In this issue:
Peter Bryant
College of Business
University of Colorado at Denver
Denver, CO 80217-3364
pbryant@castle.cudenver.edu
303-556-5833
The Amherst meetings of CSNA and the Numerical Taxonomy group are just finished. In my view we had a successful conference, and many thanks are due to the organizers Mel Janowitz and Pierre Legendre. I trust you will find a summary of the program elsewhere in the newsletter. That should give you the flavor of what we discussed. We've received a number of favorable comments from folks new to CSNA meetings. The informal nature of the meetings, relative lack of parallel sessions, chance to meet folks in other disciplines with shared interests, and the chance to discuss work in progress seem to be what they find appealing.
If those (or other) features of CSNA meetings appeal to you, why not make it a point to mention the next CSNA meetings to a colleague sometime, and urge them to consider coming and/or submitting a paper. The 1997 meeting is planned for June 12-14, 1997 at American University in Washington, D. C., while the 1998 meeting will be held jointly with the Psychometric Society in mid- June, 1998 at the University of Illinois at Champaign/Urbana. More information on these meetings will appear in subsequent newsletters.
Finally, let us know how we could meet your professional needs better. We continue to experiment with electronic delivery of our services, and we want your feedback. We also are beginning to look at some of our administrative arrangements (regular versus affiliate memberships, for example), and would welcome any ideas or suggestions of ways to improve.
Dawn Iacobucci
Department of Marketing
Kellogg Graduate School of Management
Northwestern University
2001 Sheridan Road
Evanston, IL 60208
On behalf of the Classification Society of North America, we are delighted to accept into membership the following new members: Ada Alpert, Zhenmin Chen, Curt Stenger, Katherine Faust, James Hershey, Hohn MacCuish, Stephen G. Sireci, Thomas J. Smith, Pierre Chandon, Lars Bergman, Tammo H.A. Bijmolt, Andre Hardy, Martijn Berger, Tom T.K. Chau, D.P. Eveleigh, D.J. Kiewiet, Huib de Ridder, T.J. Euverman, Lynette Hunt, Janos Podani, Dr. Ing. Juergen Formella, Frenkel ter Hofstede, Willem J. Heiser, Greet Looseveldt, Bassam Michel El-Khouri, J. Padmore, Yutaka Tanaka, Kiyoshi Tanaka, and Magda VuylstedeWauters.
F.R. McMorris
Department of Mathematics
University of Louisville
Louisville, KY 40292
frmcmo01@homer.louisville.edu
(502)852-6826
Special things in this issue include a provocative and interesting Forum column by David Banks, and a student's view of CSNA by Dave Dubin, a student at Pitt.
Due to the length of this issue, I have not included the Bookshelf or Random Conference announcements. These should resume with the September Newsletter.
WORKING WITHOUT A NET
David Banks, Dept. of Statistics
Carnegie Mellon University, Pittsburgh, PA 15123
banks@stat.cmu.edu
People, even smart people, can be seduced by the semblance of elegant universality. This is a commonplace in politics, where theories of Marxism, Nazism, and Jeffersonian democracy have paralyzed the critical faculties of otherwise brilliant minds. It also happens in science; there are biologists whose mental evolution stopped with the theory of natural selection, physicists who lazily lean upon Hamilton's principle of least action, and psychologists whose whole Weltanschauung derives from Skinner. In statistics, there are some who think that information theory is the Philosopher's Stone, while others satisfy their craving for dogmatism with hits of Bayesianity. Computer science has recently developed a new drug, called neural nets.
Neural nets are a compelling idea. The strategy is to mimic the way human brains are organized, which both flatters scientists and (in principle) avails them of the millions of years of genetic experience. There is a great literature on the history, methods, and philosophy of neural nets; rather than recapitulate what has been widely disseminated already, I refer readers to Cheng and Titterington (1994) and Ripley (1994).
I come to bury, not to praise. There are already a number of of standard criticisms of neural nets. These include the difficulty of interpreting their results in scientifically informative ways, the danger of overtraining (a problem long known to statisticians as ``overfit''), and the improbability of getting sensible answers when a new case arrives that has covariates just a little different from any seen before. I don't mean to minimize these objections, but prefer to emphasize other kinds of problems that arise.
