Multidisciplinary Scientific Document Classification Based on Heterogeneous Graph Neural Networks

Abstract

In the era of extensive academic data, effectively categorize scientific publications from all fields. To solve the challenge of organization in multi-subject research articles, this paper presents a comprehensive approach using several graph neural networks. Create a large-scale, multi-relational graph that integrates relational structures and content elements in the suggested way for joint learning from topological and semantic viewpoints. In order to take into account connections between citations, semantics, and other meta-data in both global and detailed ways, the new approach simultaneously develops various message-passing techniques. The classification accuracy of this approach greatly surpasses that of traditional and deep learning baseline models, according to the experiment results of the dataset of 45,216 documents and 24 divisions. With a macro-F1 score of 0.833 and a total accuracy of 87.4%, the model outperformed the previous homogenous GNN approach by 4.1 percentage points. Increased cluster separation for both main and minor subjects is further demonstrated by embedding analysis, confirming the discriminative nature of the hybrid representation. According to the aforementioned tests, combining sophisticated Graph Neural Networks with a heterogeneous structure can enhance semantic abstraction and generalization for extensive scholarly work classification. According to the aforementioned research, graph neural networks can be used in large-scale scientific ecosystems to improve automated knowledge management's accuracy and efficiency.