0
$\begingroup$

I use a logistic regression model for reranking some documents where a normalized features of some candidates may have negative real value so that its predicted value may get lower score(low likelihood for positive class).

I am doing this on Weka within a Java program without using arff file.

Can I use a negative value for a given feature vector? Or should I normalize the dataset so that all negative value be zero?

The reason why I ask this is that I intend to use such a negative valued feature for negative boosting in information retrieval viewpoint. What can be best when I want to embed such negative boosting feature in learning ranking model based on logistic regression.

$\endgroup$
0
$\begingroup$

Yes you can, for a feature with positive and negative values, logistic regression will find a coefficient that works best with all values. You should not set all negatives to zero, as you will be losing information.

If you want to add some negative boosting, then it may be worthwhile to create a new categorical variable with two values: 'Positive' or 'Negative'. In this way, the regression can assess the impact of the feature being positive or negative, and can also use the numerical values for prediction as well.

I think it's important to understand the contextual significance of a positive vs negative value. Is the feature continuous from -inf to +inf (or whatever your limits are)? If two candidates have a score of 1 and -1, is the difference between them similar to two candidates that have a score of 88 and 90? In both cases the score difference is 2. I don't know the context of your model, but I think this is an important consideration.

$\endgroup$
  • $\begingroup$ Thank you very much for your advice. I want to upvote but my reputation doesn’t allow me to. $\endgroup$ – qualia universe Nov 23 '17 at 1:26

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.