Computer Science Department, New Jersey Institute of Technology (NJIT), Newark, NJ 07102, USA
Abstract:
Clustered SVD-CSVD, which combines clustering and singular value decomposition (SVD), outperforms SVD applied globally, without first applying clustering. Datasets of feature vectors in various application domains exhibit local correlations, which allow CSVD to attain a higher dimensionality reduction than SVD for the same normalized mean square error. We specify an exact method for processing k-nearest-neighbor queries for CSVD, which ensures 100% recall and is experimentally shown to require less CPU processing time than the approximate method originally specified for CSVD.