Abstract:

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 propose 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sensitive support vector machine is conducted to train the reconstructed dataset. Finally, the detection decision function is obtained. Experimental results show that the proposed method can improve the accuracy of detecting shilling attacks and has a strong generality.

Key words: recommender system；shilling attack；unbalanced data；costsensitive learning； undersampling；support vector machine