Multi-source Heterogeneous Data Aggregation Method Based on Adversarial Domain Adaptation

Abstract

In recent years, with the deployment of ubiquitous sensing, the aggregation method of massive multi-source heterogeneous data has become a hot research topic. At present, although the adversarial domain adaptation in transfer learning can achieve effective results in processing tasks such as data classification, there are few methods that can well apply the adversarial domain adaptation network to the scenario of multi-source heterogeneous data aggregation. The existing adversarial domain adaptation methods are mostly applied to the transfer of homogeneous features in single source domain. However, the samples in actual application scenarios are often heterogeneous data from multiple sources. To achieve feature alignment and the aggregation of multi-source heterogeneous data at the same time, a multi-source heterogeneous data aggregation method based on adversarial domain adaptation is proposed in this paper, by embedding a mapping neural network for heterogeneous data in the adversarial network, and adding weight parameters to measure the contribution of multiple source domains' features and classes. The feasibility of the network structure is analyzed theoretically, and the effectiveness of the method is verified through experiments.