Announcement/Call for Participation

Attribute-Value and Relational Learning:

Crossing the Boundaries

A Workshop at the 
Seventeenth International Conference on Machine Learning (ICML-2000)
Stanford University
June 29-July 2, 2000

The purpose of this workshop is to investigate the boundaries between learning with attribute-value representations and relational ones. There are several reasons for being interested in these boundaries.

In recent years, people from the attribute-value learning community have used richer representations for learning to tackle practical problems that are hard to represent within the attribute-value formalism. One example of such work concerns multi-instance learning, in which each example corresponds to a set of tuples in a single relation. This formulation is more expressive than the usual attribute-value setting, which requires each example to be a single tuple, but less expressive than the typical relational setting, which allows for multiple relations as well.

Another line of research working toward the boundaries concerns propositionalization in inductive logic programming. Various researchers in this area have proposed ways to derive propositional features from relational problems and then used these features successfully in attribute-value learners.

Boundaries between attribute-value and relational learning have not only been crossed in traditional symbolic machine learning but also in areas such as probabilistic reasoning, case-based reasoning, and even reinforcement learning. For instance, some work has extended methods for learning in Bayesian networks to handle relational representations, and methods for analogical reasoning often employ a first-order or relational representation.

As in the knowledge representation community, many researcers in machine learning and data mining are concerned with the boundaries between relational and propositional representations. The issue under investigation is often the trade-off between complexity of the algorithms and expressiveness of the representation languages.

This workshop hopes to present recent research results in this area, to make progress on understanding the boundaries, to bring together researchers from different communities, and to stimulate fruitful discussions among participants.

A list of topics of interest includes (but is not limited to):

• multi-instance learning
• propositionalization of relational formalisms
• using intermediate representations for learning (e.g., some description logics)
• frameworks that capture the relation among attribute-value and relational learning
• empirical and/or theoretical results on the trade-off between complexity of algorithms and expressiveness of representations
• extensions of probabilistic and case-based learning to support relational formalisms
• background knowledge in attribute-value and relational learning
• hybrid approaches that merge the two approaches to learning
• extensions of attribute-value learning and specialisations of relational learning to handle sequential and hierarchical data
• neural network learning with structured data

Workshop Organization and Submission Requirements

The workshop will feature a number of invited presentations by people who have contributed to the area. A preliminary list of invited speakers includes:

Lorenza Saitta (Universita di Torino, Italy)

Michele Sebag (Ecole Polytechnique, Palaiseau, France)

Ashwin Srinivasan (University of Oxford, UK)

Mark Craven (University of Wisconsin, USA)

Daphne Koller (Stanford University, USA)

Russ Greiner (University of Alberta, Canada)

Researchers wishing to present their own results at the workshop should submit an extended abstract, no longer than 2000 words, to skramer@informatik.uni-freiburg.de, deraedt@informatik.uni-freiburg.de, subo@informatik.uni-freiburg.de, preferably in HTML, PDF or Postscript. The abstracts will be put on the Web before the workshop. Submissions will be judged mainly on their relevance to the workshop topic, i.e., they should make explicit their contribution to the exploration of the boundaries between attribute-value and relational learning. Abstracts that focus solely on either side of the boundary will not be accepted for presentation.

To guarantee a true workshop atmosphere, the workshop will be restricted to 50 participants. Researchers interested in participating should send an email (including address and email) to subo@informatik.uni-freiburg.de to register  for the workshop. If more than 50 persons are interested in participating in the workshop,  participants will be selected on a first-come first-serve basis.

Important Dates

Submission deadline: April 17, 2000
Notification of acceptance: May 15, 2000.
Workshop held: July 2, 2000

Schedule

8:45 -	9:00	Opening Remarks (L. De Raedt, S. Kramer)
9:00 -	9:30	Invited Talk: L. Saitta
9:30 -	9:50	P. Flach, N. Lavrac The Role of Feature Construction in Inductive Rule Learning
9:50 -	10:20	Invited Talk: M. Sebag: Genetic Programming, Background Knowledge and Stochastic Grammars
10:20 -	10:30	Discussion
10:30 -	11:00	Coffee Break
11:00 -	11:20	J. Lloyd A Logical Setting for the Unification of Attribute-Value and Relational Learning
11:20 -	11:50	Invited Talk:  M. Craven
11:50 -	12:10	E. Armengol, E. Plaza On Sorts and Relations in Feature Term Learning
12:10 -	12:20	Discussion
12:20 -	13:45	Lunch Break
13:45 -	14:15	Invited Talk: D. Koller: Learning Probabilistic Relational Models
14:15 -	14:35	L. Getoor, D. Koller, N. Friedman From Instances to Classes in Probabilistic Relational Models
14:35 -	15:05	Invited Talk: R. Greiner: Exploiting Common Relations: Learning One Belief Net for Many Classification Tasks
15:05 -	15:25	J. Cussens Attribute-Value and Relational Learning: A Statistical Viewpoint
15:25 -	15:40	Discussion
15:40	16:10	Coffee Break
16:10	16:40	Invited Talk: A. Srinivasan: Heretical ILP?
16:40 -	17:00	F. Costa, P. Frasconi, V. Lombardo, G. Soda Learning to Rank Structured Alternatives: An Application to Incremental Processing of Natural Language
17:00 -	17:20	Y. Chevaleyre, J.-D. Zucker Noise-Tolerant Rule Induction from Multi-Instance Data
17:20 -	17:40	J. Ramon, L. De Raedt Multi-Instance Neural Networks
17:40 -	18:00	B. Hammer Neural Networks Classifying Symbolic Data
18:00 -	18:15	Discussion
18:15 -	18:30	General Discussion and Closing Remarks

Chairs

Luc De Raedt Machine Learning and Natural Language Processing Lab  Institut für Informatik  Albert-Ludwigs-Universität Freiburg  Am Flughafen 17,  Gebäude 079  D-79110 Freiburg i. Br., Germany E-mail: deraedt@informatik.uni-freiburg.de

Stefan Kramer Machine Learning and Natural Language Processing Lab  Institut für Informatik  Albert-Ludwigs-Universität Freiburg  Am Flughafen 17,  Gebäude 079  D-79110 Freiburg i. Br., Germany E-mail: skramer@informatik.uni-freiburg.de

Organizing Committee

• Luc Dehaspe (Katholieke Universiteit Leuven, Belgium)
• Saso Dzeroski (Jozef Stefan Institute, Ljubljana, Slovenia)
• Roni Khardon (Univ. of Edinburgh, UK)
• Bernhard Pfahringer (University of Waikato, New Zealand)
• Lorenza Saitta (Universita di Torino, Italy)
• Michele Sebag (Ecole Polytechnique, Palaiseau, France)
• Ashwin Srinivasan (University of Oxford, UK)
• Hannu Toivonen (Nokia Research Center, Finland)
• Mark Craven (University of Wisconsin, USA)
• Oded Maron (PHZ Capital Partners, USA)
• Daphne Koller (Stanford University, USA)
• Russ Greiner (University of Alberta, Canada)

Related Events

• Tenth International Conference on Inductive Logic Programming (ILP2000), July 24-28, 2000, London.
• AAAI 2000 Workshop on Learning Statistical Models from Relational Data, July 31, 2000; Austin, Texas.