lippolt@dnlunx.UUCP (Ben Lippolt) (09/06/88)
Hello Netlanders,
We have the following problem:
We have about 200 items. Each item belongs to one of 13 classes and
is described by 12 attributes. The values an item has for these 12
attributes are not absolute, however, but are expressed relative to
the other items. Like this:
attr1 attr2 attr3 attr4
3 2 4 3
2 3 3 4
1 4 2 2
4 1 1 1
The numbers refer to items 1 to 4. Let's say that item 1 belongs to
class 1, item 2 to class 3, item 3 to class 6 and item 4 to class 3.
There is a correlation between the class an item belongs to and the
positions it has for each attribute. We can see, for instance, that
item 3 is ranked above item 1 for all attributes and that the class
of item 3 is higher than the class of item 1. If we look at items 2
and 4, we see that of these two for some attributes item 2 is ranked
higher and for some attributes item 4. Both items belong to the same
class.
What we want to do now, is to check the consistency of the
correlation between the class an item belongs to and the relative
positions it occupies for the twelve attributes. For instance, an
item that is ranked very high for each attribute should not belong to
the same class as an item that is ranked very low for each attribute.
We want to start with e.g. 50 items and check for each new item that
we add whether its class and positions are consistent with the other
items.
Our questions are:
Can we use an inductive tool for this problem? Are 50 cases, with
12 attributes each, enough to start working with? Can we find
inconsistencies, which might be rather vague, with such a tool? Is
it possible to incorporate fuzzy logic in an inductive tool? Which
tool should we use?
Any comments are highly appreciated.
Ben Lippolt (..!mcvax!dnlunx!lippolt, or lippolt@hlsdnl5)
Marlies van Steenbergen (..!mcvax!dnlunx!marlies)
PTT Research, Neher Laboratories.leverich@randvax.UUCP (Brian Leverich) (09/09/88)
If you're actually trying to "solve" the problem you posed rather than conduct AI research, you might want to review the standard statistical approaches to categorizing data. The bean estimators have been doing this stuff for decades. Even if your main target is AI research, contrasting the performance of AI and statistical approaches is generally interesting and occasionally flat-out embarrassing. Cheers. -B -- "Simulate it in ROSS" Brian Leverich | U.S. Snail: 1700 Main St. ARPAnet: leverich@rand-unix | Santa Monica, CA 90406 UUCP/usenet: decvax!randvax!leverich | Ma Bell: (213) 393-0411 X7769