Neural nets are often applied in classification, and sometimes in nonparametric regression. These are dual problems, and methods devised for one can be applied to the other; thus the natural competitors to neural nets are computer-intensive regression software based on a host of new strategies. Hastie and Tibshirani (1990) survey many of these (which generally come from California and are referred to acronymically), such as LOESS, ACE, AVAS, RPR (a generic formulation of the commercial CART package), and PPR. The most hyped of these new methods, MARS, appeared later; see Friedman (1991). Recently, statisticians and computer scientists have begun to gain some experience with the comparative performance of these techniques.
At first, neural nets seemed to enjoy several advantages. It was
exotic, Kolmogorov could be cited as a grandfather of the idea,
and CS geeks used it. Most importantly, it was stupid-friendly;
people who were ignorant of both statistics and the applications
domain could churn their computers to get a solution whose
complexity defied criticism. Then Barron (1993) appeared to nail
its superiority down, by proving that neural nets evaded the curse
of dimensionality, a phrase coined by Richard Bellman to describe
the superexponentially increasing difficulty of inference with
increasing dimensionality. (Three interestingly different
statements of the curse are:
1) In high dimensions, all data are sparse.
2) The number of possible models increases combinatorially fast
with dimension.
3) In high dimensions, nearly all data show multicollinearity (or
its nonparametric generalization, concurvity).)
However, Barron's result is asymptotic, and seems to have little force in practice. Also, the functions to which Barron's result applies are less general than one would hope---they don't include hyperflats, but become smoother as dimensionality increases. And it turns out that PPR also evades the curse (Zhao and Atkeson, 1992), and probably MARS does also.
The ultimate arbiter among these many competing methods must be performance. Ripley (1993, 1994) describes a number of experiments comparing methods for classification problems, and finds no marked superiority for neural net techniques---in fact, they often are bad,and confound this deficiency by being the most computationally extravagant of the new procedures. And I've been working on a research project that compares ten new-wave nonparametric regression methods, including neural nets, in a designed experiment, after the inimitable style of Glen Milligan. The experiment focuses on regression, rather than classification, and the factors include the dimensionality, the proportion of truly explanatory variables, and six target functions: a constant, a linear relation, a standard Gaussian, a correlated Gaussian, a mixture of Gaussians, and the product function. No method dominates all the others, but there are clear determinations of which methods work well in which circumstances. Humble multiple regression rarely excels, but it is never a disaster; in contrast, neural nets works well once, and is otherwise deplorable.
The one case in which neural nets are competitive with statistical regression methods was for the correlated Gaussian (but I should hastily emphasize that there are many implementations of neural nets---I used Cascor (Fahlman and Lebiere, 1990), because it's been used in previous studies and automatically chooses the number of hidden nodes via cross-validation). In retrospect, this wasn't such a surprise. Neural nets should excel when the natural axes are less useful than pseudo-axes constructed as linear combinations of the original variables; this is the same case in which PPR should work well. Also, neural nets should beat PPR when the shape of the target function is well-modelled as the sum of a small number of sigmoidal functions, which is exactly the case for the correlated Gaussian (all correlations were .8, all variances were 1.0, the center was 0, and the domain of interest was the unit hypercube in p-space).
This single excellence may also account for why neural nets have not performed so badly in practice that accumulated experience would eventually override the charm of neural nets' semblance of universality. In many applications, it is easy to believe that the true response function looks very much like a correlated Gaussian, especially one that is truncated against the range of feasible operation. Such things occur commonly in industry---the product is lousy when the pressure or temperature are too low, but becomes rapidly better as one increases both, up to some maximum temperature and pressure that is determined by outside constraints. In such cases, it appears neural nets aren't bad (though other methods are comparable).
In closing, I have a word of practical advice for all practitioners who are trying to chose among these new-wave techniques. First, hold out a portion (say 10%) of the training data. Then use each method you're considering on the remaining 90%, and build a regression surface (or classification rule) with each. Then use the holdout data to calculate the mean squared error of prediction (or the misclassification rate). The method that performs best on the holdout data is the method you should use for the problem at hand. If you don't feel you have enough data on hand that you can spare 10%, then you oughtn't be using these methods.
References:
Barron, A.R. (1993). ``Universal Approximation Bounds for Superpositions of a Sigmoidal Function,'' _IEEE Transactions on Information Theory_, Vol. 39, 930-945.
Cheng, B. and Titterington, D.M. (1994). ``Neural Networks: A Review from a Statistical Perspective,'' _Statistical Science_, Vol. 9, 2-75 (with discussion).
Fahlman, S. E. and Lebiere, C. (1990). The Cascade-Correlation Learning Architecture. In Advance in Neural Information Processing Systems 2, ed. by D. Touretzky. Morgan Kaufman, San Mateo, CA, pp. 525-533.
