ML '94 WORKSHOP
	     Constructive Induction and Change of Representation


An appropriate representation is critical to the success of an inductive
learning task.  In difficult learning problems (eg protein folding, word
pronunciation, gene identification), considerable human effort is often
required to identify useful terms of the representation language.  In an
effort to make learning more autonomous, researchers have investigated the
problem of generating or modifying new representations automatically.

The past five years have seen a significant increase in the amount of work in
this area.  Some methods developed have been able to effect increases in
classification accuracy.  Others are able to derive features similar to those
discovered previously by humans.  Still other systems have demonstrated
impressive performance improvement through the construction of new
representations.

In spite of these successes, we are still far from understanding the range and
limitations of current methods, or the kind of representation change that
real-world domains may require.  The objective of this workshop is to examine
issues in current work and to review progress made so far.  The workshop will
also serve as a forum for the exchange of ideas among researchers actively
working on these issues.  Topics of interest include, but are not limited to,
the following:

 -  Empirical approaches and the use of inductive biases
 -  Use of domain knowledge in the construction and evaluation of new terms
 -  Construction of or from relational predicates
 -  Introduction of new terms by analytic theory revision systems
 -  Unsupervised learning and credit assignment in constructive induction
 -  Interpreting hidden units as constructed features
 -  New terms as indices in instance-based learning or case-based reasoning
 -  Constructive induction in human learning
 -  Experimental studies of constructive induction systems
 -  Theoretical proofs, frameworks, and comparative analyses
 -  Comparison of techniques from empirical learning, analytical
    learning, classifier systems, and neural networks


WORKSHOP FORMAT

The workshop will be held on Sunday, July 10th.  Attendance will be
open.  The workshop will consist of presentations of accepted papers
and a final panel discussion.  The panel will recap the workshop and
discuss the state of constructive induction and current open
questions.


SUBMISSIONS

Paper submissions should not exceed 3000 words (about six single-spaced pages,
including figures and tables, but excluding bibliography).  Four copies of
each paper should be sent to the contact address below.  Alternatively, one
copy of a postscript file may be sent via e-mail.  Each paper should include
an e-mail contact address of one of the authors.  The papers will comprise a
set of working notes, copies of which will be available at the workshop.

We encourage descriptions of work in progress as well as position papers.
Authors are encouraged to evaluate their systems on real-world domains and
to critique their methods with respect to the following questions:

- In your system, what is the relationship between the feature generation and
  induction?  Can the feature generation method be adapted to other forms of
  induction?

- How does the method evaluate or select the features that it generates?

- Is the method sensitive to the cost of the features?  Can it create
  features of unbounded expense?

- What real-world domain(s) has the method been applied to?  What
  characteristics of each domain makes feature generation useful or necessary?
  For what general class of domain might the method be useful?

- Can the method exploit existing domain knowledge?  What forms of domain
  knowledge can be exploited?

- What features are already known for the domain being addressed?  Can the
  method re-derive any of them?

- What forms (eg, propositional, relational, numerically weighted) can the
  generated features take?  Does this limit the method?


SCHEDULE

Paper submissions due                          25 April
Decisions made, submitters get feedback        22 May
Final working-note submissions due             15 June
Workshop                                       10 July



PROGRAM COMMITTEE

Tom Fawcett (chair), NYNEX Science and Technology 
James Callan, University of Massachusetts at Amherst
Chris Matheus, GTE Laboratories Inc.
Ryszard Michalski, George Mason University
Michael Pazzani, University of California at Irvine
Larry Rendell, University of Illinois at Urbana-Champaign
Rich Sutton, GTE Laboratories Inc.


CONTACT ADDRESS

Tom Fawcett
NYNEX Science and Technology
500 Westchester Ave.
White Plains, NY 10604
e-mail: fawcett@nynexst.com