Learning Robust Feature Transformation for Domain Adaptation

摘要

There is a growing importance of feature extraction in transferring valuable knowledge from a source domain to a different but related target domain. However, when the target data are contaminated by unpredictable and complex noises, the ability of most existing feature extraction methods would be limited. In this paper, we deeply investigate the robust property of Kernel Mean P-Power Error Loss (KMPE-Loss), and thus propose a novel Robust Transfer Feature Learning (RTFL) method to enhance the robustness of domain adaptation. The key idea of RTFL is to learn a shared transformation by: 1) detecting and neglecting the contaminated target points without any specific assumption on noises; 2) reconstructing the remaining clean target points using the corresponding source-domain neighborhood; 3) incorporating a relative entropy based regularization to reap theoretic advantages. Consequently, the distribution difference between two domains is accurately reduced for knowledge transfer. We propose an alternative procedure to optimize RTFL with explicitly guaranteed convergence. As an extension, the transformation based matrix in RTFL is restricted to a small dimension basis, admitting the highly reduced computation complexity. Extensive experiments in various domain adaptation tasks demonstrate the superiority of our methods.