基于代价敏感支持向量机的推荐系统托攻击检测方法

基于标准支持向量机的托攻击检测方法不能体现由于用户误分代价不同对分类效果带来的影响，提出了一种基于代价敏感支持向量机的托攻击检测新方法，该方法在代价敏感性学习机制下引入支持向量机作为分类工具，对支持向量机输出进行后验概率建模，建立了基于类别隶属度的动态代价函数，更准确地反映不同样本的分类代价，在此基础上设计了代价敏感支持向量机分类器。将该分类器应用在推荐系统托攻击检测中，并与标准的支持向量机方法、代价敏感支持向量机方法进行比较，实验结果表明，本方法可以更精确地控制代价敏感性，进一步提高对攻击用户的检测精度，降低总体的误分类代价。

Shilling attack detection method of recommender systems based on cost sensitive SVM

The shilling attack detection method based on traditional Support Vector Machine (SVM) can not reflect the influence of the great difference of misclassification cost on classification effect. The paper proposes a novel shilling attack detection method based on cost sensitive SVM. The method uses SVM as the classification tool under the cost sensitive learning mechanism, and models the probabilistic outputs of SVM, presents dynamic misclassification cost function based on class membership value, thus accurately reflecting the misclassification cost of different samples. Hence, we design the classifier of cost-sensitive SVM and apply it in shilling attacks detection. Experimental results show that, compared with traditional SVM and different classes of cost sensitive SVM (CS SVM), our proposed method can more precisely control the cost of misclassification, improve the precision of the minority class samples, and decrease the total cost of misclassification.