Scholarlink

<p>Graph Neural Networks (GNNs) have gained significant traction in deep multi-task learning (MTL) due to their superior expressive power. The Hierarchical Graph Neural Network (HGNN) enhances multi-task feature learning by constructing task-specific GNNs and class-specific GNNs to model inter-task and inter-class relationships. However, the graph structure of HGNN relies on initial features and fixed label similarities, which inadequately captures the evolving relationships within the data during training, ultimately constraining the classification performance. This limitation fails to capture the dynamic relationships within the data, ultimately restricting the model’s classification performance. To address this, we propose CC-HGNN, a novel multi-task learning framework that synergizes GNNs with contrastive learning. First, CC-HGNN introduces a cross-task supervised contrastive learning mechanism to refine node embeddings by pulling intra-class samples closer and pushing inter-class samples apart in the feature space. This process optimizes the discriminative power of node representations. Second, the optimized sample embeddings are aggregated to obtain task-level and class-level node representations; an attention mechanism then performs weighted message passing to further enhance the embeddings. The CC-HGNN is highly generalizable and can be integrated into various deep multi-task architectures. To validate the effectiveness of CC-HGNN, we conducted extensive experiments on three benchmark datasets: ImageCLEF, Office-Caltech-10, and Office-Home. The results demonstrate that CC-HGNN consistently outperforms existing methods, achieving superior classification accuracy. Additionally, visualization analyses reveal that CC-HGNN achieves clearer cluster separation in the feature space.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CC-HGNN: enhanced multi-task feature learning via cross-task contrastive learning and graph neural networks

  • 期刊论文

  • 发表日期：

    2026年3月5日

查看全文
  • Yi Xu,
  • Zhipeng Shu

摘要

Graph Neural Networks (GNNs) have gained significant traction in deep multi-task learning (MTL) due to their superior expressive power. The Hierarchical Graph Neural Network (HGNN) enhances multi-task feature learning by constructing task-specific GNNs and class-specific GNNs to model inter-task and inter-class relationships. However, the graph structure of HGNN relies on initial features and fixed label similarities, which inadequately captures the evolving relationships within the data during training, ultimately constraining the classification performance. This limitation fails to capture the dynamic relationships within the data, ultimately restricting the model’s classification performance. To address this, we propose CC-HGNN, a novel multi-task learning framework that synergizes GNNs with contrastive learning. First, CC-HGNN introduces a cross-task supervised contrastive learning mechanism to refine node embeddings by pulling intra-class samples closer and pushing inter-class samples apart in the feature space. This process optimizes the discriminative power of node representations. Second, the optimized sample embeddings are aggregated to obtain task-level and class-level node representations; an attention mechanism then performs weighted message passing to further enhance the embeddings. The CC-HGNN is highly generalizable and can be integrated into various deep multi-task architectures. To validate the effectiveness of CC-HGNN, we conducted extensive experiments on three benchmark datasets: ImageCLEF, Office-Caltech-10, and Office-Home. The results demonstrate that CC-HGNN consistently outperforms existing methods, achieving superior classification accuracy. Additionally, visualization analyses reveal that CC-HGNN achieves clearer cluster separation in the feature space.