Domain-Invariant Label Propagation With Adaptive Graph Regularization

As an effective machine learning paradigm, domain adaptation (DA) learning aims to enhance the learning performance of the target domain by utilizing other relevant but distinct domain(s) (referred to as the source domain(s)). The existing mainstream methods for DA mainly learn discriminative domain-invariant feature representations by combining the "pseudo labels" of the target domain to better achieve knowledge transfer. However, most existing methods alternate the optimization learning of domain-invariant features and the updating of the "pseudo labels" into two different stages, which makes them difficult to achieve optimal learning performance. In order to achieve joint optimization learning of updating the "pseudo labels" and domain-invariant feature representations, a framework of Domain-Invariant Label prOpagation (DILO) with adaptive graph regularization is proposed. By combining semi-supervised knowledge adaptation and label propagation on domain data, DILO jointly optimizes domain-invariant feature representations and target learning tasks in a unified framework, allowing these two objectives to mutually benefit. Specifically, by introducing the concept of soft labels, a joint distribution measurement model is established to simultaneously alleviate both marginal and conditional distribution differences between different domains; constructing an adaptive probability graph model to enhance the robustness of label propagation. Moreover, a robust sigma -norm is applied to domain joint distribution measurement and inductive learning models to form a unified objective optimization formulation. An effective optimization algorithm is proposed for addressing the optimization problem of DILO. Compared with several representative DA methods, the proposed method achieved better or comparable robustness in adaptation learning on four cross-domain visual datasets.