1
$\begingroup$

I am working with a regression problem. I have some features which are only available for a few instances. But with those few instances based on those features we can build a model which gives reasonable results. At the same time I have other features which are available for almost all instances. How should I combine those two type of features?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

I suggest two approaches:

  1. First of all, separating the regression problem into several sub-problems considering the different combinations of features you have. In other words, group those instances where you have all the features and deal with them separately. Similarly, proceed with another group of instances with another different combination of features. This solution is suitable for cases where the possible combinations of features is small.
  2. Second (and more feasible) solution consists of creating a regression subproblem for each feature. Then, the global regression problem may be solved by fusioning the information provided by each regression sub-problem separately. In this case, if some feature is missing, it is automatically not considered for classification.

HTH

Improve this answer
$\endgroup$
2
  • $\begingroup$ hi, desantos thanks for reply. The ideas are very good! For option one I wonder if we could lose information by separating the data. Actually we tried idea similar to option 2. We built 2 classifier one for each group of features, then try to stack another classifier on top of them. But it seems that way did not bring much improvement. Maybe we did not used the right classifer to stacking the results. $\endgroup$
    – gstar2002
    Commented Sep 6, 2013 at 9:48
  • $\begingroup$ Probably you are overfitting your data as you are using one classifier to aggregate the other classifiers (I dont know). But I suggest to use other fusion technique such as harmonic/geometric mean of the scores obtained by each classifier and then create a new score. Check out if this score is adequate for your requirements. $\endgroup$
    – a.desantos
    Commented Sep 11, 2013 at 10:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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