Assumptions

1. Data is (approximately) separable

   • SVM assumes that classes can be separated by a hyperplane, either in original space (linear) or transformed space (non-linear with kernels).

2. Large-margin principle

   • Assumes that the best decision boundary is the one that maximizes margin between classes.

3. Kernel appropriateness

   • If data is non-linear, assumes the chosen kernel (RBF, polynomial, etc.) maps data to a space where separation is possible.

4. Independent and identically distributed (i.i.d.) data

   • Training and test samples come from the same distribution and are independent.

5. Balanced scaling of features

   • Assumes input features are normalized/scaled, since SVM relies on distance calculations.

6. Limited noise and outliers

   • Assumes data is not heavily noisy, since outliers close to the margin can affect the boundary.