Feature presence in Maxent training data

[frank]

Hi,

I'm trying to use Maxent package for text classification, but I get confused
when I wrote the training data. I use unigram features for the experiment using their
presence as value. The value is 1 if the feature exists in document and 0
if the feature doesn't exist.

I try 2 different approach or feature representation:

First, if the feature exists in document I wrote 1_featurelabel to incorporate its existence in data and if it doesn't I wrote 0_featurelabel to give information about its non-existence.
example --> 1_a 1_b 0_c 0_d 1_e topic1

Second, if the feature exists in document I write 1_featurelabel and if it doesn't, I didn't write anything.
example --> 1_a 1_b 1_e topic1

Which one of the representations that is correct?

Thanks

[Thomas Morton]

Hi,
   The first approach is the most typical as the lack of presence is implicitly modeled as getting zero weight.  The model will automatically assign the feature a value of 1 so you don't need to encode that in your features (not that it will hurt anything the way it is).

Hope this helps...Tom

[frank]

Hi,

From your answer, I don't need to encode the lack of presence in my data. So, the second
approach is the one I have to choose. Is it right? Just to clarify.

Thanks for your help.

[Thomas Morton]

Correct.  I think I referred to them incorrectly in my last post...Tom

[frank]

Hi,

Ok. I got it. It's really helpful. Thanks.