Glycosylation is a major modification process in post-translational modification of protein. Accurate prediction of O-linked glycosylation sites is a big challenging faced by machine-learning, for the fixed-model of O-linked glycosylation is not yet known. In this paper, on the basis of the largest-ever Steentoft database up to now, a new feature——chi-square score difference table method based on position (χ²-pos) was first proposed, which combined with pseudo position-specific scoring matrix (PsePSSM) and undirected composition of k-spaced amino acid pairs (Undirected-CKSAAP) were used to present protein sequences. Then 5 support vector machines models were constructed with the same proportion of positive and negative samples. At last, by weighted voting, our results showed that the prediction accuracy, Matthew’s correlation coefficient and area under ROC curve reached 89.62%, 0.79 and 0.96 respectively. They were superior to the literature report. It also demonstrated that the combination of three different features χ²-pos, PsePSSM and Undirected-CKSAAP has extensive application prospect in protein sites prediction such as glycosylation and phosphorylation.