基于监督信息特性的主动半监督谱聚类算法-深圳大学.pdfVIP

1 Vol . 38 No . 1 20 10 1 ACTA ELECTRONICA SINICA Jan. 20 10 王 娜, 李 霞 ( , 5 18060) : , , . , , , ASSC( Active Semi-supervised Spectral Clustering) . , , . UCI , ASSC . : : TP311, TP 187 : A : 0372-2112 ( 2010) 01- 0172-05 Active Sem-i supervised Spectral Clustering Based on Pairwise Constraints WANG Na, LI Xia (College of Inf ormation Engineering, Shenzhen University, Shenzhen, Guangdong 518060, China) Abstract : Sem-i supervised clustering uses a sm all amount of supervised data such as pairw ise con straint s to aid un superv ised learning . The impro ved clustering perf ormance depends heavily on the choice of constraints . This m akes it important to explore the appropriate pairw ise con straints f or semi- supervised cluster ing. This paper presents a m ethod f or actively selecting inf ormative pair- w ise constraint s, w hich corresponds to p ick up data pairs f ar ap art in the sam e cluster and those close in diff erent clusters . An active sem-i supervised spectral clustering ( ASSC) is then developed by utilizing the selected pairw ise con straint s to adjust the distance ma- trix in spectral clustering. A s a result , the intra- cluster distance is decreased and the inter- cluster distance is increased . Exp erim ental result s on UCI benchm ark data set s and artif icial data set show th at these inform ative p ariw ise constraints lead to sub stantial perf or- m ance enh ancement o ver the random selective p airw ise constraint s spectral clu stering. Key words: spectral clustering, sem-i supervised clustering , active learning, pairw ise cons