Deep adversarial reconstruction classification network for unsupervised domain adaptation

Although the existing adversarial domain adaptation methods have been successfully applied in the unsupervised domain adaptation community, their performances may perhaps be weakened due to a significant distributional diversity of the target domain caused by the absence of the local features of samples in the target domain after the domain adaptation process. In this paper, based on the well-known autoencoder, a single-input with multi-output model called deep adversarial reconstruction classification network (DARCN) is developed to circumvent the above issue. The proposed DARCN mainly consists of the following four modules: a feature extractor for extracting the domain-invariant features along with the local features of the target domain, a predictor for estimating the label, a domain discriminator for distinguishing between the source and target domains, and a decoder for reconstructing the original data. The standard backpropagation method can be effectively used to optimize the proposed model. The experimental results indicate that DARCN realizes the improved classification performances in most cases in contrast to some existing comparative methods.