The Study Of Robust Recommendation Approach Based On The Measurement Of Suspicious Users

Collaborative filtering recommender systems can effectively solve the the problem of information overload,which have been widely applied to many fields,e.g.,e-commerce sites.However,in order to manipulate the recommendation results,some malicious users inject a large number of fake profiles into the systems' rating database to interfere with the decision-making recommendation process.The behavior has seriously affected the recommendation quality and the user's trust in recommender systems.Therefore,how to protect recommender systems from malicious attacks and provide users with reliable recommendation results has become a hot topic to be studied.In this paper,based on the thought of suspicious users measurement,we carry out some deep research from both memory-based and model-based recommendation technology,and design a series of high robustness and less loss of accuracy collaborative filtering recommendation algorithms.Firstly,aiming at the problem that the neighbor choosing reliability of the user-based recommendation algorithm is not high,a robust recommendation algorithm based on k-distance and item category information is proposed.By referring to the idea of outlier detection,the user suspicion degree can be calculated.Combining the user suspicion degree with item category information,we give a missing value imputation computing method to fill the unrated items.Based on the rating matrix which has been filled,the robust recommendation for target user is generated by incorporating conventional user-based collaborative filtering technique with the choosing neighbors according to the user's suspicion degree and similarity.Secondly,aiming at the problem that the existing computational models of trust can not measure the trust relationship between users accurately,a robust recommendation algorithm based on suspicious users measurement and multidimensional trust is proposed.According to the user profile features,the relevance vector machine classifier is established which can identify and measure the suspicious users.We mine the implicit trust relation among users based on the user-item rating data,and construct a reliable multidimensional trust model by integrating the user suspicion information.The reliable recommendation for target user is generated by combining the reliable multidimensional trust model and the neighbor model.Thirdly,aiming at the problem that the robustness of the recommendation model based on matrix factorization is relatively poor against shilling attacks,a robust recommendation algorithm based on fuzzy kernel clustering and support vector machine is proposed.According to the high correlation characteristic between attack profiles,we use fuzzy kernel clustering method to cluster user profiles in high-dimensional feature space,and the attack profiles can be gathered in the same cluster.In order to identify the attack profiles further,we use support vector machine classifier to classify the cluster including attack profiles.The robustness of algorithm is improved by combining the attack detection results with the matrix factorization process.Fourthly,aiming at the problem that the robustness of the recommendation model based on matrix factorization acquires at a cost of recommendation accuracy,a robust recommendation algorithm based on suspicious users identification and Tukey M-estimator is proposed.According to the distribution of users' ratings,the computational methods of deviation degree about the number of ratings and average similarity of neighbors are given,which can identify the suspicious users.We combine the identification results of suspicious users with the conventional user-based collaborative filtering technique to construct a reliable neighbor model.In the process of matrix factorization,Tukey M-estimator is introduced to construct robust matrix factorization model.The reliable neighbor model and robust matrix factorization model are combined to improve the recommendation accuracy and robustnessLastly,we conduct the comparative experiments between the proposed approaches and others on Movielens dataset,which demonstrate the effectiveness of the proposed approaches.