The application of traditional clustering algorithms to partition chemical datasets is well established. Recently, clustering methods which partition a dataset based upon the eigenvectors of an input matrix have gained considerable attention in computer vision, providing excellent results for a variety of tasks. Despite this, their application to chemical data has been limited. A non-overlapping spectral clustering approach (L-NOSC), which utilizes a modified Lanczos algorithm to identify the eigenpairs of a matrix, has been developed. This L-NOSC affords considerable computational advantages when compared with other spectral clustering methods that rely on a full matrix diagonalization procedure. The ability of the L-NOSC algorithm to cluster several activity datasets, described by five different descriptors, has been evaluated using the Quality Clustering Index. Finally, the performance of the algorithm has been compared to both the leading traditional clustering methods and a spectral clustering algorithm which uses a full matrix diagonalization, with promising results.

Spectral clustering of chemical data: A Lanczos-based approach

243rd National Meeting (2012)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.