摘要
推荐系统托攻击检测面临数据不均衡和代价敏感两个问题,但目前的检测方法缺乏同时对这两个问题的研究。提出了一种基于重采样和代价敏感支持向量机相结合的托攻击检测新方法。该方法首先利用基于样本重要性的欠采样技术实现训练样本的均衡,重构过程中根据边界样本对分类支持的重要性的不同加以处理,在消除大量噪声样本的同时保留了绝大多数对分类学习有用的样本;然后引入代价敏感支持向量机对重构后的样本集进行训练,最终得到系统决策函数。实验结果表明,本方法能提高对托攻击的检测精度,具有较强的推广意义。
Detecting shilling attacks for recommender systems is necessary to solve problems such as imbalanced dataset and cost sensitive, but existing methods are lack of relative studies. This paper pro- pose a new attack detection method, which combines the methods of under-sampling and cost-sensitive support vector machine together. Firstly, according to the different importance for classification to process, the training dataset is balanced by sample importance based under-sampling technique, for the sake of eliminating a lot of noise samples while retaining the most of useful samples. Secondly, cost-sen- sitive support vector machine is conducted to train the reconstructed dataset. Finally, the detection deci- sion function is obtained. Experimental results show that the proposed method can improve the accuracy of detecting shilling attacks and has a strong generality.
出处
《计算机工程与科学》
CSCD
北大核心
2013年第8期174-179,共6页
Computer Engineering & Science
基金
国家社会科学基金青年项目(11CJY008)
辽宁省社会科学规划基金项目(L10BJL026)
中央高校专项科研基金(DUT10RW302)
关键词
推荐系统
托攻击
不均衡数据集
代价敏感学习
欠采样
支持向量机
recommender systeml shilling attack
unbalanced data
cost-sensitive learning
under-sampling
support vector machine
