Research And Application Of Collaborative Filtering Based On Dynamic Multi-level Fusion Hybrid Factor

The rapid development of the Internet not only brings high efficiency to the society,but also produces a large amount of data information,but also causes the problem of data trouble.Therefore,a scheme to solve the problem of information data overload:recommendation system.The essence of recommendation system is to filter and screen massive data layer by layer,so that valuable information can be displayed to suitable user groups,that is,thousands of people and thousands of faces.If the Internet is the core of the interconnection of all things,the role of the recommendation system is to establish a medium for the interconnection of all things.Recommendation system can more easily associate users,content and services,so as to helping users avoid the tedious search process.Therefore,the recommendation system has been developed and applied faster in recent years.Collaborative filtering recommendation(CF)is the most popular and widely used recommendation algorithm in the recommendation system.The principle of the algorithm is to build and calculate the similarity between different similar user groups by analyzing user historical behavior data and user rating information,and according to the rating data of projects among different similar user groups,Output unknown rating forecasts for different projects.However,the accuracy of the recommendation results of the traditional collaborative filtering algorithm is not high,because it ignores the benefit of the influence factors in the core similarity calculation process on the similarity contribution value.At the same time,it also has some problems,such as cold start,low delay,scalability and so on.Taking neighborhood based collaborative filtering as the research goal,this paper aims to improve the problems of low recommendation accuracy,cold start and poor scalability in the collaborative filtering algorithm.Aiming at these problems,this paper puts forward improvement schemes.The research contents mainly include the following aspects:(1)Similarity calculation is the core model of neighborhood based collaborative filtering algorithm.The quality of a recommendation algorithm depends on the quality of similarity model.However,the result of collaborative filtering recommendation is not ideal when only using similarity algorithm.This paper proposes a new similarity algorithm,which uses popular weight reduction factor,score difference factor and interest forgetting factor to generate some positive and negative contribution values to the similarity.The experimental results show that the proposed method improves the accuracy of prediction results to a certain extent.(2)Traditional neighborhood based collaborative filtering recommendation algorithms ignore the impact of global context information on recommendation results and the measurement of similarity,because it only ignores the number of common ratings through the absolute rating of user groups.This paper proposes a recommendation algorithm that uses the number of common ratings divided into dynamic multi-level fusion hybrid factors to adjust the similarity of user groups.At each level,different positive and negative levels will be dynamically adjusted according to the parameters in the data set to restrict the contribution of positive and negative benefits to the similarity,so as to improve the recommendation accuracy.(3)This paper studies the industrial application recommendation system based on the current subject background,and designs and implements a set of lambda architecture under big data combined with Hadoop big data components The(real-time+ off-line)hybrid recommendation system effectively improves the scalability and recommendation delay of collaborative filtering algorithm.At the same time,the architecture process and core module design of the system are analyzed and discussed.