The Research Of Robust Recommendation Algorithm For Shilling Attack And Random Noise

Recommender systems arise at the historic moment in the era of big data.Recommender systems utilize the related information of users and items to predict the subsequent behavior of users.Recommender systems are different from traditional search engines that passively output information that users want.Recommender systems construct each user's interest model and actively recommend user the project which is interesting items for the user in list through utilizing three main kind of information.These information includes the individual information of users and items,the historical interactive information between users and systems and the external environment information when users interact with recommender systems.However,the man-made noise and random noise often exist in the original data sets of recommender systems,and the noise always affects the recommender system to obtain an accurate user's interest model.The man-made noise is users who inject false ratings data into recommender systems,also refered to as shillings,and this behaviour is called shilling attack.But the random noise is different from shilling,and it distuibutes in the whole ratings data set.Although traditional shilling attack detection algorithms have well detection accuracy in the face of the shillings,it has to analyze certain type of shilling attack model and to design corresponding attribute detection indexes.Thus it lacks of automation.For that Stacked autoencoder in deep learning is introduced to shilling attack detection.Firstly,the proposed algorithm normalizes each user's rating vector,and then greedy layer-wise training is implemented through inputing vectors to stacked autoencoder.Secondly,the algorithm fine-tunes the whole stacked autoencoder when after pretraining it.Thirdly,every user's feature vector will be directly outputed,and the way of end-to-end feature extraction is realized.So the stacked autoencoder is introduced into user feature extraction model in shilling attack detection,and lower degree of human involvement is gotten.Experiments on two benchmark data sets demonstrate that indexes precision,recall and F1 of the proposed algorithm results are all better than compared other algorithms.So this strategy reduces the step of analysis and design and enhances the automation of shilling attack detection.Finally shilling attack detection on stacked autoencoder reduces the impact of shilling on recommendation algorithm and makes that more robust.According to the distributed random noise in data set,the hypothesis of local low-rank is introduced into probabilistic matrix factorization(PMF),and then the enhanced algorithm local probabilistic matrix factorization(LPMF)is proposed.During the procedure of realizing classic algorithm local low-rank matrix approximation(LLORMA),that the entire rating matrix is divided into certain number of local rating matrix in a specific way makes the train data set sparser.But PMF performs well on large scale,sparse and imbalanced data set.LPMF divides the entire rating matrix into certain number of local matrices and combines these local optimal solutions in a weighted manner.So that it can alleviate the over fitting problem of PMF and low sparsity problem of LLORMA.Therefore,LPMF combines the advantages of the above two algorithms,and makes up for each other's shortcomings.And the experimental results show that the prediction accuracy is higher in different data sets.