News Recommendation Research Based On Hypergraph Model

With the rapid development of economy and the Internet technology, news information has obtained rapid growth and propagation. More and more people prefer to read news online or by mobile phone rather than watching TV or buying newspaper. Massive news information generates and fast communication not only brings us a wealth of information, but also brings the problem of information overload. The research and application of news recommendation system can recommend suitable news and improve reading experience for us, and solve the information overload problem.Traditional recommendation researches are focus on the access collaboration between users and items, which have achieved good results in commodity recommendation, score prediction, etc. However, the recommended news always are newly published news; they are difficult to establish a direct and stable connection with users. Therefore, in this scenario, it is easy to appear the problem of item cold-start. Recent researches often construct the recommender based on content-based algorithm. They recommended news articles by the comparison of user’s trained preference profile and the candidate news. However, content-based recommender needs to express the news article as a Vector Space Model(VSM), which would be easy to lose the semantic information and could not fully consider some important basic characteristics of news content, such as genre, news class, area, reporter and other factors. Moreover, pure content-based recommender would lead to low diversity result. In addition, it is difficult to let the series news reports, that the context of user reading news, as recommended learning’s background factors. That is, not consider the influence of news evolution problem on user behavior prediction. Because item cold-start problem, diversity problem and news evolution problem have not been solved, the accuracy of news recommendation would be affected.Based on the above limitations, we proposed a hypergraph-based news recommendation model. This model is conducive to integrate the basic attributes of news content into recommended framework, and is also easy to integrate the content-based and collaborative filtering based method, which will gain better results. The main contributions and innovations are shown as follows:(1) Propose a hypergraph-based news recommendation model, which contains definitions of various objects, news content attributes and their interrelations. Hypergraph model is good at describing relationship among heterogeneous objects. This paper abstracts various objects of the scene of news and defines the hyperedge category over the edge with universality. In this way, this recommendation model would be scalable and could be extended to other applications and researches.(2) The clustering method in traditional news recommendation has limitations as it is only based on text single object. This paper studies on hypergraph clustering which is suitable for heterogeneous object relational mining, and proposes a Hypergraph Clustering Based news recommendation(HCB).Through the study of traditional news recommendation, we find that the method based on text clustering only mines the relationship among news articles in clustering processing, but not considers user relationships. Therefore, in this paper, we propose a recommendation based on hypergraph clustering, which contains the users, news and their relationships. Inspired by spectral clustering method based on ordinary graph, the recommendation algorithm in this paper can mine user interest in clustering. Then, choose most appropriate news for user based on news selecting policy. The recommender based on traditional text clustering can only analyze objects of the same class, and news articles need to converted to vector space model before clustering. While HCB method can cluster heterogeneous objects based on their relationships. That is, the users and the corresponding interest articles are locked in the clustering stage. Through experiments of real corpus, in recommendation accuracy, diversity and stability these three aspects, HCB method is superior to method based on content and collaborative filtering, also slightly better than text clustering methods.(3) In order to obtain news recommendation results more intuitively, integrate the clustering and news selection, we propose a Hypergraph Ranking Based news recommendation(HRB) and Hypergraph Ranking Based optimized by binary decision tree(HRBopt).Using hypergraph clustering could only find the interested news clusters. The candidate sets are often large and need to be further selected. Therefore, we propose to utilize hypergraph based ranking method for solve the news recommendation issue. In this recommendation framework, each hypergraph object is mapped to a corresponding matrix element. Then, the ranking values are computed by a ranking cost function. In order to further eliminate the subjectivity on the definition of hypergraph point and hyperedge, propose to utilizebinary decision tree for further amending the selection result. From the experiments, we can see that, HRB and HRBopt are not good at diversity and coverage compared with HCB and other clustering based baselines. However, the accuracy and NDCG outperform the other baselines.(4) In order to solve the user reading interest related issues in news recommendation, we study the news story chain and its application. We propose a Hypergraph news Story chain Based news recommendation(HSB).The user may be interested in the news on the missing news story chain. In this method, we define the basic element and principle. We construct the news story chain by hypergraph-based random walk, and gather the news articles in news chain as candidate set. From the experimental results, we know that HSB was good at accuracy and NDCG compared with content-based recommendation. However, it is worse than the other hybrid methods because it ignores the collaborative factor.General y, hypergraph is convenient to modeling, clustering and ranking among the many-to-many objects. In this work, we study the news recommendation and defined the news recommended objects using hypergraph model. Using hypergraph model can obtain better results and its modeling is more simple and easy to expand.