Research On Multiple Source Transfer Learning

It is a challenging on how to effectively use the decision knowledge from multiple source domains to predict the labels of instances in the target domain.This thesis has researchd this problem,and the main work is divided into three aspects:1)In this study,aiming at the collaborative utilization of transfer information on domainwise and pointwise,a novel doubly reweighting multi-source transfer learning framework is proposed,called DRMTL.DRMTL aims to simultaneously optimize the structural risk function,domain reweighting adaptation,pointwise reweighting adaptation and manifold consistency.The merits of DRMTL include:(1)the importance of every source domain can be evaluated using the proposed novel flexible weighting index;(2)the loss between an unknown label prediction and its prediction by some source decision function for a target sample can be reweighted by using a novel domain separator;(3)the manifold structure of target domain is effectively used in this framework.Finally,a specific learning algorithm DRM-RLS is proposed using the DRMTL framework and the classical regularized least-squares classifier.Our experimental results on several real-world datasets reveal that the proposed approach outperforms several state-of-art transfer learning algorithms.2)In order to make better use of Universum data and avoid large-scale matrix inversion during model training,this thesis proposed the MTL-BDI method,which combines theε-insensitivity criterion and Universum learning.All source domain samples are used as Universum data and re-weighted according to their proximity to target classification hyperplane,so that they can better guide the determination of target classification hyperplane.At the same time,sparse regularization based on ε-insensitivity loss is used to avoid large-scale matrix inversion.The convergence of the algorithm is proved.Extensive experiments on text and image datasets validated that the proposed approach can achieve better performance than other multi-source transfer learning.3)In order to effectively identify Electroencephalogram signals in multiple source domains,this thesis proposed the MS-TL-TSK method,which combines together multiple source transfer learning and manifold learning mechanisms into Takagi-Sugeno-Kang fuzzy system.Specifically,the advantages of MS-TL-TSK include:(1)An adaptive reweighting strategy based on the theory of Shannon entropy is presented to control the transfer significance of source each domain;(2)Using the theory of pointwise transfer learning;(3)By taking into account the manifold regularization term,the manifold structure of the target domain is effectively maintained in the proposed system;(4)By inheriting the interpretability of TSK fuzzy system,MS-TL-TSK has good interpretability that would be understandable by domain experts for identifying EEG signals.The effectiveness of the proposed fuzzy system is demonstrated on several EEG multiple source transfer learning problems.