NetNews Usenet Archive 1992 #31/NN_1992_31.iso/spool/comp/ai/4693

Internet Message Format | 1992-12-23 | 3.1 KB

Xref: sparky comp.ai:4693 sci.math.stat:2674
Path: sparky!uunet!munnari.oz.au!uniwa!newsman!newsman!winzar
From: winzar@newsman (Hume Winzar)
Newsgroups: comp.ai,sci.math.stat
Subject: Re: Learning from subjective data
Date: 24 Dec 1992 18:34:40 GMT
Organization: Commerce, Murdoch University
Lines: 50
Message-ID: <winzar.98.0@newsman>
References: <[e-mail]> <[e-mail]> <[e-mail]>
NNTP-Posting-Host: 134.115.160.7

In article <[e-mail]> [e-mail] (R. Bharat Rao) writes:
>[e-mail] (Russell G. Almond) writes:
>>R. Bharat Rao ([e-mail]) writes:

>>> I was wondering if anyone knew of any work that has been done on
>>> learning from subjective data.  For instance, you may have a data set
>>> of events with a number of independent attribute (x1...xn) and a
>>> single dependent attribute y.  However, y is a subjective rating.

In addition to the Psych dept. suggested by R.G. Almond, you might also
check out the work done by Transport Economists, and transport engineers.

Specifically models of Stated Preference of different transport mode
alternatives.  The problem domain is to work out the relative influences on
alternative evaluations for transport.  With this type of data they can make
an informed guess about, say, the effects on road traffic congestion of
installing a new railway station.

The approach goes something like this:  Attribute combinations can be
formulated with respect to Mode (car, bus, train, etc) Speed, Convenience,
Cost, etc.  Maybe thousands of combinations are possible which usually are
reduced to a smaller number in an orthogonal array using fractional
factorial design technique.  Survey respondents are asked to evaluate a
subset of this array and make a rating or ranking on each, or sometimes just
choose their first preference.  The data are combined and, depending on
their type, run through a general linear package, or a Multinomial Logit
(MNL) package, to derive parameter estimates for each attribute and
attribute level.

An obvious, and critical, assumption here is that each respondent is making
evaluations that are random variations around a single shared evaluative
function.  If you do not accept this assumption, then you have to create
separate models for each person, and gather all data from each person.  (I
do that with marketing and psychoogical research but with very simplified
models.)

Check out the work of Jordan Louviere, now at Salt Lake City.  Or David
Hencher at the Transport Research Institute at the Graduate School of
Business, University of Sydney, Australia.  (There are also transport
research institutes in LA and London but I don't know any names, sorry.)

  - - - - - - - - -
 |      _--_|\     |   Hume Winzar
 |     /      \    |   Commerce School,
 |     *_.--._/    |   Murdoch University,
 |           v     |   Perth, Western Australia
  - - - - - - - - -
                       E_Mail [e-mail]
                       Phone: [phone]
                       Fax:   [phone]