CFP: ECML'98 WS - Upgrading Learning to the Meta-Level

[Melanie Hilario]

[Our apologies if you receive multiple copies of this CFP]

                                                                         
                               Call for Papers
                              ECML'98 Workshop

                    UPGRADING LEARNING TO THE META-LEVEL:
                   MODEL SELECTION AND DATA TRANSFORMATION


                     To be held in conjunction with the
                10th European Conference on Machine Learning
                      Chemnitz, Germany, April 24, 1997


                http://www.cs.bris.ac.uk/~cgc/ecml98-ws.html


Motivation and Technical Description

Over the past decade, machine learning (ML) techniques have successfully
started the transition from research laboratories to the real world. The
number of fielded applications has grown steadily, evidence that industry
needs and uses ML techniques. However, most successful applications are
custom-designed and the result of skillful use of human expertise. This is
due, in part, to the large, ever increasing number of available ML models,
their relative complexity and the lack of systematic methods for
discriminating among them. Current data mining tools are only as
powerful/useful as their users. They provide multiple techniques within a
single system, but the selection and combination of these techniques are
external to the system and performed by the user. This makes it difficult
and costly for non-initiated users to access the much needed technology
directly.

The problem of model selection is that of choosing the appropriate learning
method/model for a given application task. It is currently a matter of
consensus that there are no universally superior models and methods for
learning. The key question in model selection is not which learning method
is better than the others, but under which precise conditions a given method
is better than others for a given task.

The problem of data transformation is distinct but inseparable from model
selection. Data often need to be cleaned and transformed before applying (or
even selecting) a learning algorithm. Here again, the hurdle is that of
choosing the appropriate method for the specific transformation required.

In both the learning and data pre-processing phases, users often resort to a
trial-and-error process to select the most suitable model. Clearly, trying
all possible options is impractical, and choosing the option that appears
most promising often yields to a sub-optimal solution. Hence, an informed
search process is needed to reduce the amount of experimentation while
avoiding the pitfalls of local optima. Informed search requires
meta-knowledge, which is not available to non-initiated, industrial
end-users.

Objectives and Scope

The aim of this workshop is to explore the different ways of acquiring and
using the meta-knowledge needed to address the model selection and data
transformation problems. For some researchers, the choice of learning and
data transformation methods should be fully automated if machine learning
and data mining systems are to be of any use to non specialists. Others
claim that full automation of the learning process is not within the reach
of current technology. Still others doubt that it is even desirable. An
intermediate solution is the design of assistant systems which aim less to
replace the user than to help him make the right choices or, failing that,
to guide him through the space of experiments. Whichever the proposed
solution, there seems to be an implicit agreement that meta-knowledge should
be integrated seamlessly into the learning tool.

This workshop is intended to bring together researchers who have attempted
to use meta-level approaches to automate or guide decision-making at all
stages of the learning process. One broad line of research is the static use
of prior (meta-)knowledge. Knowledge-based approaches to model selection
have been explored in both symbolic and neural network learning. For
instance, prior knowledge of invariances has been used to select the
appropriate neural network architecture for optical character recognition
problems. Another research avenue aims at augmenting and/or refining
meta-knowledge dynamically across different learning experiences.
Meta-learning approaches have been attempted to automate model selection (as
in VBMS and StatLog) as well as model arbitration and model combination (as
in JAM). Contributions are sought on any of the above--or other--approaches
from all main sub-fields of machine learning, including neural networks,
symbolic machine learning and inductive logic programming.

The results of this workshop will extend those of prior workshops, such as
the ECML95 Workshop on Learning at the Knowledge Level and the ICML97
Workshop on Machine Learning Applications in the Real World, as well as
complement those of the upcoming AAAI98/ICML98 Workshop on the Methodology
of Applying Machine Learning.

Format and Schedule

The workshop will consist of one invited talk, a number of refereed
contributions and small group discussions. The idea is to bring researchers
together to present current work and identify future areas of research and
development.

This is intended to be a one-day workshop and the proposed schedule is as
follows.

   9:00      Welcome
   10:00     Paper session (5 x 30mins)
   12:30     Lunch
   1:30      Paper session (3 x 30mins)
   3:00      Summary: the issues/the future
   3:15      Small group discussions (3-4 groups)
   4:00      Reports from each group
   4:45      Closing remarks
   5:00      End

Timetable

The following timetable will be strictly adhered to:

   * Registration of interest: starting now (email to: Christophe G-C,
     please specify intention to attend/intention to submit a paper)
   * Submission of paper: 6 March 1998 (electronic postscript only to 
     either organiser: Christophe G-C, Melanie H)
   * Notification of acceptance: 20 March 1998
   * Camera-ready: 28 March 1998

Program Committee

Submitted papers will be reviewed by at least two independent referees from
the following program committee.

     Pavel Brazdil, University of Porto
     Robert Engels, University of Karlsruhe
     Dieter Fensel, University of Karlsruhe
     Jean-Gabriel Ganascia, Universite Pierre et Marie Curie
     Christophe Giraud-Carrier, University of Bristol
     Ashok Goel, Georgia Institute of Technology
     Melanie Hilario, University of Geneva
     Igor Kononenko, University of Ljubljana
     Dunja Mladenic, Josef Stefan Institute, Slovenia
     Gholaremza Nakhaizadeh, Daimler-Benz
     Ashwin Ram, Georgia Institute of Technology
     Colin Shearer, Integrated Solutions Ltd
     Walter van de Welde, Riverland Next Generation
     Maarten van Someren, University of Amsterdam
     Gerhard Widmer, Austrian Institute for Artificial Intelligence Research

Accepted papers will be published in the workshop proceedings and
contributors will be allocated 30 minutes for an oral presentation during
the workshop.

Organisers

     Christophe Giraud-Carrier
     Department of Computer Science
     University of Bristol
     Bristol, BS8 1UB
     United Kingdom
     Tel: +44-117-954-5145
     Fax: +44-117-954-5208
     Email: cgc at cs.bris.ac.uk

     Melanie Hilario
     Computer Science Department
     University of Geneva
     24, Rue General-Dufour
     CH-1211 Geneva 4
     Switzerland
     Tel: +41-22-705-7791
     Fax: +41-22-705-7780
     Email: Melanie.Hilario at cui.unige.ch