Item – Theses Canada

​
OCLC number
480658355
Link(s) to full text
LAC copy
LAC copy
Author
Bennett, Andrew,1980-
Title
Density-based and consensus clustering in ecology.
Degree
M. Sc. -- Laurentian University, 2007
Publisher
Ottawa : Library and Archives Canada = Bibliothèque et Archives Canada, [2008]
Description
2 microfiches
Notes
Includes bibliographical references.
Abstract
The first chapter, 'Density versus hierarchy', explores the potential for density-based clustering methods in ecology. Clustering is fundamental to pattern recognition in science, but methods vary widely and can strongly affect the patterns detected. Non-hierarchical and hierarchical clustering methods are well known to ecologists, but both force a spherical structure on groups. New to ecology, density-based methods define partitions with local densities instead of group centers, allowing analysts to perceive oddly distributed groups in noisy data. To explore the full parameter space of one density-based method (DBSCAN), I developed a tool called the K-curve. I use K-curves on ecologically relevant synthetic data and find that density-based methods equal or surpass center-based methods. Important concepts are illustrated both with the 2-dimensional spatial point pattern of a tropical tree population in Panama and with a similar pattern in a population of Blanding's turtles in central Ontario. Results suggest that density-based clustering can define meaningful spatial ranges for either plants or animals, weighted by the probability of occupancy. An important problem in every cluster analysis is the choice of parameters, but existing algorithms to optimize partitions are not compatible with the oddly-shaped groups extracted by density-based clustering. The second chapter describes new consensus methods I invented to optimize partitions at multiple scales for any kind of clustering output. 1:1 consensus is appropriate for comparisons of similar partitions. Nested consensus, either one-way or two-way, compares dissimilar partitions. Tests on synthetic data find that 1:1 consensus is sensitive to important group structure at multiple scales, whether used on center-based or density-based partitions. Other tests find that nested consensus can unite density-based and center-based partitions to meaningfully describe tortuous groups. This unifying approach is appropriate for multivariate analyses in high-dimensional spaces, as encountered in studies of niche, functional type, and community. Density-based consensus clustering should be very useful in ecology, famous for odd distributions, noisy data, and multiple scales of interest. * This thesis is composed of two chapters, each a self-contained article prepared for submission to a journal. An appendix gives the background in cluster analysis necessary to fully appreciate the new methods described in the thesis. Matlab functions to perform the analyses are appended to the thesis on a CD and are also freely available from Andrew Bennett.
ISBN
9780494315200
0494315202