Introduction to Clustering Large and High-Dimensional Data
Focusing on a few of the most important clustering algorithms, this book provides a detailed account of these major models in an information retrieval context.
TABLE OF CONTENTS Introduction to Clustering Large and High-Dimensional Data
2.3. Quadratic k-means: Summary
2.3. Quadratic k-means: Summary
This section presents results of numerical experiments that illustrate the difference between the batch k-means algorithm, and the k-means algorithm (that combines batch and incremental iterations). We discuss properties of partitions generated by the two algorithms. Finally, we attempt to introduce an objective function dependent distance on the set of partitions generated by the k-means clustering algorithm.
2.3.1. Numerical experiments with quadratic k-means
The numerical experiments presented below compare performance of the batch k-means algorithm (Algorithm 2.1.1) to those of the k-means clustering algorithm (Algorithm 2.2.1). Both algorithms are applied to the same initial partitions. The initial partitions are generated by the Principal Direction Divisive Partitioning (PDDP, to be discussed in Section 5.1) with the maximal cluster size 25. We run the experiments on Medlars Collection and Cranfield Collection (available from http://www.cs.utk.edu/~lsi/). As a reference point we use the quality of partition generated by PDDP. A percentage of the decrease in the objective function Q is reported under improvement in Tables 2.1 2.3. In all the experiments we set tol B = tol I, and denote the tolerance by tol.
| Algorithm | Input parameters | # of clusters | Q | Improvement |
|---|---|---|---|---|
| PDDP | cluster size ?25 | 63 | 706.62 | 0% |
| batch k-means | tol = 0.005 | 63 | 671.68 | 5% |
| k-means | tol = 0.005 |
Copyright Jacob Kogan 2007 under license agreement with Books24x7
