Research On New Transfer Clustering Algorithms

Cluster analysis has a long history.Over the past decades,its importance and cross-cutting characteristics with other research directions have been affirmed.Clustering is one of the important research contents of data mining,pattern recognition and other research directions,and plays an extremely important role in identifying the internal structure of data.Among them,the prototype-based clustering points out the solution direction for the poor performance of clustering algorithm in the face of insufficient data.Transfer clustering participates in the clustering process of target domain by source domain(auxiliary domain)knowledge or data,which makes the traditional clustering algorithm use more information to achieve data calibration of target domain.In particular,transfer prototype-based clustering as a representative transfer clustering technology shows good performance.Meanwhile,with the continuous enrichment of large data and the diversification of transfer scenarios,it is difficult to apply transfer prototype-based clustering universally.Therefore,the theory and method of transfer prototype-based clustering are still far from perfect,and need to be further completed urgently.In view of the above challenges,the following research work has been carried out in this paper.Firstly,this paper proposes a possibility matching based knowledge transfer prototype clustering framework,which can extend the single-perspective clustering analysis to the clustering analysis assisted by relevant domain information.At the same time,compared with the fuzzy knowledge matching prototype-based clustering framework,the framework introduces the possibility measure transfer mechanism to transfer source domain knowledge more rationally,so as to further improve performance of prototype-based clustering.Secondly,with the diversification of transfer scenarios,the source domain and target domain feature space are no longer distributed in the same way,that is heterogeneous distribution(X_s?X_t),which makes the current transfer clustering cannot effectively involve the source domain knowledge directly in the target domain clustering task.In this paper,a new heterogeneous feature transfer learning algorithm is proposed,which can transfer learning in heterogeneous feature space.Combining with transfer prototype clustering,a framework of heterogeneous transfer prototype fuzzy clustering is proposed to explore clustering analysis in heterogeneous transfer scenarios.Finally,several specific transfer prototype-based clustering algorithms are implemented based on the above two frameworks.The experimental results show that the transfer prototype-based clustering framework based on probability knowledge matching in homogeneous scenes shows better performance than the existing related algorithms.Heterogeneous transfer prototype-based fuzzy clustering framework in heterogeneous scenarios effectively expands the application scenarios of transfer prototype clustering.