Research On Rating Prediction Algorithm Based On The User Embedding

The rating prediction as the key task of recommender algorithm has been widely concerned for a long time.Typically,since the recommender algorithms analyze from large-scale user-item interaction information(ratings)to obtain the items that may be of interest to a specific user,the accuracy of rating prediction task will directly affect the recommender effectiveness of the model.However,the existing rating prediction algorithms tend to call only a small portion of the huge interaction data,ignoring the potential user association information in the rating and trust data,thus causing problems such as low prediction accuracy and cold start of the model.Therefore,this dissertation addresses the issues that related to how to fully explore the potential model association relationships between users from different industries and living environments to enrich the model prediction analysis algorithm to model users,improve the prediction accuracy when the model performs the rating prediction task,and effectively alleviate the cold start problem of new users of the model to a certain amount to better meet the personalized needs of different types of users,etc.The following work was done:1.In order to further utilize explicit user feedback(ratings)to enrich user modeling,this dissertation proposes a singular value decomposition model based on the user cooccurrence.The model quantifies the potential correlation between users by using the point mutual information between different users,and obtains the user co-occurrence matrix from the user common preference matrix analysis.Based on the existing user bias and implicit parameters of singular value decomposition,the user co-occurrence matrix is added to improve the modeling accuracy of users and the accuracy of the model rating prediction.The model was tested on four datasets such as Movielens-100 K and Each Movie.2.At present,some researchers doing prediction-related work widely believe that the user groups in social information networks tend to accept the information or products recommended by their trusted friends,so the trust relationship between users in different networks has received widespread attention as the most reliable form of friend relationship expression in social information networks.In order to effectively alleviate the cold start problem of model users that is widely found in current recommender algorithms,this dissertation proposes to incorporate information with user trust into the rating prediction model to explore more correlations between different users based on the theoretical research results of user co-occurrence matrix.By combining the analysis of real user trust and history rating interaction information,not only can the accuracy of model rating result prediction data be significantly improved,but also effectively alleviate the prediction model user cold start problem.The model was experimentally measured on the Film Trust and Epinions datasets,and good experimental results were achieved.In this dissertation,through in-depth analysis and investigation of users' explicit feedback behavioral information(ratings)and socialized relationship information(trust),combined with a large number of experiments to support,the final effect basically met the experimental expectations.