Related attributes

gvenez

How do I solve the problem of related attributes?

For example I have a dataset with the following structure:

Economic Code 1	Percentage	Economic Code 2	Percentage	Economic Code 3	Percentage	Success
EC12	60%	EC13	30%	EC14	10%	Yes
EC13	60%	EC15	20%	EC12	20%	No
EC19	50%	EC13	50%	EC14	0%	Yes

The first and second attribute are related. The second attributes shows how much percentage of weightage should be given to the first. Similarly for the third and fourth. An attribute should be neglected if the value of weightage is zero.

How do I go about solving this? Any tips or suggestions are welcome.

IngoRM

Hi,

well, there are a lot of options but two of them are pretty simple and worth a try:

• If the values EC12, EC13 etc. are actually ordered, you could try to translate those ordinal values into numerical ones and directly calculate a new attribute by, for example, multiplying the weights with the numerical values.
• The second option is much easier and would definitely my first choice: why do you bother at all? Just let the modeling scheme sort this out and simply put you data into a classification scheme.

The rational for the second suggestion is simply that many data mining schemes are capable of handling those feature interactions themself. Let's take a decision tree for example. The resulting model could look like

if "Percentage of EC 1" > 80%
--- if "EC1" = EC12 then Yes
--- else No
else if ...

I hope you get the point. This is even more true for other learning schemes which take multiple attributes into account at the same time. Those often lack understandibility though, so this might not be an option then.

Cheers,
Ingo