Clustering Datasets with Singular Value Decomposition
Douglas, Emmeline P.
MetadataShow full item record
Spectral graph partitioning has been widely acknowledged as a useful way to cluster matrices. Since eigen decompositions do not exist for rectangular matrices, it is necessary to find an alternative method for clustering rectangular datasets. The Singular Value Decomposition lends itself to two convenient and effective clustering techniques, one using the signs of singular vectors and the other using gaps in singular vectors. We can measure and compare the quality of our resultant clusters using an entropy measure. When unable to decide which is better, the results can be nicely aggregated.