A Hierarchical Filter Transfer Learning Algorithm for Heterogeneous Domains

A main goal of heterogeneous transfer learning algorithms is to solve the domain adaptation problem of different feature spaces. However, some existing heterogeneous transfer learning methods usually only extract common features from the source domain and target domain, ignoring specific features, which may damage the performance of transfer learning. Therefore, a hierarchical filter transfer learning algorithm (HFTLA) for heterogeneous domains in is proposed. First, a nonlinear mapping is constructed to learn the potential relationship between the features of different domains. Then, the feature space can be aligned by learning common features and specific features, which can ensure the integrity of the features. Second, a hierarchical filter framework is developed to play different roles in different stages of transfer learning. In the pretransfer phase, a knowledge filter based on genetic principle is designed to increase the diversity of knowledge with different genetic operators. In the post-transfer phase, a guided filter is established to achieve a coupling balance between source knowledge and target information. Finally, experimental results on heterogeneous domains illustrate the effectiveness of HFTLA.