Entropy minimization and domain adversarial training guided by label distribution similarity for domain adaptation

In domain adaptation, entropy minimization is widely used. However, entropy minimization will bring negative transfer when the pseudo-labels are inconsistent with the real labels. We hope to increase pseudo-label accuracy to counter negative transfer in entropy minimization. To this end, we introduce domain adversarial training into entropy minimization. Furthermore, we consider the misalignment caused by domain adversarial training under severe label shift. Therefore, we propose method called entropy minimization and domain adversarial training guided by label distribution similarity (EMALDS). Through domain adversarial training which focus more on class-aligned divergence, our method improves pseudo-label accuracy and reduce negative transfer in entropy minimization. Extensive experiments demonstrate the effectiveness and robustness of our proposed method.