Suppose we have an input feature that is highly predictive of the outcome we want to predict. However, the feature is missing on 99% of the samples in the data set. What is the best way to use this feature in building a prediction model? Any ideas would be greatly appreciated!
1
-
$\begingroup$ The best way would be to redo the data collection stage, and this time make sure to capture the value of this variable. There's no magic, if you don't have the data you can't use it ;) $\endgroup$– ErwanOct 5, 2020 at 11:44
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
3
Its best to remove such a variable. Reasons are following:
- Artificial imputation can add bias and result cannot be justified because 99% data for the particular variable was artificially created.
- The variables/features that you choose for building the predictive model should have low correlation with the target/outcome variable/feature. Because, variable that are highly correlated with the target/outcome variable reduces the model predictive performance.
- You want a model which has low bias and low variance.
-
$\begingroup$ regarding your second point: is it not actually ideal if we have a feature that is highly predictive of the target outcome? At the end of the day, that's how we want a good model to do: accurately predict the outcome. $\endgroup$ Oct 5, 2020 at 15:45
-
$\begingroup$ Also, about your 3rd point: how do input features being correlated impact the model variance? It is not very clear to me $\endgroup$ Oct 5, 2020 at 15:46
-
$\begingroup$ @user2348674 reading your questions gives me a strong feeling that your fundamentals in applied statistics are unclear. I recommend the book Applied Predictive Modelling (Kuhn & Johnson, 2013). Read it, will help you clarify your doubts. $\endgroup$– mnmOct 5, 2020 at 23:40