Research On Collaborative Filtering Recommendation Algorithm Based On User Trust Network Model

With the rapid development of the Internet,information has risen sharply and the phenomenon of“information overload”has occurred,which has aggravated the difficulty of information selection.In this context,the recommended model came into being.A collaborative filtering recommendation strategy is widely used in personalized recommendation service.It is one of the most successful recommendation technologies at present,but it is limited by the sparse scoring data and the cold start problem of users,which affects the performance of recommendation system.In order to solve such problems,this paper enhances the reliability of similarity calculation by integrating the trust relationship between users into the collaborative filtering algorithm.Specifically,according to the structure of user trust network,this paper proposes six different link prediction methods based on structural similarity for the characteristics of directed networks,and then constructs a user trust network model.By analyzing the global topology of the trust network formed by link prediction,we propose to measure the user trust weight by means of trust factor,integrated with the scoring factor established from the user scoring mode to obtain the comprehensive influence of the user.Then,the weights in the scoring prediction task are determined,and a collaborative filtering algorithm based on the user's comprehensive influence is formed.Experiments show that the proposed algorithm outperforms a recent published comparative algorithm^[1]not only in time complexity but also in performance,which improves the accuracy of scoring prediction.In view of the above limitations of trust relationship prediction based on network structure,we transform the process of trust network modeling into regression problem and classification problem in machine learning.In order to establish our regression process,we adopt the implicit semantic decomposition method to transform the user direct trust matrix into the product of the trust matrix and the trusted matrix,thus determining the optimization problem with the mean square error as the target;The model is solved by the alternating least squares method and achieves the way of predicting the degree of trust between users.Finally,combining the obtained trust network model with the collaborative filtering algorithm,a collaborative filtering algorithm based on implicit semantic regression model of trusted users is formed.Through experimental comparison,our improved algorithm has further improved the recommendation effect.At the same time,in order to establish our classification process,we use embedding vectorization technology to characterize the input of trust users and trusted users,and then use the feature vector formed by above two input vectors as the input of logistic regression model,and establish the optimization problem with the maximum likelihood function as the target;The regression coefficient and the embedding matrix in model are solved by the stochastic gradient descent method and achieve the way of predicting the unknown trust relationship of any two users.Combining the obtained trust network model with the collaborative filtering algorithm,a collaborative filtering algorithm based on the implicit semantic classification model of the trusted users is formed.In terms of performance evaluation of the algorithms,compared with the comparative algorithm,the algorithm improved by 3.62%in the MAE indicator and 4.20%in the RMSE indicator,which significantly improves the performance of the recommended system.In summary,the research results of this paper fully demonstrate the technical advantages brought by the recommendation algorithm combined with user trust relationship and the effectiveness of the proposed algorithm.