Page 226 - Classification Parameter Estimation & State Estimation An Engg Approach Using MATLAB

7

Unsupervised Learning

In the previous chapter we discussed methods for reducing the dimen-
sion of the measurement space in order to decrease the cost of classifica-
tion and to improve the ability to generalize. In these procedures it was
assumed that for all training objects, class labels were available. In many
practical applications, however, the training objects are not labelled, or
only a small fraction of them are labelled. In these cases it can be worth
while to let the data speak for itself. The structure in the data will have to
be discovered without the help of additional labels.
An example is colour-based pixel classification. In video-based surveil-
lance and safety applications, for instance, one of the tasks is to track the
foreground pixels. Foreground pixels are pixels belonging to the objects of
interest, e.g. cars on a parking place. The RGB representation of a pixel can
be used to decide whether a pixel belongs to the foreground or not. How-
ever, the colours of neither the foreground nor the background are known
in advance. Unsupervised training methods can help to decide which pixels
of the image belong to the background, and which to the foreground.
Another example is an insurance company which might want to know
if typical groups of customers exist, such that it can offer suitable
insurance packages to each of these groups. The information provided
by an insurance expert may introduce a significant bias. Unsupervised
methods can then help to discover additional structures in the data.
In unsupervised methods, we wish to transform and reduce the data
such that a specific characteristic in the data is highlighted. In this

Classification, Parameter Estimation and State Estimation: An Engineering Approach using MATLAB
F. van der Heijden, R.P.W. Duin, D. de Ridder and D.M.J. Tax
Ó 2004 John Wiley & Sons, Ltd ISBN: 0-470-09013-8