Research On Detection Method For Shilling Attacks Based On User Rating Behaviors In Recommender Systems

As an effective approach to alleviate the problem of "information overload",collaborative filtering recommender system is widely used in the e-commerce field,which plays a very important role in the improvement of website traffic,conversion ratio of commodity,and customer loyalty.However,collaborative filtering recommender system has its vulnerability,and malicious users can deliberately inject a large number of fake ratings into the system to influence or manipulate the recommendation results for their interests.Therefore,it is an urgent problem to detect various shilling attacks that threaten the security of recommender system,in order to ensure recommendation quality and guarantee the credibility of recommendation results.Based on the user rating behaviors,we carry out some deep research on multi-source features extraction,multi-view ensemble detection,automatic detection,user relationship graph-based detection,and so on.Firstly,aiming at the problem that the detection features extracted from single source cannot fully character the user behaviors,a multi-source features extraction method is proposed based on user ratings.In detail,based on the idea of information fusion,the temporal popularity of item is defined,and then the wavelet transform method is utilized to filter the noises and unstable signals.From the fusion view of item popularity and rating time,4 user features are extracted.From the rating inter-arrival time,2 user features are extracted based on corrected conditional entropy and “rest-rating” model.From the fusion view of rating values and rating time,2 user features are extracted.From different popular item sets,2 user features are extracted.Secondly,in view of the redundant detection features and the unbalanced classification problem in supervised methods,multiple detection views are constructed by optimal feature set partitioning method based on the above multi-source features,and then multi-view ensemble method is proposed base on multiple kernel learning,which can automatically determine the weight of each classification view.At the same time,a feature set repartition strategy is introduced for increasing the diversities of base classifiers.Thirdly,an automatic detection method is proposed based on marginalized stacked linear denoising autoencoder for the problem of high knowledge cost and poor generaliza-tion ability in traditional detection methods.In marginalized denoising autoencoder,different corruption rates for items are calculated according to the ratings' distribution in the common attack models.Also,the ratings sparsity is used to weight the mapping matrix to extract low-dimensional representation.And then the frame of marginalized stacked linear denoising autoencoder is designed for automatic features extraction.According to the robust and low-dimensional features,an AdaBoost-based detection method is proposed.Fourthly,an unsupervised detection method based on users' relationship graph is proposed for the problem that some prior knowledge needed in traditional unsupervised detection methods cannot be acquired and the detection accuracy is not high.According to the number of co-rated items,deviation degree of filler rate,difference degree of ratings preferences,the calculation method for edge weight in users' graph is proposed.And then the stacked denoising autoencoders are used to extract the graph features in order to reconstruct the users' graph.In the reconstructed users' graph,a shilling attack detection method is proposed based on community discovery algorithm and community attributes.Finally,on Netflix and Amazon datasets,the experiments are carried out to evaluate the proposed methods in detecting various simulated shilling attacks and actual attacks.The results are compared with the existed methods to verify the effectiveness of the proposed methods.