When using kernelized SVM, the kernel function must be specified. Common choices for kernels include:
- Linear: The linear kernel is the simplest choice and works best when classes can be linearly separated.
- Polynomial: The polynomial kernel is a possible choice when data is not linearly separable and maps the data into higher dimensional space by taking a power of the original data. It requires the degree of the polynomial to be specified.
- Radial Basis (RBF): The RBF kernel projects the original data into infinite-dimensional space and is a common choice for non-linear decision boundaries. It requires a gamma parameter that controls the influence carried by individual observations. It is often the default choice used when there is no prior knowledge about the decision boundary.
- Sigmoid: The sigmoid kernel represents the original feature space in the same way as a neural network perceptron model with a tanh activation function, but in complex non-linear decision boundaries, the RBF kernel is usually preferred.