Friedman, J.H. (1991). ``Multivariate additive regression splines,'' _Annals of Statistics_, Vol. 19, 1-66.
Hastie, T.J. and Tibshirani, R.J. (1990). _Generalized Additive Models_, Chapman and Hall, New York.
Ripley, B. D. (1993). Statistical aspects of neural networks. In _Networks and Chaos--- Statistical and Probabilistic Aspects_, O. E. Barndorff-Nielsen, J. L. Jensen, and W. S. Kendall, eds. Chapman and Hall, London, pp. 40-123.
Ripley, B. D. (1994). Neural networks and related methods for classification (with discussion). _Journal of the Royal Statistical Society, Series B, Vol. 56, 409-456.
Scott, D.W. and Wand, M.P. (1991). ``Feasibility of Multivariate Density Estimates,'' _Biometrika_, Vol. 78, 197-206.
Zhao, Y. and Atkeson, C. G. (1992). Some approximation properties of projection pursuit networks. In _Advances in Neural Information Processing Systems 4_, J. Moody, S. J. Hanson, and R. P. Lippmann, eds. Morgan Kaufmann, San Mateo, CA, pp. 936-943.
1996 Annual Meeting
Classification Society of North America
and the Numerical Taxonomy Group
June 13-16, 1996
University of Massachusetts
Amherst, MA 01003
SCIENTIFIC PROGRAM
Program Chair and local host: Melvin F. Janowitz, University of Massachusetts
Friday, June 14, 1996
KEYNOTE ADDRESS: Phillipa Pattison, University of Melbourne, ALGEBRAIC BASES FOR THE ANALYSIS OF BINARY DATA
INVITED TALK: Donald Foster, Vassar College, SHAKESPEARE'S WHO DONE IT: THE CASE OF "A FUNERAL ELEGY"
SESSION 1A: PATHFINDER NETWORKS AND PROXIMITY GRAPHS
Organizer and Chair: Don Dearholt, Mississippi State University
Donald Geman, Joseph Horowitz and Dechang Wang, University of
Massachusetts, STOCHASTIC MODELING OF MAGNETIC RESONANCE IMAGES
WITH APPLICATIONS TO TISSUE CLASSIFICATION
Date: Tue, 13 Aug 96 12:19:30 -0300
From: "F.R. McMorris"
Donald W. Dearholt, Mississippi State University, K-LOCAL IMAGE
GRAPHS AND OPTIMAL ROUTING IN DYNAMIC NETWORKS
Penny Rheingans, University of Mississippi, INTERACTIVE COMPUTER
GRAPHICS TECHNIQUES FOR THE EXPLORATION OF MULTIVARIATE DATA
SESSION 1B: CLUSTERING AND CLASSIFICATION
Brigitte Jaumard, Pierre Hansen and Nenad Mladenovic, GERAD,
Montreal, Quebec, Canada, MINIMUM SUM OF SQUARES CLUSTERING IN A
LOW DIMENSIONAL SPACE
Pierre Hansen, Nenad Mladenovic, GERAD, Montreal, Quebec, Canada
and Bernard van Cutsem, and Bernad Yeart, IMAG, Grenoble, France,
AN ORDINAL CLASSIFIABILITY TEST FOR SINGLE-LINKAGE CLUSTERING
Peter Bryant, University of Colorado, Denver, ON THE MINIMUM
DESCRIPTION LENGTH (MDL) PRINCIPLE FOR HIERARCHICAL
CLASSIFICATIONS
James F. Palmer, SUNY College of Environmental Science and
Forestry, PERCEPTUAL CLASSIFICATION OF LANDSCAPES
INVITED TALK: Donald Geman, University of Massachusetts, TREE
STRUCTURED SHAPE RECOGNITION
SESSION 2A: IMAGE ANALYSIS AND ESTIMATION
Charles A. Bouman, Jau-Yuen Chen, and Jan P. Allebach, Purdue
University, STOCHASTIC MODELS FOR FAST MULTISCALE IMAGE SEARCH
Basilis Gidas, Brown University, OBJECT RECOGNITION VIA
HIERARCHICAL/SYNTACTIC MODELS AND DYNAMIC PROGRAMMING
Alfred O. Hero III, University of Michigan, OPTIMAL DETECTION OF A
TARGET STRADDLING A LINEAR BOUNDARY IN CLUTTER
Sridhar Lakshmanan, University of Michigan at Dearborn ; Bruce
Hauss and Steve Hershkowitz, TRW Space and Electronics Group,
Redondo Beach, California, TARGET IDENTIFICATION IN LOW RESOLUTION
ISAR IMAGES USING DEFORMABLE TEMPLATES
Jayashree Subrahmonia, IBM TJ Watson Research Center; Zhibin Lei
and David B. Cooper, Brown University, USE OF SELF AND MUTUAL
INVARIANTS FOR RECOGNITION OF COMPLEX OBJECTS REPRESENTED AS THE
UNION OF IMPLICIT POLYNOMIAL PATCHES
SESSION 2B: Applied Statistics
Bernard Harris, University of Wisconsin, A COMPARISON OF VARIOUS
COMBINATORIAL TESTS FOR VERIFICATION OF CLUSTERING
Pascale Rousseau, University of Quebec at Montreal, COMPARISON OF
THREE ESTIMATORS OF LOGISTIC REGRESSION MODELS
William D. Shannon, Washington University School of Medicine and
David Banks, Carnegie Mellon University, A DISTANCE METRIC FOR
CLASSIFICATION TREES
Stanley L. Sclove, University of Illinois at Chicago and Simon A.
