摘要：

Abstract:

【Purpose/significance】This study aims to identify potentially high-impact papers by predicting citation counts of academic

articles. Despite early research considering factors such as literature, authors, and journals, it often overlooks the textual semantic in⁃

formation in paper metadata and suffers from data longevity issues【. Method/process】This study collected research data from ten Chi⁃

nese journals in the field of Information Resource Management published from 2008 to 2017. We constructed 16 formal features of the

papers and utilized pre-trained language models to build semantic features of paper metadata to enhance the predictive performance of

the model. Additionally, this study introduced a deep forest model for predicting citation counts.【Result/conclusion】The results show

that the deep forest model generally outperforms other experimental algorithms in predicting citation counts. In baseline experiments,

the deep forest F1 mean is 2.65% higher than the mean of other methods. After incorporating semantic features of papers, the effective⁃

ness of the citation prediction model improved significantly, with F1 results surpassing the average of other models by 1.39%. Seman⁃

tic features demonstrated stable performance enhancement across different models, with F1 means in scenario one being 3.4% higher

than the mean of other methods, and in scenario two being 1.86% higher than the mean of other methods.【Innovation/limitation】By in⁃

tegrating semantic features with traditional bibliometric characteristics, this study employs a deep forest algorithm to predict and iden⁃

tify the potential academic impact of papers during their early publication stages. Using academic papers from the field of Chinese Li⁃

brary and Information Science as samples, the findings may not encompass all scientific disciplines.