Research On News Recommendation Algorithm Based On Topic Matrix Decomposition Model

Today is the information explosion era, every day we are in the browser with a great variety of news, a lot of times, we have had the "Select difficult disease", which is the "information overload" problem, therefore, recommended smart news emerged. By providing users with a personalized reading list, improve the user's reading experience. The current network news reading has become one of the most popular application of the Internet.Cold start problem is a personalized recommendation has been present and has not been well addressed, including the User cold start and Project cold start, however, Content-based recommendation can solve the cold start program. On this paper, firstly, we study the issue of the user cold start problem, it is recommended to generate the initial set of candidate recommended by popularity, then the use of Topic model and Content-based recommendation and user context information which combine while taking advantage of negative feedback based supplement strategy, after iteration construct user interest model. This paper proposes the idea of "appropriate pause and interesting", to build multi-level training iteration users interested in new user news recommendation model provides a good solution for users cold start problems.The core of news recommendation is user identification and user personalized modeling. Read the news on the Internet, many news portals and news applications do not require users to login or register, and users in order to facilitate, more users are worried that their information was leaked out and do not want to register to read the news only after, This makes the news recommendation cannot get users to provide their own personal information as well as some of the explicit interest, At the same time, the user in the process of browsing the news, but also inadvertently left a lot of footprints". For example, the location of the IP, the time to enter the system, browse what content and other implicit data. Under such a premise, this paper makes a research on news recommendation. This paper is accurate and complete characterization of user behavior, implicit access to the user's news browsing log into long-term, near-term, short-term, real-time four time dimensions to build user interest model. Because the news is dynamic, and the user's interest is dynamic, user interest will change with the time and situation. This paper introduces the MAC address, topic model, context information, we put forward a real-time user interest model based on subject. The model is divided into three layers, which are offline layer, middle layer and online layer. The model is effective in real-time processing.The news attribute as the inherent attribute of each article. In a very long time ago, the news recommendation algorithm has been used as a basis for the recommendation, and also achieved very good results. However, most of the traditional recommendation algorithms based on the news attribute assume that these properties are not related to each other. In fact, there are more or less the influence of the property. Indeed, there are a few researchers taking into account the practice of this attribute, Achieved a general effect. This paper is based on the previous studies and semantic analysis of topic model and the matrix decomposition model of news attribute, Based on the identification of user and user interest model, some of the relevant topics such as user topics and news topics as well as news attributes are obtained. Proposed a Topic matrix decomposition news recommendation algorithm of it is fuse the user themes, news themes, user behavior attributes, news attributes and situational information, has been more accurate way to the user for news recommendation. The experimental results show that the proposed recommendation algorithm has better performance than the traditional recommendation algorithm, and the diversity index is superior to the traditional algorithm.