Sherman, University of Nebraska Medical Center, FINITE-MIXTURE
MODELING OF NUMERICAL DATABASES, WITH APPLICATION TO DETERMINATION
OF PROTEIN STRUCTURE
David Banks, Carnegie Mellon University, FROM CLUSTER ANALYSIS TO
NONPARAMETRIC REGRESSION
Michael Sutherland, University of Massachusetts, ALL ABOUT BOUTS
Banquet Speaker: Herman Friedman, Fordham University,
CLASSIFICATION AND CLUSTERING: A PERSPECTIVE ON APPLICATIONS
Saturday June 15, 1996
INVITED ADDRESS: Bruno Leclerc, Centre d'Analyse et de
Mathematique Sociales, Ecole des Hautes Etudes en Sciences
Sociales, A SURVEY ON THE CONSENSUS OF CLASSIFICATION TREES
SESSION 3A: Information Retrieval
Michael J. Kurtz, Harvard-Smithsonian Center for Astrophysics,
HISTORY AND STATUS OF THE NASA ASTROPHYSICS DATA SYSTEM ABSTRACT
SERVICE
James Allan, University of Massachusetts, CLASSIFYING TEXTS USING
RELEVANCE FEEDBACK
David Dubin, University of Pittsburgh, CLUSTERING TENDENCY AND THE
CLUSTER HYPOTHESIS IN INFORMATION RETRIEVAL
SESSION 3B: Statistical Applications
Melvin Prince, Marist College, A TAXONOMY OF COMPARATIVE CONSUMER
EXPENDITURES
Michel Goulet, La Confederation de caisses populaires et
d'economie Desjardins du Quebec and David Wishart, Clustan, Ltd.,
Edinburgh, Scotland, CLASSIFYING A BANK'S CUSTOMERS FOR IMPROVING
THEIR FINANCIAL SERVICES
Robert Sokal, State University of New York; Neal L. Oden, The
EMMES Corporation, Potomac, MD;Jeff Walker, State University of
New York and Diane M. Waddle, Duke University Medical Center,
USING DISTANCE MATRICES TO CHOOSE BETWEEN COMPETING THEORIES AND
AN APPLICATION TO THE ORIGIN OF MODERN HUMANS
SESSION 4A: Classification in Social Network Analysis
Martina Morris, Pennsylvania State University; Mark Handcock, New
York University; Maria Wawer, Columbia University; and Ron Gray,
Johns Hopkins University; USING MOBILITY NETWORKS TO ESTABLISH
TREATMENT AND CONTROL CLUSTERS IN A RANDOMIZED COMMUNITY TRIAL
Katherine Faust, University of South Carolina, ACTOR AND EVENT
CENTRALITIES FOR TWO-MODE NON-DYADIC NETWORKS
L. Koehly, University of Texas; N. Contractor, M. Heald and S.
Wasserman, University of Illinois, PERCEPTUAL CONGRUENCE AND
INCONGRUENCE AMONG NETWORK MEMBERS IN COMMON CLUSTERS
Ece Kumbasar, University of California, Irvine, MODELS FOR
COGNITIVE REPRESENTATIONS OF NETWORK STRUCTURES
W.H. Batchelder, University of California, Irvine, STATISTICAL
MODELS FOR CONSENSUS AGGREGATION OF DIGRAPH SOCIAL NETWORK DATA
Invited address: David Swofford, Smithsonian Institution,
ADVANCES IN METHODOLOGY FOR RECONSTRUCTING EVOLUTIONARY TREES:
BRINGING THE TECHNOLOGY BACK TO THE BIOLOGISTS
SESSION 5A: Consensus Theory
William H.E. Day, Memorial University of Newfoundland, A NEW
CONSENSUS RULE FOR COMMITTEE ELECTIONS
B. Monjardet, University of Paris I and CAMS, ON THE COMPARISON OF
THE BORDA-SPEARMAN AND CONDORCET-KENDALL APPROACHES FOR
CONCORDANCE AND CONSENSUS OF PREFERENCES
Gary D. Crown, Wichita State University; Robert C. Powers,
