Call for Participation in NIPS 2002 Workshop:
"Beyond Classification and Regression: Learning Rankings,
Preferences, Equality Predicates, and Other Structures"
Organizers:
Rich Caruana & Thorsten Joachims
Date and Location:
Fri Dec 13 or Sat Dec 14, Vancouver/Whistler, BC.
Submissions:
Extended abstract by Sun Nov 10
URL: http://www.cs.cornell.edu/People/tj/ranklearn
http://www.nips.cc
Description:
Not all supervised learning problems fit the standard classification
and regression function-learning model. Some problems require that we
predict things other than values or classes. For example, sometimes
the magnitude of the values predicted for cases are not important, but
the ordering/ranking induced by these values is important. Sometimes
we don't know a priori what classes exist, but we need to learn which
items belong to the same class. Sometimes it is important to retrieve
the top 10 objects, and no one cares what ordering is predicted for
the remaining 100,000 objects. Sometimes the quality of learning will
be judged by measures such as Precision and Recall or ROC that are not
well optimized by standard value prediction models.
Mismatch from the value prediction learning model can arise not only
with the predictions, but also can arise in the form of the training
examples. For example, when a user indicates that one document is more
relevant than another, or that two documents should be in the same
class, but does not assign values to the documents themselves, the
notion of what constitutes a training example is turned on it's
head. In these situations a training example is a relation on pairs of
what traditionally would have been considered independent training
cases. This change has deep ramifications.
This workshop aims to explore supervised learning problems that go
beyond the usual value prediction model. In particular, it addresses
problems where either (a) the goal of learning or (b) the input to the
learner, are more complex structures than in classification and
regression. Examples of such problems include:
- learning a partial/total ordering
- learning equality / matching
- learning to optimize non-standard criteria such as ROC area or
precision/recall
- using relative preferences as training examples
- ordinal regression
- learning to cluster with feedback or constraints
- learning graphs and other structures
- problems that require these approaches (e.g., text retrieval, medical
decision making, protein matching)
The goal of the workshop is to create a forum for discussing recent
methods and results, and to inspire research on new algorithms and
problems.
Invited speakers include:
- William Cohen
- Ralf Herbrich
- Guy Lebanon
- Dragos Margineantu
- Mike Mozer
- Alex Niculescu
- Foster Provost
- Yoram Singer
