CALL FOR PAPERS
Intelligent Data Analysis - IOS Press
SPECIAL ISSUE on
INCREMENTAL LEARNING SYSTEMS CAPABLE
OF DEALING WITH CONCEPT DRIFT
*****************************************************************
Please distribute this announcement to all interested parties.
Special issue Editors:
Miroslav Kubat, University of Miami, USA
João Gama, University of Porto, Portugal
Paul Utgoff, University of Massachusetts, USA
Suppose the existence of a concept description that has been induced
from a set, T, of training examples.
Suppose that later another set, T', of examples become available.
What is the most effective way to modify the concept so as
to reflect the examples from T'?
In many real-world learning problems the data flows continuously and
learning
algorithms should be able to respond to this circumstance.
The first requirement of such algorithms is thus incrementality,
the ability to incorporate new information.
If the process is not strictly stationary,
the target concept could gradually change over time,
a fact that should be reflected also by the current version of the
induced concept description.
The ability to react to concept drift can thus be viewed as a natural
extension
of incremental learning systems.
These techniques can be useful for scaling-up learning algorithms to
very large datasets.
Other types of problems were these techniques could
be potentially useful include: user-modelling, control in dynamic environments,
web-mining, times series, etc.
Most of evaluation methods for machine learning (e.g. cross-validation)
assume that
examples are independent and identically distributed. This assumption
is clear
unrealistic in the presence of concept drift.
How can we estimate the performance of learning systems under these
constrains?
The objective of the special issue is to present the current status
of algorithms,
applications, and evaluation methods for these problems.
Relevant techniques include the following (but are not limited to):
1. Incremental, online, real-time, and
any-time learning algorithms
2. Algorithms that learn in the presence
of concept drift
3. Evaluation Methods for dynamic instance
distributions
4. Real world applications that involve
online learning
5. Theory on learning under concept drift.
Submission Details:
We are expecting full papers to describe original, previously unpublished
research,
be written in English, and not be simultaneously submitted for publication
elsewhere
(previous publication of partial results at workshops with informal
proceedings is allowed).
We could also consider the publication of high-quality surveys on these
topics.
Please submit a PostScript or PDF file of your paper to:
[EMAIL PROTECTED]
Important Dates:
Submission Deadline: 1 of February 2003
Author Notification: 1 of July 2003
Final Paper Deadline: 1 of September 2003
Special Issue:
_______________________________________________
