An adaptive multi-task learning method with task-specific attention for node classification and link prediction

Abstract

Multi-task learning has shown great potential for joint modeling of node classification and link prediction on graph-structured data. However, many existing multi-task GNNs still rely on rigid shared representations with limited adaptive fusion capability, making it difficult to effectively integrate global aggregation and local attention and to properly disentangle task-specific representations, which in turn leads to negative transfer. In addition, task imbalance during optimization can cause one objective to dominate training, resulting in unstable convergence and degraded performance. To address these challenges, we propose an adaptive multi-task learning framework for node classification and link prediction, called AMTGG, which adaptively coordinates both tasks within a unified architecture. Specifically, AMTGG constructs shared representations via parallel GCN and GAT branches, and introduces learnable fusion coefficients to adaptively combine their complementary global–local cues. Then, task-specific attention layers and gating mechanisms are employed to automatically select task-relevant subspaces from the shared features, thereby alleviating representation interference between objectives. To balance the training dynamics of the two tasks, GradNorm is adopted to dynamically adjust task weights, while EMA smoothing is further applied to stabilize weight evolution and improve convergence. Since link prediction involves modeling the interaction between node pairs, which is often underemphasized in multi-task frameworks, our model incorporates subgraph extraction and line graph transformation to enhance edge-level structural learning. Extensive experiments on seven benchmark graph datasets demonstrate that AMTGG enhances model robustness through feature masking regularization and noise injection, and achieves synergistic improvements on both node classification and link prediction.