Research On Extraction Of The Backbone In Information Recommender Network

Nowadays, in face of a huge number of data, people find it’s difficult to effectively distinguish the information they need. Information booming lead to lower efficiency of online systems, which is called information overload. With a large amount of digital information, two important tools for information filtering search engine and recommender system have been extensively used. As an information filtering technology, personalized recommendations can uncover the potential interests of users based on historical information of their behaviors. However, when the recommender system encounters largescale data, the complexity of computing will burden, which makes system fail to tackle the full data. Meanwhile, applications often require real-time computing. So if we could extract the information backbone to compress the data size without affecting the recommendation accuracy, it can significantly benefit the real online scenarios. Therefore, we study the relevance between nodes or links in the network and the performance of recommendation based on the structure characteristics of the network. And we also try to extract the information backbone which preserved the performance of recommendation.The main works are as follows:1. In terms of identifying important nodes and links, we summarize the extraction methods of the information backbone in complex network and recommender network.Then we describe the basic theory, common algorithms, evaluation indexes and the structure of the recommender network. From the point of view of the importance of user node, we compare several characteristics of the user node to find the relevance with the performance of recommendation.2. We propose an infromation backbone extraction methhod based on similarity subgraph in the bipartite network. Several topology-aware characteristics of the recommender network are studied. We combine the similarity of users and objects to define the weights of links and find that the links with higher weight have more relevance with the recommendation accuracy. Experimental results on three real datasets show that the recommender systems achieve more than 90% of the accuracy of the top-L recommendation by relying on only 20% links extracted by our method. And the diversity of the recommendation can also be preserved. The number of neighbors has no significant impact on the results. Moreover, the structure of the information backbone is studied in detail and the topology-aware characteristics are better maintained. The information backbone of our extraction method can significantly compress the data so as to improve the performance of recommender system while effectively preserve the recommendation accuracy.3. We propose a mixed method based on betweenness centrality and temporal information to extract the information backbone in the bipartite network. In this thesis,by introducing the betweenness centrality which is the topological features of the complex network, the relevance between recommendation and the betweenness of user nodes,item nodes and links are studied. The results of experiments show that the higher the betweenness of items and links is, the more relevance they have with the recommendation.Accordingly, we propose a hybrid algorithm with temporal information and topology information. Experimental results on two real networks show that we can increase the precision of recommender system by adding the weight of the betweenness of links. On the contrary, we can increase the diversity of recommender system by adding the weight of the temporal information of links. By setting unique weight to every link according to their temporal and topology information, the algorithm improves the recommendation accuracy and diversity while remarkably shortens time consume and reduces the data storage usage.