面向通用异构域适应的部分最优传输模型

Partial Optimal Transport for Universal Heterogeneous Domain Adaptation

摘要: 通用异构域适应指在源域和目标域的数据分布不同且标签集部分重叠时，通过迁移学习的方法，将源域的知识迁移到目标域。然而现有方法大多依赖同构特征空间假设或全局对齐，难以在通用异构域场景下实现精确的语义匹配与私有类拒绝。为此，本文提出一种面向通用异构域适应的部分融合最优传输模型。首先，针对异构空间中跨域距离难以直接度量问题，引入关键点引导机制，构建具有语义一致性的跨域关系代价函数。在此基础上，引入Gromov-Wasserstein距离保持域内几何结构，实现异构特征空间的分布匹配，引导跨域样本结构上的几何对齐。其次，为在共享部分标签情况下精准对齐公共类别并拒绝私有类别，将模型扩展为部分最优传输形式，通过限制总传输质量，仅对公共类别进行选择性匹配，从而有效拒绝目标域中的私有类样本。此外，引入了一种基于标签的组稀疏正则项，以诱导传输计划在类别维度呈现群稀疏性，进一步抑制私有类样本的误匹配。最终，本文构建了一个融合几何结构对齐、语义引导匹配与私有类拒绝的统一优化模型，并设计了一种结合Majorization-Minimization框架与Frank-Wolfe方法的高效求解算法。在三个跨域数据集上的实验结果表明，所提模型在公共类分类和私有类检测上均优于主流对比方法，验证了其在处理通用异构域适应问题上的有效性。

Abstract: Universal heterogeneous domain adaptation aims to transfer knowledge from a source domain to a target domain with distinct data distributions and partially overlapping label sets. However, most existing methods rely on homogeneous feature space assumptions or global alignment, making it difficult to achieve precise semantic matching and private class rejection under universal heterogeneous domain scenarios. To address these issues, this paper proposes a partial fused optimal transport model for universal heterogeneous domain adaptation. First, to directly measure cross-domain distances in heterogeneous spaces, a keypoint-guided mechanism is introduced to construct a semantically consistent cross-domain relational cost function. Building on this, the Gromov-Wasserstein distance is incorporated to preserve intra-domain geometric structures, enabling distribution matching across heterogeneous feature spaces and guiding geometric alignment of cross-domain samples at the structural level. Second, to accurately align common classes and reject private classes under partially shared labels, a partial optimal transport formulation is adopted, which restricts the total transported mass to selectively match only common categories, thereby effectively rejecting private-class samples in the target domain. Third, we introduce a label-based group-sparsity regularization, inducing structured sparsity in the transport plan at the category level to further suppress private-class mismatches. Finally, we formulate a unified optimization model that integrates geometric structure alignment, semantically guided matching, and private-class rejection. Then an efficient solver based on the Majorization-Minimization framework combined with the Frank-Wolfe method is designed. Experiments on three cross-domain datasets show that the proposed model outperforms competing methods in both public-class classification and private-class detection, demonstrating its effectiveness for general heterogeneous domain adaptation.