Citation Counts Prediction of Publications Based on Gated Graph Convolutional Neural Network Model with Multi-Feature Fusion

Predicting citation counts is the process of estimating how frequently a paper will be cited based on the related information. It is significant for analyzing research trends and evaluating the impact of paper. Current research has employed multi-feature or citation relationships to predict citation counts but often ignores the interactions between them and fails to integrate their complementary information effectively. In this work, we propose a multi-feature fusion based gated graph convolutional network (MFF-GGCN) model, which combines textual features, statistical features, and citation relations. In MFF-GGCN, RoBERTa is employed to extract textual features for capturing semantic information from papers. And through a comprehensive analysis of the paper metadata, key statistical features are obtained to describe the metadata attributes and structural information of the paper. The citation relationships are represented as an adjacency matrix to capture the potential influence among papers. Finally, the adjacency matrix as well as the feature matrix consisting of textual features and statistical features are input into the gated graph convolutional network to regressively predict the citation counts of papers. The ex-periments are conducted on the LMANStat dataset, and the results show that the proposed model achieves an MAE value of 4.742 and an RMSE value of 6.903, which outperforms the eight baseline models.