Hu Zewen, Cui Jingjing, Xu Rong, Gu Yilin
[Purpose/significance] This study aims to develop an automated prediction framework for identifying poten⁃tially high-impact papers by integrating multidimensional feature index with deep learning methods, enhancing the comprehensive characterization and accurate prediction of academic value and impact, thereby providing methodologi⁃cal support and technical pathways for optimizing research evaluation, allocating academic resources, and informing science and technology decision-making. [Method/process] In order to realize the accurate prediction of potentially high-impact papers in the literature of social science field, this study firstly constructs a three-dimensional feature in⁃dex system of scientific and technological papers from the three dimensions of the papers' own features, content fea⁃tures and cited features. Then, this study designs and realizes the automatic prediction framework of potentially highimpact papers, which integrates the three-dimensional feature index of papers and deep learning model, to realize the deep learning prediction of potentially high-impact papers in the massive literature from the major social sciences dis⁃cipline named economics. Finally, the differences and advantages in features of potentially high-impact papers are sys⁃tematically compared and analyzed. [Result/conclusion] There are significant differences among various dimensional index in scientific and technical literature. The constructed three-dimensional feature index system and feature vector space, combined with deep learning prediction models, can comprehensively measure the value and influence of scien⁃tific and technological papers. At the same time, the superior prediction effect can promote the automatic prediction and recommendation application of potentially high-impact papers in massive literature. Artificial neural networks and TabNet perform well in prediction accuracy and precision, but are inferior to traditional machine learning models in metrics such as recall, P-R area, and AUC value. Using citation features of papers outperforms using the paper's own features or content features to predict high impact papers. Potentially high-impact papers exhibit significant advantag⁃es in multidimensional feature index, including papers' own features, thematic features, citation dynamics, etc.
