基于相关性建模的图注意力多示例多标签学习

Graph Attention-Based Multi-Instance Multi-Label Learning with Correlation Modeling

摘要: 多示例多标签学习(Multi-Instance Multi-Label Learning, MIML)在图像分类、文本标注等任务中展现了独特优势，但现有方法在建模样本间复杂依赖关系和标签间相关性时仍存在不足。本文提出一种结合图注意力机制与对比学习的相关性建模多示例多标签学习方法(GAMIL-C)。首先，将每个包中的实例构建为图结构，利用图注意力网络(Graph Attention Network, GAT)建模实例间的结构关系，并引入边特征以刻画实例间的相似性。其次，在标签空间中通过相关性感知模块捕捉标签间的高阶依赖，以提升标签预测的准确性。进一步地，我们引入对比学习策略，在实例表征和标签表征之间构建判别性约束，从而增强模型的泛化能力。实验在多个公开数据集上验证了该方法的有效性，与主流基线方法相比，在预测精度和宏平均指标上均取得显著提升。结果表明，该方法能够有效建模实例与标签间的双重相关性，并通过对比学习进一步提升模型鲁棒性，为复杂场景下的多示例多标签学习提供了一种新的解决思路。

Abstract: Multi-Instance Multi-Label Learning (MIML) has shown great potential in tasks such as image classification and text annotation. However, existing approaches still struggle to effectively capture the complex dependencies among instances and the correlations among labels. In this paper, we propose a novel correlation modeling approach for MIML based on graph attention networks and contrastive learning (GAMIL-C). Specifically, we first construct a graph structure for the instances within each bag and employ a Graph Attention Network (GAT) to model the structural relationships among instances, where edge features are incorporated to characterize their similarities. Then, a correlation-aware module is introduced to capture high-order dependencies in the label space, thereby improving label prediction accuracy. Furthermore, a contrastive learning strategy is incorporated to establish discriminative constraints between instance and label representations, which enhances the generalization capability of the model. Extensive experiments conducted on several public datasets demonstrate that our method significantly outperforms strong baselines in terms of prediction accuracy and macro-averaged metrics. The results indicate that the proposed method effectively captures the dual correlations between instances and labels, while contrastive learning further improves model robustness, providing a new perspective for MIML in complex scenarios.