Research on the Prediction of Highly Cited Papers Based on PCA-BPNN

With the increase in scientific research investment, the number of papers has increased significantly, and the evaluation of the impact of papers has received extensive attention from scholars. The citation frequency is the most convenient and widely used index to measure the academic influence of papers. Still, the citation frequency can only measure the real impact of papers some period of time after those have been published. Therefore, to be able to identify highly cited papers at the early stage of publication, this paper collects data on 1025 academic papers published under the library and information discipline of the Web of Science library in 2007 and then extracts 24 predictive characteristics from three aspects: papers, authors, and journals. On this basis, 7 principal component vectors are constructed by feature screening based on PCA. Also, combined with the BP neural network model, the PCA-BPNN highly-cited paper classification prediction model is constructed and finally compared with the other 5 models. The results show that the PCA-BPNN model built in this paper has better prediction performance and provides an effective model for the prediction of paper influence.