Multi-Source Multi-Target Domain Adaptation Based on Evidence Theory

Domain adaptation usually confronts the multiple-source and multiple-target domain issue. In such cases, the reduction of distribution discrepancy across domains and the combination of information in diverse domains are two major subproblems. Here, we propose a new method called multi-source multi-target domain adaptation based on evidence theory (MMET) to improve the accuracy. In MMET, we first develop a joint first- and second-order statistical distribution alignment approach to reduce distribution discrepancy. For a certain target domain, the other target domains are merged with it to yield multiple new target domains. Then, patterns in this target domain will obtain multiple domain-invariant feature representations by pairwise aligning the distributions of the source domain and each new target domain. For a query pattern in this target domain, it will obtain multiple soft classification results after employing the new distribution alignment approach. In order to integrate useful information in different target domains, the weighted average fusion (WAF) rule is used to locally combine the soft classification results, and multiple pieces of WAF results will be produced because of multiple source domains. For integration of information in different source domains, evidence theory (ET) is employed to globally combine these WAF results. MMET was compared with a variety of advanced methods, and the experimental results show that MMET can significantly improve the accuracy in each target domain.