Design And Research On Recommendation Algorithm Based On Similar Game And Personalization

In the early days of the Internet era, classification index can effectively help users to find the information needed. With the increasing of the Internet data, a bottleneck appears that classification index is difficult to cover all the data, which can be solved by search engine. The outbreak of Internet data leads to data overload in all fields. Limited by keywords, search engine is difficult to deal with growing personalized requirement. When searching keywords can not be quantified, such as “the best”, ads tend to fill search results. And then, recommendation system has been widely studied and used.Recommendation algorithm is the soul of the recommendation system. General recommendation algorithms can be divided into online and offline from the perspective of real-time. Offline recommendation is the data base of online recommendation, and online recommendation is the further application of offline recommendation. From the perspective of the characteristics of users and items, offline recommendation can improve the accuracy of the recommendation. Offline data are relatively stable, which is easy to model and evaluate. Offline recommendation generally requires a longer running time. If the idle period is ignored, “interest drift” problem will appear. Then online recommendation is needed. Online recommendation can adjust the recommended results that are from offline recommendation, which can effectively alleviate the “interest drift” problem and maintain the accuracy of offline recommendation. Two aspects are studied in this paper: offline recommendation and online recommendation, specifically including the following contents:Firstly, this paper improves the traditional item-based collaborative filtering algorithm by using distributed computing framework. A data relation extraction scheme based on Dpark and a matrix storage scheme based on BeansDB are proposed by analyzing the performance of Dpark, a distributed computing framework, and the characteristics of BeansDB, a distributed memory database.Secondly, on offline calculation, this paper puts forward an offline recommendation algorithm introduced user curiousness degree and item popularity degree which emphasizes “personalization”(By using public music data set in the experiment, this paper mainly discusses the “music popularity”, and will directly uses the term “music popularity degree”). This paper discusses the definition and theoretical basis of user curiousness degree and item popularity degree respectively, and proposes improved recommendation model based on the two measures. The principle of the model is explained in detail. And after the implementation of the algorithm, compared with traditional item-based collaborative filtering algorithm, the performance has been analyzed.Thirdly, on online calculation, a detailed description has been made on the phenomenon of “interest drift” and conventional alleviation methods are discussed, and the online recommendation algorithm introduced “similar game theory” is proposed. Combining the results of offline recommendation algorithm, the algorithm does online recommendation during the interaction of system and users, which effectively alleviates the “interest drift” problem in offline recommendation and enhances the recommendation algorithm in real-time and flexibility, and puts forward a feasible method to realize the algorithm flow of the similar game theory. Experiment results have been demonstrated and analyzed.Fourthly, this paper proposes a kind of overall architecture of personalized music recommendation system based on the “RESTful Web Service”, realizes the combination of theory and practical system, and shows the basic server-side interfaces and the screenshots of website and mobile app, which provides a feasible technical route to apply the algorithm to practical systems.