University of Louisville; and Melvin F. Janowitz, University of
Massachusetts, REPRESENTATION FAMILIES AND NEUTRAL CONSENSUS
FUNCTIONS
Jean-Pierre Barthelemy, Ecole Nationale Superieure de
Telecommunications Bretagne, Brest, France, STABILITY CONDITIONS
FOR NON-NUMERICAL HIERARCHICAL CLUSTERING
R.C. Powers, University of Louisville, MEDIANS AND MAJORITIES IN
MULTILATTICES
SESSION 5B: Applications
Gregory W. Cermak, GTE Laboratories, Waltham, MA, CONSUMER
PREFERENCES FOR TYPE OF INFORMATION
Kathleen M. Sheehan, Educational Testing Service, Princeton, NJ, A
TREE-BASED APPROACH TO PROFICIENCY SCALING
N. Sriram, National University of Singapore, EXPERT, INTERMEDIATE
AND NOVICE CLASSIFICATIONS OF PSYCHOLOGY AND PHYSICS CONCEPTS
Stephen G. Sireci and Frederic Robin, University of
Massachusetts, SETTING PASSING SCORES ON TESTS USING CLUSTER
ANALYSIS
Sunday, June 16, 1996
NUMERICAL TAXONOMY (NT) CONFERENCE
Opening Remarks: Joseph Kunkel, University of Massachusetts
NT Invited Address, David L. Swofford, Smithsonian Institution,
ADVANCES IN METHODOLOGY FOR RECONSTRUCTING EVOLUTIONARY TREES:
BRINGING THE TECHNOLOGY BACK TO THE BIOLOGISTS
NT SESSION: Session chair: William H. E. Day, Memorial
University of Newfoundland
NT invited talk: F. James Rohlf, State University of New York,
Stony Brook, APPLICATION OF GEOMETRIC MORPHOMETRIC METHODS TO
EVOLUTIONARY STUDIES
Pierre Legendre, University of Montreal, THE BIOLOGICAL MEANING OF
DIFFERENT TYPES OF PERMUTATIONS
Mariana Kant, University of Moncton and Francois Lapointe,
University of Montreal, PREPROCESSING INFORMATION FOR THE TREE
COMPATIBILITY PROBLEM
R.E. Strauss, Texas Tech University and E. Dyerson, University of
Arizona, USE OF CONTINUOUS-CHARACTER MAPPINGS ONTO PHYLOGENETIC
TREES FOR COMPARATIVE STUDIES
R.E. Strauss and M.A. Houck, Texas Tech University, IDENTIFICATION
OF AFRICANIZED HONEYBEES VIA NONLINEAR MULTILAYER PERCEPTRONS
Dave Dubin
After this year's CSNA meeting in Amherst, I prepared a brief
report on the meeting for my fellow students at the University of
Pittsburgh's Department of Information Science. Later I was asked
whether I would mind having the report reprinted in the CSNA
newsletter.
My goal was not only to report papers and presentations of
interest to information scientists, but to urge my colleagues to
join the society. It seems peculiar to me to make the same pitch
to people who are members already (and who, for the most part,
aren't students). But I'm happy to summarize my opinions of what
makes CSNA a "student-friendly" association.
1) CSNA is an interdisciplinary association. Researchers in a
variety of subject disciplines gather to discuss methods and
approaches that they describe using different vocabularies in
their own journals and conference proceedings. The CSNA meeting
is a great place to find out if one of your methodological
problems has already been solved in a discipline you wouldn't
ordinarily have thought to read.
