Glycosylation Site Prediction (GSP) research has witnessed a growing interest in proteomics. The high ability to GSP is helpful for better understanding the function of protein, theoretically. In this research, our aim is to explore a new method for improving the performance of GSP of O-glycosylation sites. We propose to utilize Independent Component Analysis (ICA) for feature selection and dimension reduction, and then use Support Vector Machine (SVM) for glycosylation site classification, in which our method is applied for two kinds of datasets in glycosylated site and non-glycosylated site. Compared with using other subspace-based method and SVM method, experimental results show that our new approach is feasible and effective with higher prediction accuracy.