Research On The Prediction Method Of Paper Influence

Citation count is one of the most intuitive indicators to measure the influence of a paper,a highly cited paper indicates that it has a high academic influence.In order to keep up with the trend of scientific research,researchers need to constantly read the current influential papers and to read papers that will have influence in the future.Identifying potential papers in advance is helpful for researchers to select appropriate references and research fields,and to provide reference for the system of paper retrieval and recommendation.Therefore,how to find potential and influential papers from a large library of papers is a hot topic in the current research.The existing research is mainly to predict the influence of the paper after extracting the features related to papers,authors and venues.However,there are the following deficiencies in the current research:(1)Features related to academic networks have not yet been fully considered.(2)Not every feature affects the influence of the paper equally in a large number of features.Which kind of features can significantly affect the influence of the paper has not been studied in detail.(3)It does not consider the feature selection and it has not verified whether feature selection can improve prediction efficiency.On the basis of the existing researches and to solve the above problems,thisthesis proposes a new model for predicting the influence of the paper.The model mainly includes the following aspects:(1)The features related to the academic social network are extracted,including the features related to authors,papers,venues,paper network,author network and venue network.(2)The importance of each feature is analyzed using a neural network model and several sets of optimal feature group are selected through correlation-based feature selection method and manual selection method based on feature analysis results.(3)A better prediction method is identified after using different forecasting methods(Multi-Layer Perceptron,Gauss Process Regression,Multiple Linear Regression and Support Vector Machines)to predict the influence of the papers in different time periods(3 years and 5 years).This thesis used the dataset on ArnetMiner to performe several sets of comparative experiments.The experimental results show that support vector machine is most suitable for predicting citations of papers.Compared with existing methods,the model improves the prediction accuracy in this paper.