The Research And Implementation Of Paralled Personalization Recommendation Algorithm For Large-scale User Behavior Data

With the continuous development and prevalence of Web2.0, the amount of information on the Internet rose to another order of magnitude. In the flood of information, user lost its way easily, and it is difficult to know what user’s recessive demand for service providers. So, a service which can predict user’s interests and take the initiative to push the information which user is interested in has great significance for service providers. Personalized recommendation services become a hot spot in response to this demand. More and more E-commerce sites begin to start deploying recommendation service to attract users and recommend goods to users in order to enhance the effectiveness of the site. The personalized recommendation has also become a popular area in Academic research. The user behaviors studied in this paper are of Web users. Unlike behaviors that can be explicitly scored, the behaviors of Web user are implicit and cannot be easily scored which make user’s interest to items be implicit as well.Collaborative filtering is a popular technology for personalized recommendation service, which focuses on the synergies between users, but there still exist some challenging issues, e.g., cold start of new items, scalability. So, how to combine the other recommended techniques such as content based recommendation, or theories of other disciplines such as social network analysis with the collaborative filtering for solving these issues has become a hot topic.For a user behavior dataset of the information field, we propose a new hybrid recommendation algorithm which combines the user’s interest model with Slope One collaborative filtering. To resolve the scalability issue, we also implement a parallel version of our recommendation algorithm. This paper alse proposes a content-based-of SimHash similarity recommendation method to solve the cold start of new items.In dealing with recessive behavior data of users, this paper proposes a rating strategy based on time of user on the page and page content size. By taking advantage of the naive Bayesian classifier to classify items, this paper creates user’s recent preference model of item categories based on user’s ratings data and time weights. In the recommendation process, this paper combines user’s preferences deviation of item categories to Slope One collaborative filtering algorithm to optimize recommendation quality. Finally, the hybrid recommendation algorithm is tested in application data. Experimental results verify the introduction of user interest model to the Slope One algorithm can effectively improve quality of recommendation and solve the cold start of new items. Also illustrate the need to consider the influence of user’s personal interest in the field of information recommendation.