posted on 1996-11-21, 00:00authored byThompson N. Doman, John M. Cibulskis, Michael J. Cibulskis, Patrick Dale McCray, Dale P. Spangler
Clustering of chemical inventories on the basis of
structural similarity has been shown to be useful in a
number of applications related to the utilization and enhancement of
those inventories. However, the widely-used Jarvis−Patrick clustering algorithm displays a number of
weaknesses which make it difficult to cluster
large databases in a consistent, satisfactory, and timely manner.
Jarvis−Patrick clusters tend to be either
too large and heterogeneous (i.e., “chained”) or too small and
exclusive (i.e., under-clustered), and the
algorithm requires time-consuming manual tuning. This paper
describes a computer algorithm which is
nondirective, in that it performs the clustering without manual tuning
yet generates useful clustering results.