Learning With Multiple Views
Bonn, Germany, August 11th, 2005
Contact address:
multiview@ls8.cs.uni-dortmund.de
Abstract
Multi-view learning is a natural, yet non-standard new problem
setting; it describes the problem of learning from data
represented by multiple independent sets of features.
A typical example is learning to classify web pages
by either the words on the page or the words
contained in anchor texts of links to the page.
Multi-view learning methods have been studied by
Yarowsky (1995)
and
Blum and Mitchell (1998),
who noticed that having multiple representations
can improve classification performance
when in addition to labeled examples, many unlabeled examples are
available.
A recent result by
Abney (2002)
suggests that there
may be an underlying principle which gives rise to a family of new
methods: The disagreement rate of two independent hypotheses
upper-bounds the error rate of either hypothesis. By minimizing the
disagreement rate on unlabeled data, the error rate can be minimized.
In the last 2-3 years, several new supervised and unsupervised methods have
been proposed which
appear to utilize this consensus maximization principle in one way or
another. However, in many cases the contributors are not to the full
extent aware of the relationships between their methods and a possible
common underlying principle.
The workshop aims at bringing together researchers who are working on
learning problems with multiply represented instances and consensus
maximizing learning methods; our goals are to make the intrinsic
structure of this field more clearly visible and to bring
this interesting and rapidly developing area to the attention of
additional researchers.
Topics of the workshop include
- analysis of algorithms: co-training, co-EM, co-EMT, ...
- active and semi-supervised learning
- multi-view clustering and classification
- novel learning tasks (interpretability, constraints, ...)
- independence of views: quantification and relevance
- text, web, and other applications
- hierarchical, partitioning, spatial, spectral clustering
- theoretical analyses
- relation to other fields of learning (e.g., boosting)
- consensus maximization principle
- generative and discriminative models
Schedule
Please note that this schedule is preliminary. In particular, starting times of
the sessions may change.
| 9:00 - 9:25 | A Co-Regularization Approach to Semi-supervised
learning with Multiple Views | V. Sindwhani et al. |
| 9:25 - 9:50 | Analytical Kernel Matrix Completion with Incomplete Multi-View
Data | D. Williams and L. Carin |
| 9:50 - 10:15 | Active Learning of Features and Labels | B. Krishnapuram et al. |
| 10:15 - 10:40 | Spectral Clustering with Two Views | V. de Sa |
| | Coffee Break | |
| 11:00 - 11:40 | Invited Talk: Comparability and
Semantics in Mining Multi-Represented Objects | M. Schubert |
| 11:40 - 12:05 | Hybrid Hierachical Clustering: Forming a Tree From Multiple View
s | A. Gupta and S. Dasgupta |
| 12:05 - 12:30 | Estimation of Mixture Models using
Co-EM | S. Bickel and T. Scheffer |
| | Lunch Break | |
| 14:00 - 14:40 | Invited Talk: | R. Ghani |
| 14:40 - 15:05 | Using Unlabeled Texts for Named-Entity
Recognition | M. Rössler and K. Morik |
| 15:05 - 15:30 | Optimising Selective Sampling for Bootstrapping
Named Entity Recognition | M. Becker et al. |
| | Coffee Break | |
| 16:00 - 16:25 | The use of machine translation tools for
cross-lingual text mining | B. Fortuna and J. Shawe-Taylor |
| 16:25 - 16:50 | Multiple Views in Ensembles of Nearest Neighbor
Classifiers | O. Okun and H. Priisalu |
| 16:50 - 17:15 | Interpreting Classifiers by Multiple
Views | S. Rüping |
| 17:15 - 17:30 | Final Discussion | |
The Proceedings are available online:
[PDF].
Program Committee
The program committee consists of:
- Steven Abney, University of Michigan
- Steffen Bickel, Humboldt University
- Ulf Brefeld, Humboldt University
- Sanjoy Dasgupta, University of California, San Diego
- Johannes Fürnkranz, Darmstadt University
- Rayid Ghani, Accenture
- Thomas Hofmann, Brown University
- Thorsten Joachims, Cornell University
- Kristian Kersting, Freiburg University
- Stan Matwin, University of Ottawa
- Tom Mitchell, Carnegie Mellon University
- Ion Muslea, SRI
- Bernhard Schölkopf, Max Planck Institute for Biological Cybernetics
Important Dates
| Apr 1, 2005 | Paper submission deadline |
| Apr 22, 2005 | Notification of acceptance |
| May 13, 2005 | Final paper deadline |
| Aug 11, 2005 | Workshop |
Downloads
- Proceedings: [PDF]
- Call for Papers: [PS] [PDF] [TXT]
- Workshop Proposal: [PS] [PDF] (the original
workshop proposal, contains some additional information)
