Abstract: Traditional support vector machine drops significantly when only a few labeled training samples is available. To address this problem, a new SVM-KNN classification method based on semi-supervised learning is proposed. In the first stage, use the few labeled training samples to train a weaker SVM classifier. And in the second stage, make use of the boundary vectors to improve the weaker SVM iteratively by introducing KNN. Using KNN classifier doesn’t enlarge the number of training examples only, but also improves the quality of the new training samples which are transformed from the boundary vectors. Then the proposed model is used to shilling attacks detection on recommender systems, the experimental results show that the proposed method can improve the classification accuracy, effective and easy to use in the case of fewer labeled training samples.

Key words: shilling attacks detection, semi-supervised learning, support vector machine, K-nearest neighbor

摘要： 针对支持向量机方法在标记用户数据不充分的情况下无法有效实现托攻击检测的不足，提出一种基于SVM-KNN的半监督托攻击检测方法。根据少量标记用户数据训练一个初始SVM分类器，利用初始SVM对大量未标记用户数据进行分类，挑选出分类边界附近有可能成为支持向量的样本点，利用KNN分类器优化边界向量的标记质量，再将重新标注过的边界向量融入训练集，迭代训练逐步改善SVM的分类边界，最终获得系统决策函数。实验结果表明在标记用户数据较少的情况下，方法能有效提高托攻击的检测精度和效率，具有较强的推广能力。

关键词: 攻击检测, 半监督学习, 支持向量机, K最近邻