In recent years, graph neural networks (GNNs) have emerged as a dominant paradigm for learning from graph-structured data, achieving strong performance across a wide range of graph mining tasks, including node classification, clustering, and relational inference. Despite these advances, most existing GNN architectures primarily rely on message-passing mechanisms that aggregate smoothed high-order local neighborhood information. While effective for capturing local proximity patterns, such designs often struggle to model richer multi-level graph structures, particularly in scenarios where both fine-grained local interactions and holistic global organization are essential. From a structural perspective, graph representation learning inherently involves information distributed across multiple levels. Local proximity cues, such as node degrees and common neighbors, capture immediate relational patterns, whereas global structural characteristics, including long-range dependencies and overall graph topology, provide complementary high-order information. However, conventional GNN models typically entangle these heterogeneous signals within a unified aggregation process, leading to suboptimal representations. To address this limitation, we propose DAGRL, a dual attention-based graph representation learning framework that explicitly disentangles and integrates multi-level structural information. The proposed model employs dedicated graph neural components to learn structure-specific representations and utilizes a dual attention mechanism to adaptively fuse them into a unified embedding space. Extensive experiments demonstrate that DAGRL consistently outperforms existing methods across multiple benchmark datasets.
My et al. (Fri,) studied this question.