基于余弦相似自适应加权视图重构的不完全多视图聚类算法

打开文本图片集

中图分类号：TP18文献标志码：A

Incomplete Multi-view Clustering Algorthm Based on Adaptive Weighing View Reconstruction with Cosine Similarity

CHEN Yongtaia，QIU Yeʰ，WAN Minghuaʰ，YANG Guoweiʰ （a.ScholofEngineringAudit，b.SchoolofComputerScience，NanjingAuditUniversity，Nanjing815，Jiangsu，Cina

Abstract：To solvetheproblems thatmulti-view dataoften contained mising or abnormal information inpractical applicationsandtheexisting incompletemulti-viewclustering methodsfailedtoadequatelyrepresenttheoriginaldatasimilarity intheprocessofdatasimilaritymatrixoptimization，increasedthecomputationalcomplexityandignoredthediferenceof discriminant infor-mation between views，an incomplete multi-view clustering algorithmbasedonadaptive weighting view reconstruction with cosine similarity was proposed.Natural alignmentof mising views was achieved by introducing alocal preservation reconstructionterm，which avoided negativeeffectsthat might be caused by filing missing views with mean values.Inthe initializationphase，the preservation of manifold structure of original multi-view data was enhanced by computing cosine similarityinthe original multi-viewspace，andanadaptive weighting strategy was employed to capture the importance of diferent views during thecomplete viewconstruction process.The clustering experiments were caried outin4benchmark datasetsand compared with theoptimal resultsof9existing repre-sentative algorithms.Theresults showthat theaverage clusteringaccuracy，normalizedmutual informationandpurityof theproposedalgorithmare improved by 5.52% ， 8.78% and 4.77% respectively，which fully reflects the excellent incomplete multi-view clustering performance of the proposed algorithm.

Keywords：incomplete multi-viewclustering；adaptive weighting；cosine similarity；manifold structure；viewreconstruction

在现实世界中，我们经常面对的是多个来源或由多个特征描述的数据，这种数据被称为多视图数据[1]。（剩余13939字）