Dynamic Adaptive Cross-Domain Mean Approximation

Cross-domain mean approximation (CDMA) is an efficient measure of distributional differences between domains. It measures the sample distribution difference between domains by calculating the distance from a sample in one domain to the sample mean in another domain, which in turn can facilitate cross-domain migration of knowledge. However, in practical applications, the marginal and conditional distributions of data are unbalanced, and CDMA pursues equally measuring the difference between the marginal and conditional distributions without considering the difference between the two, which leads to its inefficiency in transfer learning. For this reason, this paper firstly improves CDMA by introducing an adaptation factor and designing dynamic CDMA to evaluate the edge distribution error and conditional distribution error between the source and target domains. Secondly, on the basis of dynamic CDMA, a dynamic adaptive cross-domain mean approximation (DA-CDMA) feature extraction algorithm is proposed to extract features that are invariant between domains in order to realize cross-domain migration of knowledge. In addition, in order to reduce the shift of mean value caused by individual bad samples far away from the mean center during the feature extraction process, a mean update mechanism is proposed to update the mean value within the class to increase the stability and accuracy of migration. Finally, classification experiments are conducted on publicly available migration learning datasets to verify the effectiveness of the method.