--
Data Mining and Knowledge Discovery community,
focusing on the latest research and applications.
Contributions are most welcome and should be emailed,
with a DESCRIPTIVE subject line (and a URL, when available) to (kdd@gte.com).
E-mail add/delete requests to (kdd-request@gte.com).
Nuggets frequency is approximately weekly.
Back issues of Nuggets, a catalog of S*i*ftware (data mining tools),
and a wealth of other information on Data Mining and Knowledge Discovery
is available at Knowledge Discovery Mine site, URL http://info.gte.com/~kdd.
-- Gregory Piatetsky-Shapiro (moderator)
********************* Official disclaimer ***********************************
* All opinions expressed herein are those of the writers (or the moderator) *
* and not necessarily of their respective employers (or GTE Laboratories) *
*****************************************************************************
Correction: Ron Kohavi quote in Nuggets 96:21 should be:
We are always given the finite amounts of data ...
and rarely do we reach asymptopia.
(not asympotia, sorry for the typo.
Asymptopia is the mythical land, the data miners 'utopia', where
the amount of data is infinite and all algorithms converge and
all users are satisfied ... Naturally, asymptopia can be reached only
in the limit. GPS) Previous1NextTop
>~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
(from DBWORLD)
Date: Sat, 29 Jun 1996 10:07:19 -0500
From: brodie@gte.com
(Michael L. Brodie)
To: dbworld@ricotta.cs.wisc.edu
Subject: (DBWORLD) News From Your Friends At VLDB ...
Hello (DB) World, This Is VLDB Calling!
Golly there's been such a lot going on at VLDB central. First the headlines:
* VLDB Trustees Election: welcome 4 new trustees, thank 4 retiring trustees
* VLDB Home page: new URL, and the other guys!
* Information on new VLDB Programs: Library, Tutorial, ...
* VLDB '96 Proceedings may be on the WWW
* VLDB Journal, January '96 issue on the WWW, free
And now the details
* The VLDB Trustees Election concluded March 1996. The newly elected
trustees are (alphabetical order): Michael Carey (North America),
Masaru Kitsuregawa (Japan), Tamer Ozsu (North America), Timos Sellis
(Europe).
* The newly elected VLDB Endowment positions are:
Vice-President: Dr. Arie Shoshani
Treasurer: Dr. Stanley Su
(the president was not up for election this cycle and remains Dr. Peter
Lockemann
* The retiring trustees are:
Prof. Stuart Madnick, USA
Prof. Yahiko Kambayashi, Japan
Prof. Nick Roussopoulos, USA
Prof. Yannis Vassiliou, Greece
Prof Stanley Su and Dr. Arie Shoshani competed a 6 year term, but were
re-elected to Executive Committee positions, and thus their term was
extended to 4 more years.
* The VLDB Home page has the new URL is: http://www.vldb.org/vldb
www.vldb.com is being used by the organization that runs the 'VLDB
Summit' conference. This organization has nothing to do whatsoever with
VLDB, nor did they approach the VLDB Foundation for the use of the terms
that we have used for over 20 years.
* The VLDB Home page has been updated with information on
- new trustees (and how to contact all of them - go get 'em) - the entire
trustees emeritus
- the new VLDB Tutorials Program
- the existing VLDB Library Program
- connections to massive amounts of useful VLDB-related information
including the wonderful collection of WWW pages managed by Prof Michael
Ley, University of Trier, and mirrored at:
* VLDB Tutorials
The VLDB foundation would like to reach out to countries of the world which
may never see a VLDB conference happen, by having a tutorial series in such
a country. The proposals have to come from within such countries. Please
spread the word and have the interested organizers in these countries
contact Profs. Navathe and Dittrich.
* The VLDB '96 Conference, Bombay (now restored to its original
name of Mumbai), India, is looking into experimenting with
making the VLDB'96 proceedings available on the WWW. Stay tuned.
* As announced earlier this year, the January 1996 issue of the VLDB
Journal is currently available on the WWW for free. Find it from the VLDB
home page and follow the VLDB Journal links to Springer ...
__________________________________________________________________
Michael L. Brodie (brodie@gte.com)
GTE Laboratories Incorporated
Phone: (617)466-2256 40 Sylvan Road, MS-62
FAX: (617) 466-2439 Waltham, MA 02254 USA
Previous2NextTop
>~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Date: 13 May 1996 12:59:24 +0500
From: 'Robert Golan' (Robert_Golan@mail.tcpl.ca)
Subject: Financial Data Mining From CIFEr '96 conference
Optimization Trees and Data Mining
Robert L. Grossman
Magnify, Inc.
