Discriminant Feature Transformation for Domain Adaptation

Abstract:
In recent years, more and more feature-based transfer methods have been proposed to solve the domain adaptation problems. The common feature-based transfer approaches are characterized by learning a feature transformation. However, the transformation may distort the data structure and class structure, thus lead to a non-ideal transfer result. In this paper, we proposed a novel transfer learning approach that learns a discriminative feature transformation by iterative label refinement with constraint measurement. Specifically, we jointly reduce the domain distribution discrepancy and minimize the weighted inter-class distance as well as maximize the weighted intra-class distance to adapt cross-domain features with preserved data and class structure. Extensive experiments are conducted to evaluate the effectiveness of the proposed method on 14 cross-domain tasks. The analysis demonstrates that the proposed method can significantly improve the classification accuracy over several state-of-the-art algorithms.