashed lines) of Support Vector Machine (SVM) separate cocaine-addicted (red circles) from healthy control (blue triangles) participants via two features F₁ and F₂. Left panel: The kernel function, which maps a data point to another dimension, is in the form of Φ(x). Φ(x^'), which produces a linear decision boundary. Right panel: The separating line is non-linear, since a polynomial kernel, ((Φ(x). Φ(x^'))⁴, is used to map the data. In this case, the decision boundary is non-linear, placing more cocaine-addicted participants in the correct regions. For instance, two of the cocaine-addicted participants and one healthy participant pointed with green arrows are misclassified with a linear kernel. Once trained with a polynomial kernel, the decision boundary is more flexible resulting in fewer misclassified participants. The use of the polynomial kernel increases the accuracy