|
Activity Number:
|
383
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Wednesday, August 9, 2006 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Section on Statistical Computing
|
| Abstract - #306546 |
|
Title:
|
Clustering by Intersection-Merging
|
|
Author(s):
|
Qunhua Li*+ and Marina Meila
|
|
Companies:
|
University of Washington and University of Washington
|
|
Address:
|
Department of Statistics, Seattle, WA, 98195-4322,
|
|
Keywords:
|
model-based clustering ; mixture model ; expectation-maximization ; hierarchical agglomerative clustering ; simulation annealing
|
|
Abstract:
|
We propose Intersection-Merging (IM), a wrapper algorithm to improve the initial clusterings in model-based clustering. The algorithm takes a set of clusterings obtained e.g. by EM, breaks down the clusterings into subclusterings via an intersection step, and then agglomerates them via a merging step. We introduce two versions of merging: greedy (standard IM) and by simulated annealing (IMSA). We apply the algorithm to the Gaussian mixtures of several synthetic and real data sets. The results show that both IM and IMSA improve on the starting clusterings under a variety of criteria.
|
