Research On Information Recommendation Technologies In Online Social Networks

The rapid development of Internet brings people into the information age, and the exponential growth of information amount indicates the coming of big data era. It is essential to analyze and mine the potentially valuable knowledge and regularity from the big data, which will efficiently uncover the latent information of interest to users. Nowadays, online social networks, such as e-business and SNS, have become an important part of our daily life. However, online social networks bring us the information overload problem as well as living convenience. Taking advantage of the methods in the fields of data mining and machine learning, recommender systems arise to help users find the potential objects of interest from information ocean. Therefore, recommender systems have become a significant issue in both academic and industrial communities, and received considerable attention of domestic and overseas scholars.Currently, network science has become the most active interdisciplinary science.In the discipline of network science, the mining of influential nodes is an important issue with many applications. For example, the failure of influential nodes may disable the whole network, or even another coupled network(cascade failure). The mining of influential nodes is usually conducted on unipartite networks, then how to apply this technology into recommender systems on bipartite networks, become an interesting research issue. This thesis studies the personalized recommender algorithms on bipartite networks, and then improves the algorithmic performance by introducing the information of user social networks. The major contributions are listed as follows.(1) We proposed an improved random walk algorithm by depressing the influence of large-degree objects. Experimental results on MovieLens and Netflix data sets show that this algorithm can effectively improve not only the accuracy but also the diversity with respect to information recommendation. On the other hand, we investigate the effect of weight assignment when combining both random walk(RW) and heat conduction(HC)algorithms together, and find that the new hybrid algorithm of RW and HC with balanced weights will achieve optimal recommendation results. That is to say, in order to obtain a better result of recommendation, we should assign same weights to them so that same influence can be given by object’s degree in different diffusion steps.(2) In recent years, many outstanding recommendation algorithms have been proposed to alleviate the information-overload problem. Generally speaking, adjusting parameter properly based on the information contained in testing set will lead to an optimal performance. However, in real online recommender systems, the testing information is not available to us, so how to choose an optimal parameter in recommender system becomes a big challenge. To solve this problem, in this thesis, we propose a method to estimate the optimal parameter approximately by merely using the training set’s information. Furthermore, the experimental results illustrate that we can estimate the whole system’s parameter by using only 10 percent of information.(3) Based on the coupled social networks(CSN), we propose a hybrid algorithm to nonlinearly integrate both social and behavior similarity of online users. Based on the coupled social networks, filtering algorithm considers the effects of both social similarity and personalized preference. Experimental results show that the hybrid pattern can not only provide more accurate recommendations, but also enlarge the recommendation coverage while adopting global metric. Further empirical analyses demonstrate that the mutual reinforcement and rich-club phenomenon can also be found in coupled social networks where the identical individuals occupy the core position of the online system.(4) Many researchers focus on mining the influential nodes on single-mode networks, however, exploiting the information network to help mining the influential nodes in the CSN is lack of study. In this thesis, we propose a method via mining the important nodes on CSN. In addition, we take the ranking information in recommender systems into account and obtain better accuracy and diversity for information recommendation.In real online E-commerce system, generally speaking, most of the large-degree objects are popular objects, which have large weights. If their influence is taken into account, the small-degree objects might be difficult to be recommended. Therefore, depressing the large-degree objects influence would take the full advantage of the big volume of small-degree objects. In a word, this work provides a practical solution for online recommendation, that is to say, how to promote the attention on the long-tailed products.In real online recommender system, we cannot adjust the parameter of recommendation’s algorithm to obtain an optimal performance because the testing information is unknown to us. Therefore, it is significant to estimate the parameter while the testing information is unavailable. To solve this problem, we propose a method to estimate the optimal parameter approximately by merely using the training set’s information.We provide deep insights into researching the key technologies and theory of information mining and recommendation in online social networks, and propose some effective methods in recommending the information and mining the important nodes in online social networks. Furthermore, we take the ranking information in recommender systems into account and obtain more accuracy and diversity for information recommendation. Our studies are significant to explore the solutions for information-overload and have wide potential applications. This thesis may shed some lights on the in-depth understanding of the structure and function of online social networks.