As per the example availble in the documentation of 'fitcsvm' function as shared in the question, the term 'generalization rate' refers to models ability to generalize loss for unseen data.
Hence in this context, the term is used for defining the out-of-sample 'misclassification rate'. Since the 'classLoss' comes out to be 0.0484, hence in percentage terms, 'generalization rate' is given to be 'approximately 5%'.
Thanks.
