Active Semi-supervised Domain Adaptation Via Non-maximal Degree Node Suppression

Deep neural networks can easily learn the intrinsic patterns from a large number of labeled data,but it is difficult to generalize the learned knowledge to a new domain.To alleviate the above problem,the concept of domain adaptation is proposed.It aims to mitigate the discrepancy between the source and target domains,and thus enables knowledge transfer of models from the source domain to the target domain.Although unsupervised domain adaptation methods have achieved significant performance improvement on the target domain,they still face a performance bottleneck.Thus,we need to spend a small amount of annotation cost to obtain few labeled data to maximally enhance the model performance on the target domain.In this paper,we propose an active learning-based domain adaptation framework.Specifically,we first propose an active strategy called Non-maximal Degree Node Suppression(NDNS)to select representative and diverse samples from the target domain for labeling.NDNS first creates a directed graph using the target data by defining acceptive neighbor and accepted neighbor,then selects the target samples by iteratively performing maximum node retrieval and nonmaximum node removal.In addition,we propose an Asymmetric Minimax Entropy(AME)approach to calibrate the source domain classifier to fit the target domain distribution.AME balances the label distribution of model output by introducing an auxiliary variable.Further,we reduce the KL divergence between model output and auxiliary variables to calibrate the model.Finally,for those samples with high confidence,we perturbate them and encourage the model to produce consistent predictions.Finally,we conduct comprehensive experiments on three benchmark data sets to demonstrate the effectiveness of our proposed method.