Fuzzy Transfer Learning in Heterogeneous Space Using Takagi-Sugeno Fuzzy Models

Abstract

Transfer learning is gaining increasing attention due to its ability to leverage previously acquired knowledge (a source domain with a large amount of labeled data) to assist in completing a prediction task in a related domain (a target domain with little labeled data). Many transfer learning methods have been proposed, and especially the fuzzy transfer learning method, which is based on fuzzy systems, has been developed because of its capability to deal with the uncertainty. However, there is one issue with fuzzy transfer learning that has not yet resolved: The domain adaptation methods for regression tasks in heterogeneous space are still scarce, and the relation of features in two domains have not been explored to assist the construction of target model. In this work, we proposed a new fuzzy transfer learning method, which constructs the transformed mappings for the domain-independent and domain-dependent features, separately. The existing fuzzy rules of the source domain are transferred to the target domain through modifying the input space using the mappings, and the parameters of the mappings are optimized by the few labeled target data. The experiments on real-world datasets validate the effectiveness of the proposed method and discuss the impact of some important parameters to the performance of the constructed target model.