<p>Citation intent prediction aims to identify the purpose behind an author’s citation of a specific work, serving as a core task in academic text processing. Traditional methods fail to fully utilize and effectively integrate multi-dimensional interactions among heterogeneous elements of citation contexts. To address the challenges, we propose a Heterogeneous Graph Fusion Network (HGFN) for multi-intent citation prediction. A Citation-Focused Context Optimization module is adopted to refine textual semantic information of the citation context. We construct a heterogeneous graph of all citation contexts and use the Graph Feature Fusion Attention mechanism to fuse the node representations generated by Heterogeneous Graph Transformer (HGT). For a citation, HGFN deeply fuses the textual representation and the representation of the heterogeneous subgraph of the citation, based on which a classifier is formed to predict the citation intent. Experiments on public datasets covering both multi-intent and single-intent tasks demonstrate that HGFN’s performance significantly surpasses current mainstream baseline models. This not only provides an effective and general new paradigm for citation intent recognition but also showcases its framework’s powerful potential in handling multi-source heterogeneous information.</p>