| Description |
Demonstration of learning in versions spaces
The CANDIDATE-ELIMINATION algorithm
was designed to detect relevant features in data sets and can therefore
easily be used for automatic classification tasks. The example presented
on this page shall visualize, how the algorithm is able to learn the concept
you use to classify a given set of data.
Imagine you are the personnel manager of a big company. Every day a
plenty of applications reach your desk, but you don't have enough time
to invite all of the candidates - you don't even have much time for
selection! Maybe you could teach someone to filter the application
flood before, but the requirements of the company change very
drastically and it would be much more comfortable to have an automatic
classification system, which has the power to classify candidates
depending on your actual decision concept, which is learned by the
system before by simply watching your actions.
Description of the applet
On this page you can check out such a classification system. For the
case of simplicity the criterions for classification are limited to
four attributes with just a few possible values. Every candidate is
represented by an icon showing a smiley, and by four values, one for
each attribute. You can explore the characteristics of each candidate
by clicking on the representing icon. The active icon has a red frame
and its properties are listed below. To classify one of the
unclassified candidates just click on the related icon and click
"Accept" or "Reject" afterwards. When the system has learned enough of
your concept, one active classification may indirectly affect the
status of some other candidates, which are automatically classified by
the system. The icons of all classified candidates (no matter if
classified directly or automatically) of the last turn are highlighted
by a yellow frame, so you can inspect the properties of automatically
classified candidates one after another to check, whether you would
have decided the same way. You can retract your last choice by using
the "Undo"-Button, for example to compare what happens, if you change
your mind about the last candidate. To avoid problems with
contradictive classifications there is no way to change any of the
classifications once made in another way than by "Undo". As a help to
see the yet unclassified candidates, their icons are additionally
framed with a thin white line.
The actual versions space can be inspected in the two text fields on
the bottom. The upper one shows the set of the most specific
hypotheses, which is always one, as soon as you accept one candidate.
This is a result of the fact that for each attribute there is a
unique generalization of every two attribute values.
The lower text field shows the most general hypotheses. If you refuse
to accept a candidate, this set can become very large, so you are best
off to accept a candidate first.
The procedure ends, when all of the candidates are classified.
Note: You need at least JAVA 1.1.
|
