Research On Transfer Learning Based Collaborative Fuzzy Clustering For Multi-View Data

As an effective data mining method,clustering has been widely used in numerous fields,including e-commerce,wise medical,finance,and so on.In practical production and application,traditional clustering algorithms can effectively process data from the single feature set.However,with the development of intelligent technology,the emergence of multi-view data has brought severe challenges to traditional clustering algorithms.Thus,a large number of multi-view clustering algorithms have emerged.As one kind of multi-view clustering algorithms,the multi-view collaborative fuzzy clustering algorithm can efficiently perform data clustering with multiple views.This type of algorithm aims to mine the complementary information through collaboration between multiple views to promote clustering performance.However,the collaboration strategy of the existing multi-view clustering algorithm is relatively simple,which makes it difficult to effectively mine the potential complementary information.In addition,to ensure the effect of collaborative learning,these algorithms often involve a large number of parameters,which are hard to determine manually.In order to solve the above issues,this thesis applies the idea of transfer learning into the collaborative learning between multiple views,and proposes a series of improved algorithms.Firstly,this thesis draws on the idea of transfer clustering and presents a novel multi-view collaborative fuzzy clustering algorithm.This method adopts the membership matrix of different views as valuable knowledge to learn from each other among multiple views and introduces the membership constraint of multiple views to ensure the global consistency of multi-view clustering.The experimental results verify the effectiveness of the proposed transfer learning based multi-view collaborative fuzzy clustering algorithm.Secondly,since the learning factor of the above algorithm is a global variable,the clustering result is greatly affected by the initial value of this variable.In order to reduce the sensitivity of the above algorithms to the initial learning factors,this thesis designs a dynamic transfer learning strategy to promote the in-depth mining of potential information among multiple views.As a result,a dynamic transfer learning based multi-view collaborative fuzzy clustering algorithm is further proposed.The algorithm uses local learning factors to guide the transfer learning between any couple of views.At the same time,a parameter self-adjustment strategy is designed to automatically optimize the value of the learning factor.The experimental results illustrate that the parameter self-adjusting technology based multi-view collaborative fuzzy clustering algorithm achieves better and more stable clustering performance.Finally,considering the differential impact of various views on the clustering result,this thesis studies two improved algorithms,which are the transfer learning based multi-view weighted fuzzy clustering algorithm and the dynamic transfer learning based multi-view weighted fuzzy clustering algorithm.The two novel algorithms assign weights to each view,and use the maximum entropy regularization method to automatically optimize the weight distribution.In this way,the dynamic transfer learning method not only takes into account the local differences between views,but also considers the global differences between views during the process of knowledge transfer.Experimental results show that the introduction of view weights suppresses the negative impact of poorly separable views on global clustering,thereby further improving the clustering performance.