Confidence Sample Selection and Specific Feature Enhancement for Domain Adaptation

For domain shifts, domain adaptation (DA) promotes label propagation by reducing distributional differences. In recent work, DA only enhances the discriminability of linear data by linear discriminant analysis, which ignores real-world nonlinear data. Meanwhile, these methods fail to consider the negative impact of the target low-confidence samples on the training process. Therefore, confidence sample selection and specific feature enhancement for domain adaptation (CSS-SFE) is proposed. Firstly, the framework selects target high-confidence samples through the min-max principle to reduce the impact of incorrect pseudo-label during training; Secondly, the class scatter matrix and neighbor scatter matrix are balanced to enhance the features of linear and nonlinear datasets so that the discriminability of the samples is improved; Thirdly, the framework maintains source and target specific features by learning different projection matrices, which prevents the discriminability of the samples decreasing; Fourthly, the marginal and conditional distribution alignment is further applied to reduce the domain distribution discrepancy; Finally, extensive experiments on several benchmark datasets demonstrate the superiority of CSS- SFE over state-of-the-art methods.