Research On Personalized Recommendation Technology Of Information Stream Based On Matrix Factorization

With the development of web 2.0, information stream service has gradually replaced the traditional media as the main information source for people. This kind of emerging platform deliver relevant information flow to users according to their personal subscription, while also allowing all kinds of interaction between users, and thus facilitate the generation and transmission of information. However, common characteristics of information stream system also lead to the general phenomenon of information explosion which in term caused the reading problem of user. Therefore, how to construct an effective model for personalized recommendation of information stream system and help users to filter and discover information is particularly an import problem at this time.With the study of characteristics of information stream systems, this paper proposed a recommendation model based on matrix factorization. The main idea is to extract user featured vector according to user’s own historical data or other data he may concerned by two extraction method which were also proposed by this paper. And we used the featured vector as the input data to replace the user-item scored matrix which was widely used in traditional recommendation models. Then, by using a method similar to MF model, we get the latent factor vector of all the users and featured words. Finally, we descript the specific methods to generate recommendation lists. And we proposed a solution of our model which used the stochastic gradient descent method.In addition, this paper presents an enhanced model which involved the idea of trust propagation. Meanwhile, we redefined the calculate method of precision and recall according to the characteristics of information stream system.In the end, we take experiments in real data sets of Weibo and presents the analysis of the result. We analyzes the effects of two thresholds and two kinds of user featured word vector extraction methods. And we compare our model with the random recommendation model. The result shows that model which extract featured words from user’s own information have better performance. Meanwhile, our model was better than random recommendation model regardless of which extraction method we used.