A polynomial kernel lets an SVM represent powers and interactions among original features without explicitly constructing them. The SVM fits a linear hyperplane in the resulting feature space; in the original space, that boundary can appear curved. section 8.6.2
The degree d controls the order of represented powers and interactions; c affects the influence of lower-degree terms. section 8.6.2
A second-degree kernel can represent squared features and pairwise interactions. Higher degrees allow more complex interactions, but can make the model more sensitive to training-data irregularities. Polynomial transformations affect the feature space globally, so boundary changes generally extend across the data rather than staying local. section 8.6.2 section 8.6.3
The degree also interacts with the soft-margin penalty C. Treat degrees and other model settings as candidate configurations, and compare them through cross-validation rather than training accuracy alone. section 8.6.3

