
HP Labs Technical Reports
Click here for full text:
Scale Up CenterBased Data Clustering Algorithms by Parallelism
Zhang, Bin; Hsu, Meichun
HPL20006
Keyword(s): parallel algorithms; data mining; data clustering; KMeans; KHarmonic Means; Expectation Maximization
Abstract:As data collection increases at an accelerating rate
with the advances of computers and networking
technology, analyzing the data (data mining) becomes
very important. Data clustering is one of the basic
tools widely used as a component in many data mining
solutions. Even though many data clustering
algorithms
have been developed in the last few decades, they
face
new challenges in front of hugh data sets. Algorithms
with quadratic (or higher order) computational
complexity, like agglomerative algorithms, drop out
very quickly. More efficient algorithms like KMeans
and EM, which have linear cost per iteration, also
need scaleup before they can be applied to very
large
data sets. This paper shows that many parameter
estimation algorithms, including the clustering
algorithms like KMeans, KHarmonic Means and EM,
have
intrinsic parallel structure in them. Many
workstations over a LAN or a multipleprocessor
computer can be efficiently used to run this class of
algorithms in parallel. With 60 workstations running
in parallel (on a fast LAN), clustering 28.8 GBytes
of
40 dimensional data into 100 clusters, the
utilization
of the computing units is above 80%.
23 Pages
Back to Index
