High-accuracy splice site recognition based on machine learning is the key to eukaryotic genome annotation. In this paper, we used chi-square test to determine the window size of sequences, and constructed a chi-square statistical difference table to extract the positional features, and combined with the frequencies of dinucleotides to characterize sequences. For the problem that the positive and negative samples of splice sites are extremely imbalanced, 10 SVM classifiers based on the equal proportion of positive and negative samples were built for weighted voting, which effectively solved the imbalanced pattern classification problem. Independent testing results in HS³D dataset showed that the prediction accuracy of donor and acceptor sites were 93.39% and 90.46% respectively, obviously higher than that of the compared methods. The positional features based on the chi-square statistical difference table can effectively characterize DNA sequences, and have application prospects in signal site recognition of molecular sequences.