Currently I am using Random Forest approach for Missing Values Imputation from missForest package in R.

I faced the following problem: the algorithm works much longer than any other imputation approaches like KNN or SVD. Does anybody know how to overcome this problem or maybe anyone had experience with using it in other packages?

I'd really appreciate any help cause I even failed to find the full description of the approach in the internet.

  • $\begingroup$ I would not advise anyone to use a method that does not have a full description. Also, please specify the reasons that older multiple imputation methods will not work for your problem. I know the older methods properly take into account imputation variability and imputation model uncertainty. Make sure that missForest also does this. $\endgroup$ – Frank Harrell Aug 29 '14 at 14:11
  • $\begingroup$ Thanks for your comment! What I am trying to do is: to check which imputation algorithm perform better in average for different levels of missing values. For this purpose I took Netflix dataset and applied several algorithms for different percentage of missing data. To begin with, I started with just Random Forest, KNN and SVD. Surprisingly best results were achieved for Random Forest (missFOREST). Also surprisingly KNN gave better results than SVD. Problem with multiple imputation approach is that it also takes a lot of time for the calculations. I will need to apply approach in real time... $\endgroup$ – Oleg Sep 1 '14 at 9:22
  • $\begingroup$ What is key is exactly how you measured performance in terms of bias, precision, confidence interval coverage, bias-corrected predictive discrimination (using the bootstrap or cross-validation), etc. What approach did you use? $\endgroup$ – Frank Harrell Sep 1 '14 at 12:27
  • $\begingroup$ Currently all the analysis is based on just RMSE calculation. I have the whole Netflix dataset, randomly take-out some particular percentage of data, then impute it and estimate RMSE. This procedure is done 1000 times in order and the results are averaged to avoid bias. I thought that since the dataset is huge, I am kind of guaranteed from overfitting and all corresponding staff? is there something else which should be considered during imputing? or all the measures that you mentioned must be used? $\endgroup$ – Oleg Sep 1 '14 at 14:43
  • $\begingroup$ I guess you could say that an RMSE of an appropriate quantity (predicted probability, predicted logit, etc.) is a good standard; one just doesn't know how good is good enough in terms of choosing an imputation method. $\endgroup$ – Frank Harrell Sep 1 '14 at 17:00

Your Answer

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

Browse other questions tagged or ask your own question.