Research On The Method Of Unsupervised Transfer Learning Based On Feature Distribution Discrepancy Adaptation

With the development of society,massive data has been made by us in daily life and work.Recently,along with the coming of the mobile internet,the production of data grows more and more rapidly,and the amount of data increases in an incredibly high speed.Machine learning methods can help people mine information from data.However,conventional machine learning methods assume that training set and test set follow the same statistical probability distribution.This hypothesis will result in poor performance in data mining tasks,for the training set and test set in reality always come from different probability distributions respectively.How to extract the most valuable information from large amount of data sets with different probability distributions has become a research focus in machine learning.Transfer learning tries to relax the hypothesis that the training set does not follow statistical probability distribution with the test set.As a result,the performance of the test set can be improved by knowledge migration from training set with various probability distribution.Joint distribution method assigns equal weight to the marginal discrepancy and the condictional discrepancy.However,this situation is invalid in reality.In addition,joint distribution method also assumes that the marginal discrepancy and the condictional discrepancy exists simultaneously.This hypothesis will result in neglecting the existence of single distribution discrepancy and poor performance.In this paper,we propose two novel transfer learning methods aiming at these two problems.The main work and innovation points are as follows:This paper analyzes the problem in present conventional transfer learning methods realize domain adaptation by simultaneously reducing the discrepancy of marginal distribution and conditional distribution,and varying weights between marginal distribution and conditional distribution in different tasks is neglected.In view of the above problem,this paper proposes compact cluster-based balanced distribution adaptation(CC-BDA).Firstly,balanced distribution factor assigns different weights to the marginal distribution discrepancy and the conditional distribution discrepancy respectively.And then,the balanced distribution factor is updated dynamicly in each iteration.In addition,the intra-class distance regulation of source domain and the intra-class distance regulation of target domain are simultaneously considered to enhance the separability between classes.The results on benchmark datasets indicate that the proposed CC-BDA method can effectively improve the classification accuracy.To solve the problem that the hypothesis of single distribution discrepancy existing independently and joint distribution discrepancy existing simultaneously results in poor performance of joint distribution alignment methods.A domain adaptation framework(Two-Stage balanced distribution adaptation,T-BDA)is proposed based on balanced distribution adaptation method.In the first stage,the magnitude of the marginal distribution discrepancy is estimated quickly.In the second stage,only the conditional distribution dicrepancy is considered if the marginal distribution discrepancy is little.If not,the marginal distribution discrepancy and conditional distribution discrepancy are joinly reduced.Experiments on standard datasets show that the proposed T-BDA framework can reduce distribution discrepancy adaptively and improve cross-domain classification accuracy.