The binding between peptides and MHC molecules is an important event to the cellular immunity against pathogens. The binding peptides are recognized as the epitopes, which are useful for the epitope-based vaccine design. Accurate prediction of the MHC-II binding peptides has long been a challenge in bioinformatics. Recently, most researchers are interested in predicting the binding affinity instead of categorizing peptides as "binders" or "non-binders". In this paper, we introduced a novel encoding scheme based on Locally Linear Embedding (LLE) and Wavelet Transform (WT), in which important amino acid properties were firstly selected from all properties (described in AAindex database) by using LLE, and then amino acids of peptides were replaced with these novel properties. Further, WT was adopted to extract the frequency attributes of the numerical sequences; thereby the peptides were transformed into homogeneous-length vectors. Finally, Support Vector machine Regression (SVR) was used to make quantitative prediction models based on these numerical vectors. When applied to the 16 datasets from IEDB database, our encoding scheme produced consistently better performance than other encoding schemes, indicating that our encoding scheme is an effective tool for the prediction of MHC-II binding affinity.