April 23, 1996
Classification and regression trees [Q86] and [B84] have
proved themseleves useful for a variety of data mining
problems. Our interest is to extend these basic techniques
to problems broadly related to optimization [G96a] and
[B96].
An important motivating example for us to assign a
statistically computed score to a collection of credit card
holders indicating the likelihood that the credit card
holder will make payments and then to set a credit limit for
each holder in such a way as to maximize the over-all
expected revenue from the entire collection. As the credit
line grows, so does the expected revenue from the revenue.
On the other hand, so does the risk of slow or late
payments. The goal is to set the control attribute, which
is the credit limit in this case, so as to maximize the
expected revenue. The role of data mining is to uncover
patterns in the data which lead to better optimization
algorithms or better means of estimating the statistical
attributes in the problem, such as the credit utilization.
The problem is difficult because of the number of
attributes, which may reach a thousand, and the amount of
data, which may reach a terabyte. For further details, see
[G96a]
More generally, we are interested in problems in which
there are objects containing 1) data attributes from which
are computed 2) summary attributes. From the data and
summary attributes we are interested in using data mining
techniques such as classification and regression trees and
related techniques to compute 3) statistical attributes.
Finally, we are interested in setting certain 4) control
attributes, which are functions of the statistical and
summary attributes, in order to optimize a given objective
function.
To attack problems like this, we have developed tree-based
optimization techniques [B96]. In particular, to achieve
the required scalablity we compute trees in a distributed
fashion using multiple nodes in a parallel computing
environment and then average the resulting trees.
Because of the amount of data involved we have also
developed specialized object warehouses which are
specifically designed to support data mining operations
[G96b] and specialized interfaces between object warehouses
and hierarchical storage systems [G96c].
As the amount of data grows, the number of patterns that can
be potentially discovered with data mining algorithms
combinatorially explodes. By focusing on patterns related
to the objective function associated with our optimization,
we reduce the number of patterns we must examine.
Bibliography
[B96] Haim Bodek, Robert L. Grossman and H. Vincent Poor,
``Data Mining and Tree-based Optimization,'' Magnify
Technical Report, Number 96-R7, March, 1996.
[B84] L. Breiman, J. H. Friedman, R. A. Olshen, and C.
J. Stone, Classification and Regression Trees, Wadsworth,
Belmont, California, 1984.
[Q86] J. R. Quinlan, ``The Induction of Decision Trees,''
Machine Learning, Volume 1, pp 81--106, 1986.
[G96a] R. L. Grossman and H. V. Poor,
Optimization Driven Data Mining and Credit Scoring,
in Proceedings of the IEEE/IAFE 1996 Conference
on Computational Intelligence for Financial Engineering
(CIFEr), IEEE, Piscataway, 1996, pages 104-110.
[G96b] R. L. Grossman, H. Bodek, and D. Northcutt,
``Early Experience with a System for Mining, Estimating, and
Optimizing Large Collections of Objects Managed Using an
Object Warehouse,'' Proceedings of the Workshop on Research
Issues on Data Mining and Knowledge Discovery, Montreal,
Canada, June 2, 1996, to appear.
[G96c] R. L. Grossman, D. Northcutt, and R. Swaim,
``Interfacing Object Warehouses to Hierarchical Storage
Systems,'' in Proceedings of the Goddard Conference on
Mass Storage Systems and Technologies, September 17--19,
1996, to appear.
--- Message Part 1.1: Text Body Part ------------------------------
ROBERT -- ATTACHED IS A MICROSOFT WORD FILE FOR THE ONE PAGE PAPER YOU
REQUESTED. THE TEXT IS BELOW. ONE OR THE OTHER SHOULD WORK FOR YOU. =
BEST
WISHES
JIM HALL
Applying Computational Intelligence to the Investment Process
James Hall, Ganesh Mani and Dean Barr
LBS Capital Management
311 Park Place Blvd., Suite 330
Clearwater, Florida 34619
For many large institutions, pension funds or trusts represent a major
portion of their tangible assets. Prudent investment of these funds -- so
that they will grow at a reasonable rate without undo risk -- is critical
to the long-term success of the institution. Most fund trustees rely on
professional money managers to help with these investment decisions.
Money managers make investment decisions based upon their learned or
acquired experience of correct decision making. It is a form of pattern
recognition: when event A occurred in the past, it was generally followed
by event B. However, the financial world contains huge quantities of =
data,
much of which is corrupted by noise and heavily influenced by news. For
example a decrease in jobless claims may either boost the stock market, or
cause a quick decline, depending on the prevailing view on which direction
the economy is heading (which view in itself may be hotly contested).
