Summer School on Inductive Logic Programming and Knowledge Discovery in Databases
Prague, Czech Republic, 15-17
John Lloyd, University of Bristol
A two and a half day ILP & KDD
Summer School was held at the beginning of the ILP Week in Prague'97. Later in the same
week the ILP'97 Workshop and an Area Meeting of CompulogNet on 'Representation Issues in
Reasoning and Learning' also took place.
The Summer School was intended for
students and researchers in Artificial Intelligence, Databases, Logic Programming, Machine
Learning, and Statistics. It was also intended for end-users from commerce and industry
who face problems extracting high-level information from databases. The organisers were
Saso Dzeroski and Nada Lavraťc from the Jozef Stefan Institute, Ljubljana, Slovenia. The
School consisted of a mixture of lectures and hands-on laboratory sessions. About 45
people attended, including lecturers.
The first session was presented by Nada
Lavraťc and consisted of an introduction to and overview of ILP. The main ideas of
inductive concept learning were presented and then the generalisation to ILP was outlined.
Some basic learning algorithms and a few of the more important theoretical concepts were
discussed. Well-chosen examples illustrated the main ideas. Overall this session provided
a beautifully clear foundation for the remaining sessions of the Summer School.
Next, Saso Dzeroski gave an overview of
ILP applications. These applications ranged over dynamic systems, ecology, mechanical
engineering, molecular biology, music, natural language processing, and software
engineering. This session was quite an eye-opener - I knew that ILP had had some notable
successes in applications, but I hadn't realised just how large the range of applications
was. This success with applications is a strong indication of the health of ILP.
After lunch, Stefan Wrobel gave an
introduction to Data Mining and Knowledge Discovery in Databases from an ILP perspective.
KDD is a rapidly expanding area with much industrial and commercial interest. During this
session, participants had their first chance of some hands-on experience by trying out the
Kepler system which is an extensible KDD tool developed at GMD in Germany.
An excellent first day of the Summer
School was rounded off by a visit to the extraordinary Municipal Hall of Prague which has
recently been restored in the most sumptuous fashion.
Day 2 started out with some lectures
and laboratory sessions with the Progol system led by Stephen Muggleton. A first
interesting experiment on the issue of comprehensibility involved the well-known Michalski
Trains problem. The excellent new graphical user interface of the Progol system made this
and the other experiments easy to undertake.
The afternoon of Day 2 was devoted to
work of the Leuven group led by Luc De Raedt. The Claudien, ICL and Tilde systems
developed there were presented and the participants were able to try out these systems on
some small-scale examples. The range of problems addressed by these systems is impressive
and includes classification and data mining.
Day 3 of the Summer School, which
finished at lunch time, was devoted to KDD with the lecturer being Usama Fayyad of
Microsoft Research. He gave an exceptionally interesting overview of KDD which included an
insightful discussion of many general issues and some details of successful applications.
One was left with the impression that the KDD bandwagon, which is currently in full
flight, is unlikely to be slowing down soon!
As a newcomer to machine learning, I
found the Summer School valuable. In two and a half days, I was able to get a really
excellent idea of the contribution of ILP to this fascinating area. The School had the
right balance of lectures and laboratory sessions, and the emphasis on practical
application of the research results kept the issues relevant and interesting. One could
have a few quibbles with certain aspects - more computers would have helped; in some
laboratory sessions too much time was wasted creating files and moving them around; and
not enough time was allocated to explaining the algorithms underlying some of the systems
so one was left marvelling at the output of some systems but having little idea how it was
produced! - but these are comparatively minor points. It may also be worth considering the
possibility of expanding future presentations of the School to 3 or 4 days. This would
allow participants more time to get to know the various ILP systems better.
For those who didn't attend the
workshop the slides are on the Web at: http://www-ai.ijs.si/~ilpnet. This site is
definitely worth a visit. All in all, the School was a big success and a tribute to a
growing and dynamic research community. I hope to see it given again in the near future.
Department of Computer Science
University of Bristol, U.K.