I dont have a collection of data which I try and extract features from. I have a set of real world experiments from which features are retreived.

If the tests are sparse and I received missing data - can anyone point me to procedures to deal with sparse but pertinent features?

I do not care to deal with procedures that add "default" data etc - these types do not relate to real world experiments (in my context).

The purpose of feature selection here is like all other classification problems - I need aspects to distinguish the classes/categories apart. Imputation is not a possibility here as the data I get from real world experiments are just that - real - they come from living individual organisms, so imputing them does not make sense. I need the real data from the real individual. Thanks

  • $\begingroup$ @urema As SmallChess pointed out, this is how the site works, even if it is not always 'just'. You'll get more points for every upvote then you loose for a down vote. There is absolutely no point in calling downvoters a troll. Try to see, if there is something to learn from that or not. If not, move on. Your question is insofar unclear as it does not state, how much of a problem with missing data you have and whether "leave those cases out" is an answer. Also, you don't explain, what is the purpose of the feature selection and why imputation is not an answer for you. $\endgroup$
    – Bernhard
    Oct 30, 2017 at 12:51
  • $\begingroup$ I never called anyone anything, I used troll as a verb sir. I don't have any trouble, as I am not trying any classification - I want some pointers to literature or procedures that deal with such a consideration,, so I can prepare before attempting to stab in the dark on real data. Fail to prepare and prepare to fail - I am dealing with real world experiments, therefore I want to check out controlling such missing data for individual organisms $\endgroup$
    – urema
    Oct 30, 2017 at 12:53
  • 1
    $\begingroup$ You want to "distinguish the classes/categories apart" but you are "not trying any classification"? At the same time you wonder, why someone finds this question "unclear or not useful"? Good luck! $\endgroup$
    – Bernhard
    Oct 30, 2017 at 13:08
  • $\begingroup$ Not to be pedantic, but no I am not running any classifiers on any data AT THE MOMENT - I am preparing data, designing a method of classification - I see now that getting real world experiments may provide me missing data - so I am preparing, by analysing the domain to seek for techniques for overcoming this if it occurs - Prudence $\endgroup$
    – urema
    Oct 30, 2017 at 13:18
  • 2
    $\begingroup$ Please elaborate how - in you case - handling missing real world data differs from handling missing 'other' types of data (e.g. why imputation does not apply)? By the way: imputation can also be used for 'real world' data. See for example the biomedical/epidemiological field, where risk prediction models are sometimes implemented into practice by imputing the missing data with multiple plausible values (say 10), taking the average of the predicted risk/predicted class as the final predicted risk/class. $\endgroup$
    – IWS
    Oct 30, 2017 at 13:39

1 Answer 1


Linear regression, primary components etc need complete data to work upon. The most standard variant of dealing with that are omission of observations, if data is missing at random or multiple imputation, which you say does not make sense in your setting.

Classification trees and Random forests work very well with missing data, probably even, if not missing at random.

If there are missing values, CART and CRUISE use alternate splits on other variables when needed, C4.5 sends each observation with a missing value in a split through every branch using a probability weighting scheme, QUEST imputes the missing values locally, and GUIDE treats missing values as belonging to a separate category. (citation from http://www.stat.wisc.edu/~loh/treeprogs/guide/wires11.pdf )



  • $\begingroup$ Perfection sir, this will be good reading along with the Columbia statistics Chapter 25 pdf that is available online..... $\endgroup$
    – urema
    Oct 30, 2017 at 13:09

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