These complex, highly non-linear relationships between cause and effect
based upon huge quantities of noisy data are what allow money managers to
proliferate. The situation also seem tailor made for the application of
computational intelligence.
LBS Capital Management uses expert systems, neural nets and genetic
algorithms to manage portfolios totalling $600 million. LBS makes the =
same
assumption that other active money managers make: that there are windows =
of
persistent predictability in the financial market. Identifying both the
cause-effect relationships and the times of predictability in which they
can reliably be applied allows the investment manager to outperform the
broad market. These core of these predictions is based upon a neural
network system called AXON.
AXON is a collection of over 3000 neural networks, along with a mechanism
for constructing portfolios. It operates within a universe of the 500 S&P
stocks, the 400 S&P midcap stocks and 2000 other stocks which meet certain
market capitalization and liquidity requirements. The system is based on
the premise that each stock has its own unique performance footprint. It
also assumes that a number of cross-sectional factors influence stock
prices.
AXON inputs several data items to each stock-specific neural network.
These include price-derived and volume-derived information; value and
growth factors based on fundamental data such as earnings estimates,
cashflow growth, and book value; and other quantitative factors.
Significant non-linear relationships often appear in the data, making the
output variable, alpha (excess return over risk-adjusted index return), to
some extent predictable.
The AXON models are tested by holding out about 25-30% of the data as test
and validation sets. We are able to attain correlation coefficients of =
0.6
to 0.95 on these out-of-sample tests for a majority of the stocks being
modeled. Overfitting is a major concern in this application, and it is
important to perform careful out-of-sample validation of the trained
networks. Of course, we also come across stocks that are hard to model
using the inputs and methodology currently employed by AXON. Of the 2500
stocks analyzed each week, AXON typically generates 100-150 manes on the
long side and about an equal number on the short side. Based on the
predictions made by each individual neural net model, we trade if a
significant upside or downside projection is made. The actual
implementation of these names -- which involves deciding what percentage =
of
each stock to hold in a portfolio - is done by a portfolio optimization
process.
There is a growing body of empirical evidence that neural nets are good
forecasting tools for this domain. Our experience and track record with
actual money management adds to this evidence -- since its implementaion =
in
1993, the strategy has outperformed the broad stock market and added
significant value to its investors.
> [ GPS: the reference to this work is:
> Hall, J., Mani, G., and Barr, D. 1996.
> Applying Computational Intelligence to the Investment Process.
> In {em Proc. of CIFER-96: Computational Intelligence in Financial
> Engineering}. Piscataway, NJ: IEEE Press. ]
We are working on incremental algorithms for mining association rules
on dynamic databases. We currently have designed a basic algorithm.
To justify its very existence, we need to find GOOD applications for
it. Is there anyone out there who could help us with some GOOD problems
where any non-incremental algorithms could not at all satisfy the
requirements?
Please send your comments to hongbo@buck.ac.uk
or ubacr62@dcs.bbk.ac.uk.
Hongbo
Previous4NextTop
>~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
From: askirai@netcom.com
(Iris Askira)
Subject: Comprehensibility and novelty in KDD research
To: kdd@gte.com
Date: Fri, 5 Jul 1996 11:45:06 -0700 (PDT)
Reply-To: ASKIRAI@cgs.edu
Dear Gregory,
I enclose a few questions which came-up while working on a
paper for a PhD seminar. I would be very grateful if you could distribute
them among the KDD community:
1. In looking for better understanding of the relationship between the research
objective of comprehensibility and that of novelty, I am in difficulty with
regard to references to novelty. Any such references would be
much appreciated.
2. I am also looking for literature on the integration of domain knowledge
into the algorithm findings, where comprehensibility is of interest.
3. Is there any research that considers the possibility of a superiority of
some rule induction algorithms over others in terms of better semantic
comprehensibility ( e.g. the results are more consistent with the
existing domain theory for a given problem domain ) ?