2) Not only are a variety of subject disciplines represented, but
a variety of specific tools as well. The CSNA meeting gives you a
chance to compare studies using neural networks, machine learning,
visualization, and traditional multivariate methods. You don't
find as many opportunities for that kind of contrast at a
connectionism conference or a visualization conference.
3) All three of the CSNA meetings I've attended have been run with
sensitivity to student interests and budget constraints. For
example, the Houston meeting included special student paper
session. The short course on introduction to clustering
complemented my course work very well. At the Pittsburgh and
Amherst meetings dormitory rooms were available as alternatives to
the hotel. The Houston meeting took place in a neighborhood of
several hotels, including some inexpensive alternatives to the one
chosen for the meeting. At past meetings, students have
volunteered to assist in return for a reduced registration fee.
4) I have found CSNA members to be very supportive of students and
newcomers to the field. My impression is that most of you would
rather discuss methodology objections over beer than to flame a
person publicly. Although the meetings were small, I had good
attendance and challenging questions at both of my presentations
(even when I was scheduled opposite Professor Sokal).
These are aspects of CSNA that I hope will not change in the
future,even though I won't be a student for very long. Please
feel free to share my observations with your own students.
The 1996 general business meeting of the Classification Society of
North America was called to order at 5:30 PM, June 14, 1996, in
Amherst, MA, in conjunction with the society's annual meeting.
Peter Bryant reported to the meeting on the following topics:
(1) CSNA 97 is tentatively set for June 12-14, 1997 in Washington,
D.C. at American University, hosted by Prof. Olga Cordero-Brana.
(2) CSNA 98 is scheduled for June 17-21, 1998 in Champaign/Urbana,
Illinois at the University of Illinois. The meeting will be held
jointly with the Psychometric Society, and will be hosted by Prof.
Stanley Wasserman.
(3) CSNA 96 appears to be a great success, thanks in large part to
our host Prof. Mel Janowitz.
(4) Summary of the CSNA Board Meeting held June 13, 1996 was presented:
There was general discussion of the issue of getting access to
publications of IFCS societies. Phipps Arabie pointed
out that GfKl proceedings volumes from Springer-Verlag were hard
to obtain in the US (and expensive). William H.E. Day suggested
that addressing these issues at the IFCS level might be more
productive, but J.D. Carroll reminded the group that current
discussions about the role of our Journal as a possible IFCS
publication seems to have stagnated.
The Newsletter and Service are now distributed in electronic
forms. A proposal to modify the by-laws to permit electronic
voting will be voted on this fall.
Membership issues were discussed.
Phipps Arabie reported on recent issues related to the increase in
page count for the Journal. Galleys for the 2nd 1996 issue have
now been sent out, and the backlog is being reduced.
IFCS-98 will be in Rome in July 1998.
A general discussion of the CLASS-L list server reflected a sense
to continue it as long as possible.
The meeting was adjourned at 6:05 PM
-- S. Hirtle
Chair: Stephen Sireci, University of Massachusetts
Organized by Sridhar Lakshaman, University of Michigan at Dearborn
and A. K. Jain, Michigan State University. Session chair: Sridhar
Lakshaman
Session Chair: Art Kendall, US Government
Organized by Stephen Hirtle and David Dubin, University of
Pittsburgh
Session Chair: Gary D. Crown, Wichita State University
Organized by Stan Wasserman, University of Illinois
Organized by Robert C. Powers, University of Louisville
Session chair: John Daws, New York University
Organized by Melvin Janowitz, University of Massachusetts and
Pierre Legendre, University of Montreal
::::::::
A Student Perspective on CSNA Meetings
::::::::
Dept. Information Science and Telecommunications
University of Pittsburgh
::::::::
1996 General Business Meeting
::::::::
(a) The nominating committee is at work and will have a proposed
slate of nominees later this summer.
(b) The finance committee recommends no dues increases this year.
The recent drop in membership raises concerns, though. The
committee questions the need for our current Affiliate
memberships, now that the newsletter and service are distributed
electronically, but thinks this should be reviewed and does not
recommend any immediate change.
(c) The voting procedures of the society will return to the
simple Hare system, as opposed to the "portfolio" Hare system used
for the last year.
(d) The NT group is currently debating its own activities and
possible relationships with CSNA.
(e) There have been some communication difficulties with the GfKl
concerning their meetings. The president is to see what can be
worked out.
The WWW version of the CSNA Newsletter is made available as a
service of the
Classification Society of North America.
For further information on becoming a member of CSNA, please
contact the hirtle+@pitt.edu, CSNA
Webmaster.