Multi-Source Domain Adaptation Method Based On Self-Supervised Learning

Due to the growth in data size and computational resources,traditional machine learning has grown rapidly,however,traditional machine learning requires that training data follows an independent homogeneous distribution,which is overly strict in the real-world.Domain adaptation relaxes the constraint by enabling the transfer of knowledge from a related domain(source domain)to improve learning in the domain of the current task(target domain).The previous studies mainly focus on the Single-Source Domain Adaptation,which consists of only one source domain and one target domain.Whereas,as there may be multiple domains related to the target task,the single-source paradigm often fails to meet this situation.Therefore,multi-source domain adaptation has been receiving increasing attention.Multi-source domain adaptation aims to transfer knowl-edge from multiple labelled source domains to an unlabeled target domain to achieve accurate classification on the target domain.In multi-source domain adaptation,there are domain-shifts not only between source and target domains,but also between differ-ent source domains,which makes multi-source domain adaptation more challenging.The common framework of the existing multi-source domain adaptation methods consists of two key components:learning the models on the source domains and incor-porating target samples to fine-tune the models with various alignment strategies.Gen-erally speaking,the alignment strategies are the key for improving the generalization performance of the multi-source domain adaptation methods.In this paper,we focus on self-supervised multi-source domain-based domain adaptation.The existing self-supervised multi-source domain adaptation methods often suffer an imbalanced char-acteristic among the distribution of pseudo-labels.When the pseudo-label distribution differs too much from the true label distribution,it deteriorates the self-supervised train-ing of the model.Therefore,this paper balances the pseudo-labels from two perspec-tives:·To address the imbalance between the accuracy and number of pseudo-labels,we propose Self-Supervised multi-Source Domain Adaptation with Double Balance(S~3DA-db).Firstly,the target domain samples are ranked using information scores then top-ranked samples are selected to improve the accuracy of pseudo-labels and reduce the number of pseudo-labels.Secondly,labels with low percentage of samples are selected in the first step is enhanced using label weighting and sample enrichment,which slightly sacrifices the accuracy of the pseudo-labels but greatly improves the classification accuracy of the model on such labels.Experiments on five datasets validate the effectiveness of S~3DA-db.·To address the imbalance of pseudo-label proportions distribution,i.e.,the pro-portion between the numbers of pseudo-labeled samples and true labeled samples per-label on the target domain,often has an imbalanced characteristic,we pro-pose Self-Supervised multi-Source Domain Adaptation with Label-specific Con-fidence(S~3DA-lc).Specifically,we estimate the label-specific confidences,i.e.,the learning difficulties of labels,and adopt them to generate the pseudo-labels for target samples,enabling to simultaneously constraining and enriching the pseudo supervised signals for easy-to-learn and hard-to-learn labels.We evaluate S~3DA-lc on several benchmark datasets,indicating its superior performance compared with the existing MSDA baselines.