Seminar on Computational Learning and Adaptation


 
K-Harmonic Means: A New Clustering Algorithm

Bin Zhang
Hewlett Packard Research Laboratory
Palo Alto, CA
bin_zhang2@hp.com

The most common algorithms for clustering, K-means (KM) and expectation-maximization (EM), are highly sensitive to how cluster centers are initialized. No good initialization method is known for EM or KM. We propose a new algorithm, named K-Harmonic Means (KHM). Experiments show that KHM is almost insensitive to how centers are initialized, and often maximizes the KM objective function better than KM itself. Instead of the minimum distance from a data point to the centers, as used by KM, KHM uses the harmonic mean of the distances from the data point to all centers.
This talk gives a unified mathematical and experimental analysis of KM, EM, and KHM, revealing their similarities, differences, advantages, and disadvantages, with visual animations of the convergence of all three algorithms.


Date: Thurs., November 16

Time: 4:15-5:30PM

Place: Cordura 100


Return to the seminar schedule