Research On Personalized Recommendation Based On Graph

With the rapid development and popularization of the Internet, the Internet has provided people with vast amounts of information, and information presents explosive increasing. Vast amounts of information provide people with more choices, but bring many problems at the same time. It's more and more difficult that people find useful information, the information we need would be discovered until people browse a large number of pages. Dynamic personalized recommendation system bring a viable solution for this issue, and internet usher in "Recommendation Engine" era from "Search Engine". Opening the page, the information would be present for you before when you want it. On the other hand, the construction of site is more productive, personalized recommendations show a variety of information for different users. In addition, businesses complete market positioning through personalized recommendation system and realize precise advertising and personalized marketing, thus obtaining the maximum benefit, but also avoiding the user's resentments for random recommendation.Personalized recommendation has become a popular research field and raise interdisciplinary concerns. Recommendations based graph express user-Item relationship as bipartite structure and raise more and more attention. The recommended method based graph achieves filtering information through the diffusion of energy. For two prominent problems of personalized recommendations — the cold start and the long-tailed mining. This method provides feasible solution, some areas have adopted graph-based personalized recommendation, such as advertising and dating site. In addition, some parallel architectures intended to address "Graph Computing" have been launched, which provide help for the further applications of recommendation algorithm based on graph in the actual scene. In this paper, we will propose the two-way diffusion based on graph, a complete exposition form the motivation to achievement have been present, and then algorithm performance have been proved. In general, the user's interests are affected by two aspects, one is the user's own preferences, which is reflected in the user's historical trading data,user may prefer similar items. The other is that user's interests would be affected by friends around, which is reflected in the common interests of circle of friends, a potential item may be popular in the circle of friends firstly. Thus the proposed algorithm uses the famous "PageRank" method, we add feedback procedure to strengthen the recommended accuracy based on the "heat transfer algorithm". On two famous experimental data sets, the new recommended algorithm have shown excellent performance in several indicators and long-tailed mining, experimental results have proved that our algorithm has a greater improvement than previous algorithms. For the cold-start problem, with the aid of graph our algorithm has more advantages. In addition, the proposed method can also be extended to other algorithmic model. For example, based on the proposed algorithm, if the second-order relationship and the asymmetric similarity can be take into consideration, the algorithmic performance can be further improved.