Contrastive Learning-based Bi-classifier Unsupervised Domain Adaptation Model

For deep domain adaptation issues, redundant information in feature representation causes poor model performance. A bi-classifier domain adaptation model was proposed based on contrastive learning. Based on the theory of bi-classifier learning, the input data twice was enhanced to obtain the features from two views, and the diversity of classifiers was improved by inputting features of different perspectives into different classifiers. At the same time, by closely combining the bi-classifier method and contrast learning, the model was able to capture high-level semantic representations of the data, and reduce the confusion degree between feature from different class. Finally, the samples were recognized by the proposed model at classification boundary correctly by aligning the label distribution. Experimental results verify that the contrastive loss between two classifiers can extract valid information from the data, thereby improving model performance.