4. Is there a consensus as to what rule induction algorithms are ? How
would rule induction algorithms be defined ? ( e.g. is there any
limitation on the expressiveness of the representation formalism ? )
Best regards,
Irit Askira -- Claremont Graduate School Previous5NextTop
>~~~Publications:~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
From: Jiawei Han (han@cs.sfu.ca)
Date: Wed, 10 Jul 1996 07:51:08 -0700 (PDT)
Subject: CFP: Special Issue on Data Mining -- Journal of Intelligent Information Systems (JIIS)
---------------
Call For Papers
---------------
-------------------------------------------------
Journal of Intelligent Information Systems (JIIS)
-------------------------------------------------
----------------------------
Special Issue on Data Mining
----------------------------
As a young, promising research area with broad applications, data mining
and knowledge discovery in databases has attracted great interest in the
research communities of database systems, machine learning, statistics,
high performance computing, information retrieval, data visualization,
and many others. As an example, ACM-SIGMOD'96 Workshop on Research Issues
on Data Mining and Knowledge Discovery (DMKD'96) held in Montreal, Canada
(June 2, 1996) attracted over 100 attendees. Also, KDD'96 conference, to be
held in Portland, Oregon, August 2-4, 1996, received over 200 submissions.
With such overwhelming interest in this area, the Journal of Intelligent
Information Systems (JIIS) is organizing a special issue on Data Mining.
The information on JIIS and instructions to authors are available at:
The journal published a Special Issue on Knowledge Discovery in Databases
in Volume 4, Number 1, January 1995.
We welcome research and applications papers addressing the following issues
to be submitted to this special issue.
1. Foundations, principles and methodologies of data mining, including
Data mining methods and techniques
Efficiency and scalability of KDD algorithms
Mining different kinds of knowledge from data
Integration of deductive and inductive techniques
Statistics, probability and uncertainty in data mining
Maintenance of mined knowledge and knowledge-base construction
Knowledge evolution through learning
Methods for knowledge discovery in advanced database systems (including
object-oriented, deductive, spatial, temporal, textual, multimedia,
heterogeneous, transaction, and active databases, and global
information systems)
2. Systems and implementations for data mining, including
Knowledge discovery systems, implementations, and performance
Languages and interfaces for knowledge discovery in databases
Interactive data mining and knowledge visualization
Integrated discovery systems
Systems, implementations, and performance for knowledge discovery
in advanced database systems
3. Knowledge discovery applications, including
Successful knowledge discovery application examples in industry,
administration, business, and science or engineering
New application challenges and requirements for data mining
(e.g., science, engineering, education)
The inadequacy of current knowledge discovery mechanisms
Influence of data mining to the advances of database systems
Security and social impact of data mining
IMPORTANT DATES
Submissions Due: November 1, 1996
Review Notice: January 31, 1997
Final Version due: March 15, 1997
Five hard copies of the paper, with the length limited to 20 pages,
should be submitted by November 1, 1996 to
Dr. Jiawei Han
School of Computing Science
Simon Fraser University
Burnaby, B.C.
Canada, V5A 1S6
han@cs.sfu.ca
JIIS Special Issue Guest Co-Editors
Jiawei Han, Simon Fraser University, Canada (han@cs.sfu.ca).
Laks V.S. Lakshmanan, Concordia University, Canada (laks@cs.concordia.ca).
Raymond Ng, University of British Columbia, Canada (rng@cs.ubc.ca).
apparently cart is now marketed by salford systems, san diego. send mail
to salford@weber.ucsd.edu
for further info. they have versions for windows,
dos, macintosh, alphas, other unix workstations, vms/mvs mainframes, etc.
btw, the css-cart had some bugs which the new blokes claim to have fixed.
i haven't used the salford-cart, but according to the material i got from
them it is apparently an enhancement of the original css-cart, and also
provides output in the css-cart format (if you've gotten used to it!).
--nitin
Also, Luis Torgo (ltorgo@ncc.up.pt)
provided the following address:
Salford Systems
5952 Bernadette Lane
San Diego, CA 92120
USA
Fax : 619 582 5759
Phone : 619 582 7534
They have a versions for PC and for Unix workstations.
*Description: Claudien discovers relational rules in relational
databases. Apart from the database the input consists of background knowledge written in Prolog, and
an intensional definition of a finite subspace of first order clausal logic,
specified in the Dlab formalism.
The user-defined hypothesis space is then searched for regularities
output as a set of first-order clauses.
*Comments: Dlab, Claudien's declarative language bias module is
separately available as a Prolog library ready to be plugged into other rule
discovery engines.
