Cross-domain Adaptive Algorithm Without Domain Information

With the advent of the big data era,human activities have generated a large amount of data.This is largely due to the popularity of networking devices and social networking.Since data processing is less efficient than data generation,a large number of unlabeled data cannot be processed in time.These data should not be labeled manually to obtain category information,because manual labeling can hinder data processing.Therefore,it is common practice to transfer annotation information to new data using well-trained models.However,the result of this approach is often unsatisfactory because of the existing model with domain-specificity.Thus,unsupervised domain adaptive algorithm is proposed.As one of the most popular research,this algorithm extracts transferable category information from labeled data in the source domain and transfers this information to unlabeled data in the target domain.In order to apply this algorithm to multiple source domains,unsupervised multi-source domain adaptive algorithm is developed.However,almost all multi-source domain adaptive algorithms are limited by domain-separable assumptions and cannot be applied to mixedsource scenarios.A small number of multi-source adaptative algorithms without this assumption are only suitable for specific application scenarios because of other assumptions.In this paper,a cross-domain adaptive algorithm without domain information is proposed based on the category homomorphism assumption.Under this assumption,different source domains have different category distributions and data noise patterns,but the same category of all source domains has the same data generation mechanism.That is to say,there exists a feature transformation function,which makes all source domains have a consistent decision space.If this space is similar to the hidden space of the target domain,the category information of multiple source domains can be migrated to the target domain.In order to obtain this feature transformation function,the proposed algorithm not only constrains the output feature to preserve category information but also aligns the expected distribution of source domain and target distribution.Considering the loss of category information,the proposed algorithm chooses to align the intra-class feature distribution.In order to align intra-class distribution,this paper provides two strategies to evaluate distribution discrepancy.The first strategy is to minimize the total mean square deviation,which can centralize intra-class instances to the cluster center.The second strategy is to minimize the maximum mean discrepancy,which can align the intra-class distributions of domains.Under the unsupervised scenario,the category space of the target domain is inaccessible.Therefore,all the target instances are given pseudo labels from the model.In addition,considering that the feature transformation function may be complex,the proposed algorithm uses multi-task learning to train convolution neural networks.The comparative experiments show that,without the help of source domain information,the category-homomorphism-based cross-domain adaptive algorithm is still better than other methods in general.In addition,the strategy based on maximum mean discrepancy is slightly better than that based on the total mean square deviation.The visualization of hidden features shows that the proposed algorithms can well retain the category information of data and thus achieve robust performance.