New Paper Influence Prediction Based On Academic Network

In recent years,the number of published papers has shown an exponential growth trend.It can effectively help researchers select papers from the mass of papers to mine potential influence of new papers or discover high-quality new papers.In this way,researchers can save their time and pay more attestations to the high-quality papers.However,highly cited papers account for a small proportion.Traditional methods evaluating a paper's influence by the number of citations don't work for a new paper without any citation.Therefore,it is a challenging problem to mine the potential influence of new papers and discover new high quality papers.In order to solve the problem of predicting the influence of new papers,we construct a heterogeneous network,consisting of researchers,papers and relationships.The relationships in academic networks includes the coauthor relationship between researchers,the citation between papers and affiliations between researchers and papers.Based on the data of academic network,this project is to predict the influence of new papers.It mainly addresses two challenges:the dynamic evolution of topics,topic hotness and topic-related authority caused by the continuous evolution of the corpora;joint correlations learning for new papers'influence prediction.For the challenge of the dynamic evolution of topics,topic hotness and topic-related authority,we first introduce Dynamic Topic Model(DTM),which extends the traditional state space model and learns the evolution of topic-word distributions over time.Then,we measure the topic hotness in different periods based on the probability distribution of paper-topic.Finally,the PageRank algorithm is implemented in the citation network to obtain the potential influence of papers.We measure how an author's pre-existing authority affect his new paper's influence according to the cross product between the topic-related authority of authors and topic distributions of the new paper.For the challenge of learning joint correlations of multi-factors,this paper also analyzes an author's social factors from interpersonal,inner-clique and inter-clique aspects.What's more,this paper also analyzes the venue-centric feature of conference or journal that new papers are published in.Then,Factorization Machines(FM)model is employed to learn the potential influence of a new paper by using latent vector to represent each dimension of the paper's features.We use the stochastic gradient descent algorithm to learn parameters in this model and predict the influence of the new paper with the help of joint correlations.The datasets used in this paper are crawled from the ACM Digital Library for verification.Compared with other methods,our model shows that the method is effective.In addition,this paper designs and implements an Influence Prediction System(IPS)based on mobile devices.The IPS shows the analysis of a paper's topic distribution and influence prediction.Besides,this system also ranks topic-related authority in detail.We also perform functional tests and performance tests of this system.The test results show that the system can operate normally.