=============Deadline (August 1, 1996) is approaching===================
==============================================================
FIRST PACIFIC-ASIA CONFERENCE on
KNOWLEDGE DISCOVERY and DATA MINING (PAKDD97)
Singapore, 23-24 February, 1997
(Co-located with 2nd Pacific-Asia Conference on Expert Systems/
3rd Singapore International Conference on Intelligent Systems)
The first Pacific-Asia Conference on Knowledge Discovery and Data Mining
(PAKDD97) will be held in Singapore. As the range of computer applications
is broadening, more and more data is captured and/or generated. In order to
overcome the situation of ``data rich and knowledge poor'', knowledge discovery
and data mining (KDD) is becoming the focus of many fields from Intelligent
Databases, Machine Learning to Statistics. The aims of the conference are to
cover all aspects of KDD, to bring together researchers and practitioners from
basic and applied research and information industries, and to push forward the
state-of-art of KDD. The conference technical programme will include paper
presentations, posters, invited talks, and tutorials in a two-day event.
Areas of interest include, but are not limited to:
Knowledge Representation and Acquisition in KDD
Data Mining and Data Warehousing
Data Cleaning, Preprocessing and Postprocessing
Data and Dimensionality Reduction
Knowledge Reuse and Role of Domain Knowledge
Data Mining Tools
KDD Framework and Process
Security and Privacy Issues in KDD
Mining in-the-Large vs Mining in-the-Small
Management Issues in KDD
Machine Learning, Statistical and Visualization Aspects of KDD
Successful/Innovative Applications in Science, Government, Business and Industry
The proceedings will be published by an international publisher and will be
available at the conference.
PAKDD is Organized by
Information Technology Institute & National University of Singapore
in Cooperation with
National Computer Board, Singapore
Singapore Computer Society AI Chapter
Japanese Society for Artificial Intelligence
Korea Advanced Institute of Science & Technology
Korea Expert Systems Society
Submission Information:
Tutorial proposal comprising summary, course outline and a brief biography
of the speaker(s);
4 copies of full paper (3000-5000 words)
4 copies of applications paper (about 3000 words)
Submission Address:
Dr. Hongjun Lu
Department of ISCS
National University of Singapore
Kent Ridge, Singapore 119260
You may contact Dr Lu for your conference exhibition.
Important Dates:
1 Aug., 1996 for submission of papers and proposals
15 Oct., 1996 for notification of acceptance
15 Dec., 1996 for receipt of camera-ready manuscripts
--------CALL FOR PAPER--------CALL FOR PAPER--------CALL FOR PAPER-------
Conference General Chair:
Hing-Yan Lee, Japan-Singapore AI Centre (JSAIC), Information Technology Institute
Organizing Committee:
Lynica Foo, JSAIC, Kwok-Leong Hui, JSAIC
Bing Liu, National U. of S'pore (NUS) Hwee-Leng Ong, JSAIC
Angeline Pang, JSAIC
Publicity Chair:
Huan Liu, NUS
Programme Co-Chairs:
Hongjun Lu Hiroshi Motoda
Dept of Info. Sys. & Comp. Sci. Institute of Sci. & Indus. Research
National University of Singapore Osaka University, Japan
Programme Committee:
Arbee L. Chen, National Tsing Hua U., Taiwan
David Cheung, Hong Kong U.
Roger Hsiang-Li Chiang, Nanyang Technological U., S'pore
Son Dao, Hughes Research Lab., USA
David Dowe, Monash U., Australia
Jiawei Han, Simon Fraser U., Canada
Se Jung Hong, T.J. Watson IBM Lab., USA
Steven H. Kim, KAIST, Korea
Masaru Kitsuregawa, Tokyo University, Japan
Jae-Kyu Lee, KAIST, Korea
Bing Liu, NUS, S'pore
Huan Liu, NUS, S'pore
Peter Milne, CSIRO, Australia
Riichiro Mizoguchi, Osaka U., Japan
Shinichi Morishita, IBM Tokyo Research Laboratory, Japan
Raymond Ng, UBC, Canada
Anne Ngu, The University of New South Wales, Australia
Shojiro Nishio, Osaka U., Japan
Hwee-Leng Ong, JSAIC
S. Seshadri, IIT Bombay, India
Rudy Setiono, NUS, S'pore
John Shafer, IBM Almaden Research Center, USA
Pengfei Shi, Shanghai Jiaotong U., China
Atsuhiro Takasu, NCSIS, Japan
Takao Terano, The University of Tsukuba, Japan
Robert Veranas, NUS, S'pore
Xindong Wu, Monash U., Australia
Beat Wuthrich, HKUST, Hong Kong
Suk-Chung Yoon, Wildener University, USA
Philip Yu, T.J. Watson IBM Lab., USA
Bo Zhang, Tsinghua U., China
KDD Nuggets Index