Geoscience Reference
In-Depth Information
The final sections of this chapter are devoted to a recent geographical example of the use of
GPGPU, which is an emerging area of parallel computing with great potential for GC research.
GPU has primarily been used in the area of geostatistics as described earlier, but classification of
large data sets is a problem that lends itself to speed-up through parallel implementation.
3.7 GPGPU GEODEMOGRAPHIC INFORMATION SYSTEM
This section provides an example to illustrate how
k
-means, a common algorithm used in the
creation of geodemographic classifications, can be enhanced to run in parallel over a GPU. We
then evaluate how this parallel implementation can be further enhanced by different normalisa-
tion procedures and offer further performance improvements upon the standard
k
-means clustering
that utilises a CPU only. Although of obvious general theoretical interest, this evaluation tackles a
specific applied problem in GC of geodemographic classifications: that is, none of the established
algorithms provide sufficiently efficient means to create national or regional geodemographic clas-
sifications within an acceptable time frame. As such, in assessing the advantages of this approach,
we evaluate the ways in which the parallel
k
-means clustering algorithm can be applied to the com-
putation of geodemographic classifications online and in real time.
This work exploits the CUDA feature of recent NVIDIA graphics card: a general-purpose parallel
computing architecture that uses programs written in C or C++. A typical CUDA-enabled graphics
card has a number of GPU and a memory capacity capable of storing a large amount of data. For
example, the GeForce 8400M GT graphics card has 16 GPU and 512 MB of internal memory. CUDA
requires that the computational problem be programmed in the C language for parallel processing.
Our case study can be seen in the wider context of parallel implementations of
k
-means using
CUDA: for example, Takizawa and Kobayashi (2006) have proposed a parallel
k
-means solution
for solving image texture size problems, and Hall and Hart (2010) have proposed a parallel solution
for solving the problem of limited instance counts and dimensionality in the analysis of complex
shapes. However, these implementations only work in specified environments, and there are as yet
no global parallel
k
-means solutions that are suitable for creating geodemographic classifications.
Geodemographic classifications provide summary measures of socio-economic conditions in
small neighbourhood areas, typically based upon weighted combinations of census variables (Harris
et al. 2005). There are multiple geodemographic classifications that have been devised as general-
purpose indicators to gauge the levels of social similarity between neighbourhoods in individual
countries and even some that transcend international boundaries.* Classifications are usually struc-
tured into a series of hierarchical levels that represent the typical characteristics of an area. The
characteristics of each class within a typology are usually summarised by a label (e.g.
city living
),
a verbal
pen portrait
and other descriptive material such as photographs, montages and videos to
give end users of the classification a clearer understanding of the characteristics of the underlying
populations. In recent years, concerns have been raised over both the need for data that are up to
date (Adnan et al. 2010) and also the applicability of closed source general-purpose classifications
for public service applications such as health or education.
Geodemographic classifications have been developed in a range of national settings. Burrows and
Gane (2006) review Jonathan Robbin's pioneering work on computer-based geodemographics in the
United States, where he created PRIZM using funding from the US Department of Housing and Urban
Development. PRIZM was focused on the allocation of housing grants between cities that have a his-
tory of rioting (Weiss 2000) and is now owned by the Nielsen Company (Burrows and Gane 2006).
Harris et al. (2005) describe some of the geodemographic tools that have been developed in many
countries of the world, including Australia, China, Denmark, Finland, France, Germany, Greece,
Japan, the Netherlands, New Zealand, Spain, Sweden, Norway, United Kingdom and United States.
*
For example, Mosaic Global (http://www.experian.co.uk/business-strategies/mosaic-global.html), which integrates data
from 24 national classifications.
