Identification And Recommendation Of Serendipitous Scientific Collaborators

Academic collaboration has become a universal phenomenon in science.The emergence of academic recommender systems has effectively helped scholars to find valuable collaborators.However,most existing recommender systems focus on improving the recommendation accuracy,and ignore the serendipity of recommendation results.Serendipitous collaborations may contribute to the innovation of scientific technology or the breakthrough of scientific achievements.Therefore,we focuses on studying the characteristics and forms of serendipitous scientific collaborators.Moreover,we propose two methods involving the identification and recommendation of serendipitous collaborators,respectively.Firstly,we summarize the concepts of serendipity in related works,and then define serendipitous collaborators from three components of relevance,unexpectedness and value.The corresponding intuitive concepts of them are similarity on network structure,topic diversity,and influence of the collaborator,respectively.Secondly,we design two valuable methods based on the definition of serendipitous collaborators.They are the identification model based on the clustering algorithm,called RUVMod,and the serendipity-biased recommendation method based on the network embedding algorithm,called Seren2 vec,respectively.RUVMod clusters all the collaborators of target scholars,and divides all collaborators into 8 categories by analyzing three indices in definition of each collaborator cluster.The cluster with lower relevance,higher unexpectedness and higher value is classified as serendipitous collaborators class.In addition,Seren2 vec integrates serendipity into co-author network,and improves the traditional network embedding algorithm.The process of representing scholars into vectors is serendipity-biased.We compute the cosine similarity between vectors in order to recommend serendipitous collaborators for target scholars.The two methods are experimentally analyzed in the co-author network constructed from DBLP data set.Experimental results show that RUVMod outperforms other methods,which effectively identifies the serendipitous collaborators.Seren2 vec improves the serendipity of recommendation list and maintains adequate accuracy simultaneously.