ICDM'04 Panel
Fosca Giannotti
ICDM 2004 Panel Chair
Data Mining -- Where to go?
Organizer: Arno Siebes
Question 1:
Is there a future in industrial applications?
-- When data mining started, part of the hype was that DM
would make data analysis simple, with results that could
be easily understood. Is that still true for, e.g., ensembles of
support vector machines?
-- In marketing applications it is enough to be better than before
(or the competition). Have we gone too far into the search for
perfect models in lieu of simple good-enough models?
-- The quality measures we use in algorithms are often inspired
by algorithmic advantages. Are monotone constraints actually
important for applications?
Question 2:
Will Data Mining go down under its own succes?
-- Data Mining is running the risk of being shattered
The number of application areas with their own workshops,
conferences and journals is proliferating. Take
Bioinformatics as an example. Lots of interesting work is
done their. However, it is an autonomous community. Some
of our community also publish their results, but people from
Bioinformatics don't seem to cross-over into our community.
With a few more of such areas, Data Mining will cease to be
a homogeneous community.
-- Similarly, if you want to keep up with all the research going
on in Data Mining proper, you also have look at the database conferences
and journal, the computational Statistics conferences and journals
and so on.
Question 3:
Are there any interesting problems left?
-- All major conferences have seen a huge rise in the number
of submissions. But many of these papers are epsilon papers.
If we tinker a bit more on some aspect, we can get epsilon
better results on a certain data set under certain conditions.
Researchers have to focus on smaller and smaller subareas
to be able to reach any results. This means that conferences
give hardly anymore new inspiration and progress is stalled.