Research On Researcher Recommendation In Scientific Social Website

In Web2.0 era, users in social network platform have the authority to publish information, exchange ideas freely, which attract increasing number of people to sign up in this kind of websites for participating scientific communities and communicating about the professional knowledge. Due to the lack of professional social networking professional atmosphere, some sicentific social networking sites had been emerged,which were researcher-oriented, such as ResearchGate, Academia.edu in foreign countries and Baidu Academic, Scientific Research in China, since around 2007. Up to December 2016, over 12 million certified scholars have registered the ResearchGate,they browse other scholars' homepages, look for interested literature and experts, as well as participate in those academical discussions, questions and answers. These kind of websites are available for all researchers being from various fields, they can discuss topics with others easily, seeking potential opportunities for cooperation. Previous researches show that such platform promotes the scientific research effectively. It is obviously that finding similar research scholars and potential collaborators are two of the important reasons why researchers use scientific social websites.However, the scientific social website encounters the similar troubles of information overload and information asymmetry with the common social wensites.Fortunately, it is an effective methods to construct a personalized recommendation model based on the academic knowledge and the scientific cooperation network of researchers.Furthermore, a new trend in the current information processing and retrieval system is the acquisition of contextualized content. Considering the contextualized information into information processing will be beneficial for improving the recommendation accuracy and mitigating the information overload, and then be better adapt for target users' special needs which are independent of their history records. Therefore, this paper analyzes the social motivation of researchers in scientific social websites, and draws the difference of two kinds of recommendation scenes. It is believed that researchers are interested into the scholoars who share the common research orientaion and research preferences, and be willing to establish long-term social relations with them. Moreover,many researchers are characterized with context, hoping to find collaborators with specific requirements. For example, they have relevant experience of projects or papers on this context topic. Based on the above analysis, this paper proposes two resercher recommendation models, namely, researcher recommendation model based on the similar research interest, and collaborators recommendation model based on the specific context.Aiming at the two recommendation situations, this paper designs two corresponding and reasonable solution strategies respectively.In the researcher recommendation model based on the similar research interest, two sub-models are constructed: expert profile model and academic behavior network model.In the expert profile model, this paper employs the language model to represent the expert knowledge according to the experts' academic profession,academic orientation,academic achievements and so on, and then uses gengerated probability to calculate the similarity of the scholar's knowledge based on the Bayesian decomposition. In the academic behavior network model, firstly, the Adamic-Adar method and the Shortest Path method are used to measure the similarity and path distance of the scholar nodes in the cooperative network after minning the collaboration relationship in academic behavior network, secondly, Jaccard coefficients are used to represent the relationship degree of researchers' organization in collaboration nework from the perspective of global academic and local academic respectively. Finally, the Comb strategy is applicated to integrate the above measurements. After that, those experts of top-K highest similarity are recommended researchers. In the collaborators recommendation model based on the specific context, this paper designs two criterias to evaluate the quality of potential collaborators: expert academic quality evaluation and academic social network quality evaluation. In the expert academic quality evaluation, the introduction of situational filtering and context after the filter to the recommended method, the quality of researcher's academic achievements(number of researches, journal level,citation), titles of researchers, G index are used to evaluate their academic ability, at the same time,introducing the contextual pro-filtering and post-filtering methods. Firstly, it selects those reserchers whose reserches contain the content of the context by contextual pre-filtering strategy and generate the initial candidate collaborators set, and then use the adjusted Dirichlet Distribution method on the context. The Kullback-Leibler difference is used to calculate the knowledge match between the target researcher and the candidate researchers in the initial set while the standardized expert academic ability score is used as the weight value in the matching calculation. In the academic social network quality evaluation, this paper constructs a multi-network, including four types of relationships:paper cooperation, project cooperation, patent cooperation and attending the common meetings. At the first, it calculates of the counts of the four types, and then introduces the time factor of the year when relationship is built to amend the score of the cooperation quality, Finally, the two scores are integrated to express the score of cooperation intentions.In order to clarify and complete the models this paper proposed, a global system framework of the scientific social network is also constructed. Builing a simple crawler with Python2.7 + Selenium + Scrapy to collect data, and obtains 2,000 researchers as the model simulation data collection, and collects more than 80,000. The application process of the two models are given in detail in the chapter 5. The experiment results show that the model has the feasibility of applying and good recommendation effect.