Collaborative Filtering Recommendation System Based On Hybrid Algorithm

Collaborative filtering recommendation is the main tool of personalized recommendation system.However,collaborative filtering recommends some problems such as sparse data and cold start.In order to deal with these problems,hybrid collaborative filtering algorithms often combine additional information with collaborative filtering algorithms.Matrix factorization stands out in collaborative filtering because of its extensibility and flexibility of containing information.Existing CF recommendation algorithms that fuse side information into MF framework have some shortcomings.First of all,the existing MF method of side information fusion is usually divided into two stages,with two stages of complexity and additional instability.Secondly,the popular CF recommendation calculation based on deep learning is also quite complex.Finally,the current CF recommendation often ignores the timing information,and the user’s preference is related to the timing information.Thesis presents a new matrix decomposition recommendation algorithm and a new matrix decomposition recommendation algorithm based on dynamic time.The main contributions of thesis are as follows:(1)Based on the characteristic attributes of users and items,thesis proposes a UISVD++model which directly integrates the characteristic attributes of users and items into the matrix factorization framework.The model projects the user’s age attribute and the movie’s type attribute directly into the same potential factor space as the user and item respectively,enriching the item and user representation in the matrix factorization.At the same time,the model also captures the implicit relationship between users and items from the user’s rating information,and completes the task of predicting the rating in a single computing stage.(2)On the basis of integrating user age and commodity type,thesis studies a matrix factorization model FTSVD++,which integrates feature attribute and dynamic time.This model introduces the time factor into the score baseline predictor and maps the side information into the hidden factor space to form a dynamic time matrix factorization model.Compared with the experimental verification,the model with dynamic time is more stable and accurate in the test set.(3)The recommendation model proposed in thesis was compared with two datasets,which verified that the recommendation model directly adding side information can effectively improve the performance of the recommendation system,and the dynamic time factor fusion model can effectively improve the recommendation effect.In addition,ablation experiment has also proved that adding side information of both user and item can more effectively alleviate the cold start and data sparse problems.