Re: [Wekalist] measure the performance
by Nanda, Subrat (Research)
The following can be possible indicators of a clustering algorithm's performace:
1. As you have put it, execution time and memory spends can be vital in an environment where time is a constraint.
2. If you are doing supervised clustering, a clustering algorithm can be judged by various performance parameters like:
True False Rate,
Mis Classification errors,
RMS error, etc
3. If you are doing unsupervised clustering, then a clustering algorithm can be judged by thee parameters:
a. Level of human intervention such as supplying the 'k' in a k-means, or fuzzy functions in fuzzy clustering algorithms or
b. Thresholding in agglomerative/heirarchial clustering, etc.
c. Homogeneity in the clusters returned: Were all objects put in the same cluster really similar!
d. Sensitivity of the clustering convergence to externaly supplied information: eg, k value, etc
e. Stability of the returned clusters: Are the cluster memberships robust/rugged if you add/delete any attribute?
I am sure there can be more algorithm specific and your application specific methods to judge a clustering algorithm. But these should atleast give you some idea.
[mailto:firstname.lastname@example.org]On Behalf Of francisco
Sent: Monday, July 04, 2005 10:27 PM
Subject: [Wekalist] measure the performance
I liked to know, how can I measure the performance of the cluster algorithms!
For example, the execution times and the memory spends.
Do you know some way or some program!
Francisco Prata Paz
Wekalist mailing list