KDD Nuggets  -- September 23, 1993

Contents: 
	* Wray Buntine: "interestingness" should be defined via decision theory
	* Jerzy W. Grzymala-Busse: Re: KDD success
	* Sandra Oudshoff: tools for information harvesting?

The KDD Nuggets is an informal list for the dissemination of information
relevant to Knowledge Discovery in Databases (KDD), such as announcements 
of conferences/workshops, tool reviews, application success/failure stories,
interesting ideas, outrageous opinions, etc. 

If you have such a contribution, please email it to kdd%eureka@gte.com 

Mail requests to be added/deleted also to kdd%eureka@gte.com.
(Note: If you received this message you are on the list!)
--------------------------------------------------
Date: Thu, 12 Aug 93 12:09:15 PDT
From: Wray Buntine <wray@ptolemy.arc.nasa.gov>

>  We should elaborate some clear definitions of fundamental terms of KDD
>  (e.g. "pattern"), clarify some vague concepts (e.g. "interestingness of
>  finding"), identify the main subgroups of applications (discovery in
>  databases (relational, time series ...), scientific discovery, image
>  database discovery ...) and describe the (methodical) differences between
>  these subgroups. 

	Sounds great, but I just hope this doesn't end up in creating
yet another jargon.  So far we have machine learning, statistics,
neural networks, statistical physics, information theory and
computational learning theory people addressing learning type things
with all their own vocabulary.  So I'd hope some unification is
possible.
	For my part, I think "interestingness" is best explained using
the language of decision theory,  i.e. utility theory and probability
theory.  This is certaintly the case for what I've seen at KDD'91 and
'93.  There is a well developed theory AND practice of decision theory,
so don't need to reinvent the wheel here, and it certainly will
lead to cleaner specification of "interesting" plus derivation
of matching learning algorithms.

Wray Buntine
	
------------------------------------------------------
Date:     Thu, 9 Sep 93 14:57:05 CDT
From: Jerzy@cs.ukans.edu
Cc: "Jerzy W. Grzymala-Busse" <jerzy@cs.ukans.edu>
Subject:  Re: KDD success


	A prototype of an expert system to assess preterm labor risk for 
pregnant women was developed recently by a team under general guidance 
of M. Van Dyne (Intellidyne, Inc.), who was also responsible for 
development of the main system.  J. Grzymala-Busse (U. of Kansas) has 
been responsible for rule induction from databases using system LERS 
(Learning from Examples based on Rough Sets).  L. Woolery (U. of 
Missouri) has been an expert in the domain and analyzed data 
statistically.  The project was funded by the National Institutes of
Health, grant number 1 R43 NR02899-01A1.

	In the project three databases (provided by a local perinatal
center and two national home uterine monitoring companies) 
containing information about preterm and fullterm delivery were 
processed.  Every database was divided by half, 50% entities were used 
for rule induction by LERS and remaining 50% for validation.  The sizes 
of databases are the following: 52 variables and 3204 entities (first 
database), 77 variables and 2436 entities (second database) and 82 
variables and 7224 examples (third database).

	Rules generated from LERS on the databases were combined into
a single rule base in a prototype expert system. Each rule was given a
priority score. Because of low accuracy rates, rules generated from the
third database were not included in the prototype expert system. Data
from  each of the three databases were tested on the prototype expert
system combined rule base to determine the final predictive accuracy
rates.

	Currently used manual tools to assess preterm labor risk have 17 
- 38 % accuracy in determining preterm risk (McLean, Walters & Smith 
(1993).  The system developed in the first phase of the project has the 
following accuracies for predicting preterm or fullterm delivery: 
	89% (first database), 
	59% (second database), and 
	53% (third database).

---------------------------------------------------
From: oudshoff@sun032.research.ptt.nl (Sandra Oudshoff)
Subject: tools for information harvesting?
Organization: PTT Research, Groningen, The Netherlands
Date: Thu, 9 Sep 1993 09:18:00 GMT

Dear fellow netters,

Our AI group here is very interested in buying a commercial product
that can be used for information harvesting, i.e. extracting rules or trends
automatically from (large) amounts of data. We have heard about 
"Information Harvesting" from the company with the same name. 

If you have any information about / experience with similar tools, 
I would greatly appreciate it if you shared that with me through e-mail.
I`ll summarize on the net if there is interest. Any and all info or help
will be greatly appreciated.

Thanks in advance,

Sandra Oudshoff

e-mail to: a.m.oudshoff@research.ptt.nl