Doug Fisher
Special Issue Editor
http://cswww.vuse.vanderbilt.edu/~dfisher/mlj-unsup.html
Several forms of unsupervised learning extract
relationships from data that can be then exploited for
inference. The primary unsupervised techniques include
clustering, learning (usually Bayesian) belief networks,
and learning association rules. The unsupervised "pattern"
or "concept" learning methods that are of most interest in
this special issue differ from supervised concept learning
methods in that there is no single, dependent variable,
dimension, or predicate that is the *a priori* focus of
inference. Rather, an unsupervised method may support
inference along more than one dimension (variable, property),
typically many dimensions/properties.
Authors are encouraged to submit papers in the primary
unsupervised learning paradigms of clustering, belief-
network learning, and association-rule learning (and
possibly others) for consideration as contributions to
the Special Issue on Unsupervised Learning of the
journal, Machine Learning. Articles that relate
different paradigms are especially welcome.
Review Criteria
Each submission will be reviewed by two to three
reviewers, as well as the Special Issue Editor. Each
submission should clearly describe an unsupervised
learning algorithm from one of the major
unsupervised paradigms mentioned above, or if the
learning approach is from some novel paradigm, then
it should explain the relationships to one or more
of these established paradigms. In addition, the
successful submission will describe an inference
procedure or other performance task that operates
on learned knowledge. Inference of unobserved variable
values and/or joint probability distributions are
examples of inference tasks.
Experimental and/or formal analysis should support
the merits of the unsupervised method's ability
to improve the performance task(s) along dimensions
such as accuracy and cost, though application-oriented
contributions (see below) may relax the formality of
experimental and theoretical demonstrations if
practical gains are documented in the submission.
The requirement for well-defined performance and
learning procedures does not preclude the possibility
of interactive approaches, in which human and machine
share learning and inference responsibilities.
Cognitive modeling submissions are also possible,
where fits to human data are desired.
In all cases, submissions should be scholarly, with
thorough links to related research, and report
methods/results previously unreported in the archival
(typically journal, book) literature. However, exceptional
surveys, particularly those that cross paradigm boundaries,
may be considered at the discretion of the Special
Issue Editor and the Executive Editor.
Except in extraordinary circumstances, submissions should
not exceed 25 pages (journal-formatted pages), excluding
references.
Research areas
Submissions should report novel methods/results, including
along any of the following dimensions.
pattern and concept representations and presentations,
including visualization strategies
data representation (e.g., relational data, continuous data,
hierarchical data)
search strategies (e.g., anytime algorithms, incremental
algorithms, optimization strategies,
human-machine interactive approaches, parallelism)
performance tasks, including complex forms of problem
solving
analytical and analytical/inductive hybrid approaches
cognitive modeling
Application areas
Submissions may also contribute primarily in terms of
application (e.g., data mining in a particular area).
Appropriate application-oriented papers will report
well-defined, generic learning and performance procedures,
which need not be novel relative to the literature, but
such papers are expected to describe successful
applications of unsupervised methods for a particular
application task, and provide informed recommendations
on the class of applications for which the reported
methods appear appropriate. Careful thought
should be given to whether an application-oriented
paper can be best treated as a technical note submission
(i.e., 10 or fewer journal-formatted pages), by
appropriately relying on previous research publications.
Of course, submissions with both strong research and application
contributions are welcome.
Important Dates
Jan. 8, 1999 Title, authors, and abstract should be sent
electronically to
mlj-unsup@vuse.vanderbilt.edu
with an intent-to-submit cover page. This is an
important deadline that will facilitate reviewing,
but minor changes to title, authors, and abstract
are acceptable prior to the submission of full
papers. Submissions of title/authors/abstract
should be in plain text.
Feb. 8, 1999 Full submissions should be received by
mlj-unsup@vuse.vanderbilt.edu
Full submissions should be in postscript
format, and may be COMPRESSed if larger
than 1M. Other formats may be accepted upon
request. Submissions must be in English.
April 30, 1999 Decisions sent to authors; papers accepted
with no more than minor revisions accepted to
the special issue.
June 14, 1999 Final versions of accepted papers should be received
by the Special Issue editor in the format specified
for full submissions, using Kluwer style guidelines.