Abstract
Spectral methods have been used successfully in many clustering applications. While different algorithms lead, in general, to different clustering results, each single algorithm also produces distinct clusterings
depending on parameter initialization. Criteria for the automatic selection of these parameters have been addressed. We show in this paper that these methods, and new criteria herein presented, cannot solve this problem.
We then propose to combine clustering ensembles produced by a spectral algorithm in order to obtain robust solutions. Based on recently proposed evidence of the accumulation technique, which uses the single link method to extract combined data partitions, we further explore this idea using other methods. Experimental results tesitfy to the better performance of the combination strategy as compared to individual results produced by spectral methods, overcoming the problem of parameter selection
Users
Please
log in to take part in the discussion (add own reviews or comments).