I am working on the KDD 2009 Cup Data Set (The Small one) and I have a question about preprocessing data. It has a lot of columns with null values, some of them have more than 90% of missing. Reading some papers on the challenge I noticed that even though these columns have this amount of null competitors usually did not excluded them. I was doing some validation and in order to argue that if I should drop it or not, I calculated the mean of target variable comparing null vs non-null, something like this:
You can see for example, that for variable Var118, for churn and appetency targets they do not have a significant difference on mean target values. However, for up_selling, we can notice some discrepancy. My question is: is that a correct approach? I am having second thoughts because my data set is extremely unbalanced, and maybe my target means will be biased. Anyone have had some experience on this type of problem? I would love discuss more about that.