Heterogeneous Academic Network Based Scholarly Impact Assess And Prediction

In recent years,scientific and technological competition has become increasingly fierce,and the global scientific research level has been continuously improved.The number of academic entities such as papers,authors,and publications generated by scientific research is increasing exponentially.The most commonly used method in current academic influence research is to explore and analyze heterogeneous academic network composed of academic entities and their relationships,discover various implicit characteristics,and use them to evaluate and predict the influence of entities.At present,despite the emergence of many research results on academic influence assessment and prediction,which have been used in academic and educational fields such as research funding,talent selection,discipline construction,academic recommendation,etc.,their research still faces many challenges.For example,the capture of dynamic characteristics in heterogeneous academic networks,citation changes in classic papers,citation differences in papers from different research areas,human citation manipulation,citation inflation,ranking bias between new and old papers,and the mutual reinforcement between entities.In response to the above issues,this dissertation starts with the research object,and based on the logical relationship between the research object from a single academic entity to multiple academic entities,from evaluating current academic influence to predicting future influence,conducts research on four aspects: evaluating the influence of papers,predicting the influence of papers,evaluating the influence of multiple academic entities,and predicting the influence of multiple academic entities.A series of academic influence evaluation and prediction methods based on heterogeneous academic networks are proposed.The main contributions are as follows:(1)Evaluation of the current influence of papers.The existing paper evaluation methods do not consider the dynamic characteristics of venue influence,as well as the time factor and topic content at the same time,leading to the inaccurate evaluation results.To solve this problem,a method for evaluating the influence is proposed.A method for evaluating the influence of papers is proposed to address the issue of inaccurate evaluation results due to the lack of consideration for the dynamic characteristics of venue influence,as well as the lack of consideration for both time and topic content in existing evaluation methods.1)In order to accurately describe the dynamic characteristics of the venue’s influence,each venue in every year is treated as a separate entity,and both of its current performance and past performance for a number of years is considered when calculating the influence of the venue.2)Combining the venue influence with the publication time,the citation weight and random walk probability in the citation network are dynamically assigned to improve the performance of paper influence evaluation and effectively alleviate the ranking bias of papers.3)The personalized jump probability is dynamically assigned according to the influence of the paper in the same topic to further improve the paper’s evaluation performance.4)The influence of papers is evaluated based on the mutual reinforcement between papers and venues.Finally,the influence of papers in the ACL dataset is evaluated and analyzed,the experimental results indicate that the evaluation method proposed has higher accuracy.(2)Short-term prediction of paper citations.The existing prediction methods are difficult to accurately predict the future citation counts due to the lack of citation features of papers published shortly.To deal with this issue,novel prediction methods of paper citation counts are proposed in this dissertation.1)In the citation network,we analyze the difference of overall citation trend between papers in different research areas,and classify the research areas of the papers by using the content relevance,the number of citing times and the cited times between papers and different research areas.In order to improve the prediction accuracy of citation counts of papers,the overall citation trend of papers in the same field is used for fitting and prediction.2)For the same area,we analyze the changes in citation trends between different papers,and classify/layer the papers according to their existing citation counts.In order to further improve the prediction performance of papers,the citation trend of papers at different levels is used for fitting and predicting.3)To integrate the advantage of all the individual models through the stacking technology,a novel combined model based on a series of academic features is proposed to further improve the prediction performance of papers.Finally,examined by the short-term prediction and analysis of the paper citation counts in the DBLP data set,the experimental results show that the prediction method proposed in this dissertation has better prediction performance.(3)Evaluation of the influence of papers,authors and venues.Existing methods for assessing the influence of papers,authors,and venues are prone to human manipulation because they do not distinguish the citation strength and do not consider the topic similarity at the same time,making it difficult to accurately assess the influence of papers,authors,and venues.Aiming at this problem,a multi-entity impact assessment method is proposed.1)Use stacking technology to classify citation strength,and combine citation strength,which measures citation importance,and topic similarity,which measures topic similarity,to build a weighted academic network to improve the performance of impact assessment.2)Consider recent citation bonus of papers and authors to alleviate the ranking bias of new papers(for example,papers published for within 4 years)and old papers(for example,papers published for more than 10 years)and further improve impact assessment performance.3)Based on the weighted academic network,we evaluate the influence of papers,authors and venues to improve the rationality of academic multi-entity evaluation and the robustness against human manipulation.Finally,the ACL dataset is used to evaluate the influence of papers,authors and venues.The experimental results show that the proposed method has higher accuracy in evaluating and better robustness on resisting human manipulation.(4)Prediction of the future influence of papers,authors and venues.Existing methods for predicting the influence of papers,authors and journals do not consider the dynamic characteristics of author influence,citation changes in classic papers,and citation inflation of in co-author network),resulting in the inaccurate prediction results.Thus,a multi-entity influence prediction method is proposed.1)In order to accurately describe the dynamic characteristics of the author’s influence,each author in every year is treated as a separate entity,and both of their current performance and past performance is taken into account.2)In order to accurately predict the future influence of classic old papers and new papers,we consider both publication age and recent citations of all the papers involved at the same time,as are the authors and venues.3)Assign weights to author citation relationships and author cooperation relationship according to the author’s order in the paper to avoid citation inflation.4)By considering the author year and the publication year,a weighted academic network is constructed to predict the future influence of the paper,the author and the venue,so as to improve the prediction performance of papers,authors and venues.Finally,based on ACL dataset,the experimental results show that the prediction method proposed has better prediction performance.