Research On Transfer Learning Based Fuzzy Clustering

In the field of data mining,due to the advantages of simplicity,flexibility,and effectiveness,the clustering technique is widely used in various data mining tasks.But insufficient target data often result in bad clustering results,which are obtained by the traditional clustering algorithm.Among machine learning methods,the transfer learning is utilized to improve the training results when the target data are insufficient.This method improves the training results in the target domain by transferring useful knowledge from the source domain with good training model.Referring to the idea of transfer learning,the transfer clustering is proposed to handle the related clustering problems.Transfer clustering improves the clustering result of the target data by using the knowledge obtained from the source domain.In this thesis,based on the transfer clustering techniques,the existing centralized clustering and distributed collaborative clustering methods are improved to promote the clustering performance.The main innovative works and achievements are as follows.1.To improve the centralized clustering algorithms,the transfer learning based entropy-weighted fuzzy clustering algorithm(TEWFCM)is proposed for the high-dimensional data,and the transfer learning based kernel fuzzy clustering algorithm(TKFCM)is proposed for the non-linear data.In these algorithms,the cluster prototypes and dimension weights from the source domain are employed as knowledge and transferred to assist the clustering of target data.Experiment results prove the efficiency of the proposed methods.2.To improve the distributed collaborative clustering method,the transfer learning based distributed collaborative fuzzy c-means clustering algorithm(TCFCM)and the entropy-weighted transfer collaborative fuzzy clustering algorithm(W-TCFCM)are proposed for the data in distributed peer-to-peer networks.In these algorithms,the knowledge learning between the neighbor nodes is added into the traditional distributed collaborative clustering algorithms to improve the clustering results and accelerate the convergence of algorithms.Experiment results show that these proposed algorithms obtain higher clustering accuracy and less iterations compared with traditional distributed collaborative fuzzy clustering algorithms Soft-DKM and CDFCM.3.To enhance the universality of proposed algorithms,more efforts are made to further improve the transfer collaborative clustering.In the above transfer learning based distributed collaborative clustering algorithms,an important hyper-parameter is the learning factor.It is a global scalar which means the learning degree between neighbor nodes.The value of this parameter usually needs to be artificially set in advance.The experiments find that clustering performance of these algorithms are highly affected by the value of the learning factor,which greatly reduces the universality of the algorithms.In this thesis,a parameter-adjustable method is designed and two kinds of parameter-adjustable transfer distributed collaborative clustering algorithms(A-TCFCM and AW-TCFCM)are proposed.These methods define a learning factor between any two neighbor nodes instead of the global learning factor.Then the learning factors between neighbor nodes are adjusted according to the variation trend of group center of neighbor nodes in two successive iterations and the difference of data centers between neighbor nodes in an iteration.Experiments show that these improved methods achieve more stable clustering results rather than being influenced by the value of the learning factor.