[R] Confused - better empirical results with error in data

[Noah Silverman]

Hi,

I have a strange one for the group.

We have a system that predicts probabilities using a fairly standard svm 
(e1017).  We are looking at probabilities of a binary outcome.

The input data is generated by a perl script that calculates a bunch of 
things, fetches data from a database, etc.

We train the system on 30,000 examples and then test the system on an 
unseen set of 5,000 records.

The "real world" results on the test set looked VERY good.  We were 
really happy with our model.

The, we noticed that there was a big error in our data generation script 
and one of the values (an average of sorts.) was being calculated 
incorrectly.  (The perl script failed to clear two iterators, so they 
both grew with every record.)

As an quick experiment, we removed that item from our data set and 
re-ran the process.  The results were not very good.  Perhaps 75% as 
good as training with the "wrong" factor included.

So, this is really a philosophical question.  Do we:
     1) Shrug and say, "who cares", the SVM figured it out and likes 
that bad data item for some inexplicable reason
     2) Tear into the math and try to figure out WHY the SVM is 
predicting more accurately

Any opinions??

Thanks!