Research On Academic Recommendation Strategy Based On Network Representation Learning

In the era of scholarly big data,the whole academic network is full of tremendous amount of academic information.Thus,how to excavate valuable information from the network becomes an urgent problem.To address this problem,academic recommendation systems arise in this environment.Paper recommendation and collaborator recommendation are two key research issues of academic recommendation.However,previous graph-based recommendations are mostly based on inflexible handengineered features,which leads to large computational complexity and poor recommendation effect in high-dimensional sparse environment.In addition,most graph-based recommendations lack the effective use of text information.In order to solve these problems,the main work of this paper is to integrate the text information of the paper into the network representation learning to achieve unsupervised feature design,and propose collaborator recommendation and paper recommendation methods respectively to solve the problem of information overload.(1)Topic-aware Network Embedding for Scientific Collaborator Recommendation(TNERec): TNERec jointly learns representations from scholars' research interests and network structure of collaboration network.TNERec first extracts scholars' research interests based on topic model and then obtains scholars' vectors based on network representation learning.Finally,recommendation list is generated according to the similarity of scholars' vectors.Experimental results on the real-world data set show the effectiveness of the proposed method compared with the baseline methods.(2)Vector Representation Learning of Papers with Text Information and Structural Identity for Recommendation(VOPRec): VOPRec firstly represents the text information with word vectors to find papers of similar research contents,and then transforms the hidden structural features of the citation network into vectors to find papers with similar network topology structure.After bridging text information and structural identity with the citation network,vector representation of papers can be learned with the network representation learning.Finally,the recommendation list is generated according to the similarity between the paper vectors.Through the APS data set,VOPRec outperforms the baselines measured by precision,recall,F1,and NDCG.