IEEE Transactions on Knowledge and Data Engineering
Special Issue on
"Building Probabilistic Networks: Where Do the Numbers Come From ?"
Submissions due on September 15th, 1998
----------------------------------------------------------------------------
Probabilistic networks are now well established as practical representations
of knowledge for reasoning under uncertainty, as the increasing number of
successful applications in such domains as diagnosis, planning, vision, and
natural language processing demonstrates. A probabilistic network (also
referred to as belief network or causal network) consists of a graphical part,
encoding a domain's variables and the qualitative dependences among them, and
a quantitative part, encoding probabilities over these variables. The task of
eliciting the graphical part from domain experts is comparable to knowledge
engineering for other AI representations and, although it may require
significant effort, is generally considered quite doable. The task of
obtaining the probabilities often appears more daunting to those embarking on
constructing a probabilistic network: "Where do the numbers come from?" is a
commonly asked question. This question is the central theme of the special
issue.
For building probabilistic networks for data-rich application domains, methods
are available for estimating the probabilities for the network from data.
When there is little or no data available, one might resort to methods
developed and used by decision analysts to elicit judgmental probabilities
from domain experts. However, application of these methods to the
construction of probabilistic networks is hampered by, among various obstacles,
the large size and complex dependence structure of probabilistic networks as
compared to most conventional decision analytic models. In fact, effective
use of probabilistic networks requires careful tradeoff between the desire for
a large and rich model to obtain accurate performance and the costs of
construction, inference, and maintenance. Researchers and practitioners in
the use of probabilistic networks have developed effective and practical
techniques to handle thise problem. Few of these techniques, however, have
become widely available general purpose tools and little is known as to their
compatibility.
The main objective of this issue is to identify methods of eliciting
probabilistic information that are general enough to be shared, combined, and
developed further. We invite papers from researchers with experience in
building probabilistic networks. We are seeking information about how they
obtained the structure and the probabilities of their networks, what problems
they encountered, how they solved them, what worked, what did not, and why.
We strongly encourage submission of papers that discuss practical experiences,
although we are open for important theoretical contributions that are
directly applicable in practice. Examples of technical topics that the issue
will focus on are: various sources of probabilistic information; granularity
of the model versus granularity of the domain knowledge; good and bad
experiences with collecting data; identification of domain characteristics
that allow for applying specific elicitation methods; elicitation of
qualitative knowledge and combining it with quantitative information; the
effects of rough versus fine-tuned probability assessments on performance;
the sensitivity of probabilistic reasoning to the precision of the
probabilities; bringing experiences with existing AI knowledge elicitation
techniques to building probabilistic networks; computer aids for probability
assessment; treatment of distributions that vary over time.
----------------------------------------------------------------------------
Submission guidelines
----------------------------------------------------------------------------
Papers should be limited to a total of 30 pages. Papers should be printed on
8.5" x 11" or A4 sized paper using 12 point type (10 characters per inch) with
a one inch margin and no more than 43 lines per page. The first page must
include the names and full postal and e-mail addresses of all authors, an
abstract (not more than 200 words), and a list of keywords reflecting the
topics addressed by the paper. Electronic submissions are strongly encouraged;
the preferred format is PostScript. If electronic submission is absolutely
impossible, please send six clearly legible hard copies of papers to the
primary contact mentioned below; FAX submissions are not acceptable. Each
submitted paper will be reviewed by at least three reviewers and will be
judged on significance, originality, and clarity. In selecting the papers, we
will also consider breadth of coverage of the main theme of the special issue.
----------------------------------------------------------------------------
Important dates
----------------------------------------------------------------------------
Submission deadline : September 15th, 1998
Notification of acceptance : December 15th, 1998
Final manuscripts due : March 1st, 1999
Special issue : Mid 1999
----------------------------------------------------------------------------
Guest editors
----------------------------------------------------------------------------
If you have any questions about the special issue, please contact
one of the guest editors:
Marek J. Druzdzel
University of Pittsburgh, School of Information Sciences,
135 North Bellefield Avenue, Pittsburgh, PA 15260, U.S.A.
e-mail: marek@sis.pitt.edu
Phone: +1-412-624-9432
Linda C. van der Gaag (primary contact)
Utrecht University, Department of Computer Science,
P.O. Box 80.089, 3508 TB Utrecht, The Netherlands
e-mail: linda@cs.ruu.nl
Phone: +31-30-2534113
============================================================================