基于特征校正的多对抗域适应方法

Multi-adversarial domain adaptation method based on feature correction

领域自适应可以通过对齐源域和目标域的分布将有标签的源域信息迁移到没有标签但相关的目标域。然而,现有的大多数方法仅对源域和目标域的低层特征分布进行对齐,无法捕获样本中的细粒度信息。基于此,提出了一种基于特征校正的多对抗域适应方法。该方法在引入注意力机制以突出可迁移区域的基础上,通过部署特征校正模块对齐两个域之间的高级特征分布,进一步缩小域差异。此外,为了避免单个分类器过度拟合其自身的噪声伪标签,还提出了双分类器协同训练,并利用图神经网络特征聚合的特性生成更精准的源域标签。在3个迁移学习基准数据集上的大量实验证明所提方法的有效性。

Domain adaptation can transfer labeled source domain information to an unlabeled but related target do-main by aligning the distribution of source domain and target domain.However,most existing methods only align the low-level feature distributions of the source and target domains,failing to capture fine-grained information within the samples.To address this limitation,a feature correction-based multi-adversarial domain adaptation method was pro-posed.An attention mechanism to highlight transferable regions was introduced in this method and a feature correc-tion module was deployed to align the high-level feature distributions between the two domains,further reducing domain discrepancies.Additionally,to prevent individual classifiers from overfitting their own noisy pseudo-labels,dual classifier co-training was proposed and the feature aggregation property of graph neural networks was utilized to generate more accurate source domain labels.Extensive experiments on three benchmark datasets for transfer learn-ing demonstrate the effectiveness of the proposed method.