ARN-GAT: an adaptive node and relation modeling framework for patent knowledge graph topic identification

The rapid growth of patent data has created an urgent need for effective methods to identify key technological topics within large-scale Patent Knowledge Graphs (PKGs). This capability is essential for tracking innovation trends and supporting strategic decision-making. However, extant approaches are encumbered by three major limitations: uniform node representations that overlook structural heterogeneity, insufficient modelling of multi-relational semantics, and prohibitive computational costs in estimating graph robustness. In order to address these shortcomings, we propose ARN-GAT, an adaptive relational neural framework built upon GATv2. ARN-GAT introduces three innovations. Firstly, a layer-wise dynamic skipping mechanism is employed to alleviate oversmoothing. Secondly, the DiGRAF adaptive activation function is used to enhance nonlinear expressivity across heterogeneous node types. Thirdly, a relation–node joint adaptive gating mechanism is used to dynamically weight semantic relations while filtering noisy edges. The collaborative operation of these components results in the production of more discriminative node embeddings. Furthermore, the framework integrates graph robustness metrics into an end-to-end ranking objective, thereby enabling stable identification of structurally critical nodes. Extensive experimentation on four real-world PKG datasets demonstrates that ARN-GAT consistently outperforms state-of-the-art baselines in terms of Top-5% accuracy identification and significantly reduces ranking volatility. The proposed framework demonstrates strong generalisation and scalability, offering an effective solution for key node identification in large-scale knowledge graphs, with potential applicability to evolving settings.