<p>Link prediction is a fundamental problem in social network analysis and mining, with broad practical applications. While deep learning–based graph embedding techniques have achieved significant success, most existing methods focus on undirected graphs, despite the fact that many real-world networks, such as citation networks and social media followership, are inherently directed. In directed graphs, edge direction carries essential semantic information, yet current methods typically use local aggregation strategies that overlook global structure. Additionally, conventional negative sampling strategies further limit predictive accuracy. To address these challenges, we propose Status-Aware Dual-Channel Graph Attention Embedding (SADGE), which introduces a status-aware dual-channel attention mechanism that initializes node importance using PageRank and dynamically refines status scores during training, capturing both hierarchical structures and directional asymmetry by enhancing information flow from low-status to high-status nodes while suppressing the reverse. Additionally, we propose a virtual status negative sampling strategy that generates more challenging negative samples by considering both status discrepancies and embedding similarities, significantly improving the model’s discriminative power. Evaluating SADGE on five benchmark directed networks demonstrates its superior performance compared to state-of-the-art methods, particularly in sparse networks and those with many low-degree nodes, thus validating its robustness and effectiveness in directed graph representation learning.</p>