Research On Heterogeneous Transfer Learning Based On Fuzzy System

Transfer learning is to enhance the performance of the model in the target domain by using the knowledge in the labeled data of the source domain,and the target domain contains only a few or unlabeled training samples.Because it takes a lot of manpower and material resources to manually tag data in the real world,using the tagged source domain to assist the target domain task can greatly reduce the consumption of manual tagging.At present,the mainstream transfer learning method assumes that the source domain and the target domain contain the same feature space,but it is difficult to collect the source domain and the target domain with the same feature space in practical applications.Therefore,the development of heterogeneous transfer learning has solved these limitations.In heterogeneous transfer learning,the feature space of the source domain and the target domain are different.When dealing with cross-domain transfer tasks,how to make up for the difference in feature space or distribution,and reduce the information loss caused by constructing the common feature space,will still face huge challenge.This paper proposes two new heterogeneous migration learning methods for the above problems.Compared with the existing heterogeneous transfer learning algorithms,the algorithm proposed in this paper has better heterogeneous transfer learning capabilities.The main works of this paper are as follows:1)The first work is for a new method of heterogeneous feature transfer with a small number of label samples in the target domain.Although the existing heterogeneous feature transfer learning methods have made important developments,they still face the following challenges: The process of feature transfer learning is poor in interpretability,which hinders its wider application.In the face of the above challenges,this paper proposes a heterogeneous feature transfer learning framework based on fuzzy inference rules,and implements a heterogeneous feature transfer learning method based on fuzzy logic inference based on this framework.The proposed method uses the antecedent mapping and subsequent mapping of the TSK fuzzy system to project the source domain and target domain features into a common feature subspace,and achieves effective feature transfer learning in the hidden space of fuzzy features,thus showing a better performance.The ability to transfer and learn heterogeneous features.A large number of experiments show that compared with the traditional heterogeneous transfer learning method,the learning ability of this method is highly competitive,and at the same time has better interpretability,so it has a wide application prospect.2)The second work is for a new method in heterogeneous domain adaption with a small amount of labeled examples in the target domain.Heterogeneous domain adaptation is a branch of heterogeneous transfer learning.Most of the existing heterogeneous domain adaptation methods are semi-supervised methods.These methods require that the target domain contains some labeled samples.However,such data sets are scarce in heterogeneous domain adaptive tasks.In order to solve the above problems,we propose an unsupervised heterogeneous domain adaptive method based on fuzzy rule learning.On the one hand,the proposed method is based on the rule learning of the TSK fuzzy system to learn the features of the source domain and the target domain respectively.By learning two feature transformation matrices,the source domain and the target domain are projected into a common feature subspace.On the other hand,in order to reduce the information loss caused by feature transformation,the proposed method adopts a variety of information preservation strategies and maximizes the correlation between the source domain data and the target domain data in the common feature subspace.Through experiments on several real domain adaptive data sets,the results show that the proposed algorithm has certain advantages over the existing heterogeneous domain adaptive methods.