Study On The Based Temporal Dynamics Collaborative Filtering Of Recommender Systems

Internet booming has invaded every aspect of people’s life. It brings convenience to people’s life and generates large amount of data at the same time. Faced with the overwhelming information, the user is difficult to retrieve the information they are interested in to meet their demand, which results in the "information overload" phenomenon. To solve this problem, recommender system is established. By analyzing the information provided by the user and their behavior, establishing users’ interest models, and matching the users’ taste, recommending information which the users are interested in actively. It can help users to find information which is a potential interest to them and to enhance users’ loyalty to the system. It greatly promoted the interaction of information, particularly in the e-commerce site. Recommender algorithm is the core of Recommender System. Collaborative filtering technology is the most extensive and mature technology. Collaborative Filtering technology does not require the users’ profile and domain freedom. It based on the users’ historical behavior and historical data to infer the user’s behavior patterns and interest hidden preferences. Then do some relatively recommending. Due to efficiency and scalability, it has been successfully used in many large commercial sites. This paper focuses on two core algorithms of collaborative filtering- neighborhood model and latent factor model. After the deep research of collaborative filtering and the summary of the previous work, an integrated model of multiple models was proposed. It has been proved on Netflix dataset that this model make a larger contribution to precision of Recommender System.This paper analyzes the neighborhood model and latent factor model which are the two main models of collaborative filtering. Neighborhood model focuses on item-based model and user-based model and latent factor model focuses on singular value decomposition(SVD) technique. This paper studies with the issues such as interpretability, scalability and new users of the Collaborative filtering technology and analysis the merits of neighborhood model and latent factor model respectively. For the lack of interpretative issues of SVD model, asymmetrical SVD model was introduced in detail. This paper focuses on the scalability issues of the global optimizationneighborhood model. In order to significantly reduce the complexity of time and space,this paper modified Pearson correlation similarity measure rules by considering the user and item bias. By neighborhood prune unlikely relationship of items to form pruning global neighborhood models. But pruning will reduce the accuracy of the model. So this paper proposed a factorization neighborhood model which is similar to an asymmetric SVD model but has a different meaning. At the same time, it can integrate the model of factorization user-user relationship and cause no loss of the accuracy under the circumstances of reducing the time and space complexity. This model has improved the problems, such as interpretability, scalability and new users problems. This paper also analyzed other three factors that affect the accuracy of recommended techniques, such as implicit feedback, temporal dynamic factor and the confidence level. It analyzed and summarized its forms in various models respectively. Among which the implicit feedback and temporal dynamic factor on the recommendation have much greater improvement of precision of Recommender System than other factors. Combined with the previous summary, research analysis and improvement, this paper has proposed an integrated model, which including baseline estimate, pruning global neighborhood model, SVD ++, factorization user-user relationship global neighborhood model and temporal dynamic effects factor.Finally, this paper had experiments on Netflix dataset and these experiments measured the recommended effectiveness according to root mean square error(RMSE).This paper has a comparison of the accuracy among each neighborhood model and each SVD model and has a comparison of the runing times among each factorization neighborhood model. It also has a comparison of the accuracy between the model append temporal dynamic factors and the model without temporal dynamic factors.Experiments on Netflix dataset verified the following conclusions that the factorization item-item relations or user-user relations can significantly reduce the running time of system and the temporal dynamics factors can significantly improve the accuracy of system. Finally, compared with the accuracy of the new model we proposed and the TimeSVD++ model, this paper verified that the new model can greatly improve the accuracy of system.