I created my first randomForest model (regression) in order to predict sensitivity to a particular drug in breast cancer cell lines. Total features were 35000 and I used default mtry for regression (N/3) and 30000+1 for ntree.
Then, I determined the optimal number of predictors that minimized the MSE error using the ‘rfcv’ function in the randomForest library (50 replicates with five-fold cross validation). This resulted in a selection of 23 predictors.
Then, I created a second randomForest model based on these 23 features in order to recompute variable importance values, using default mtry=N/3=7.
Results of this second RF model are very good but I noticed that if I lower the mtry parameter I can get a lower MSE error, reaching the best results at mtry=1, but that would be like using univariate decision trees and I'm not really sure if this is desirable.
Can someone kindly give me some advice?
Thank you
A basic comment comment concerning model fitting: on any one dataset it is possible to build a model that fits the dataset perfectly, but as this dataset is just some (random, hopefully) sample of the total population, this "perfect" model likely does not fit any other same at all well. I suggest that with a large dataset, you take a sample (say, 50% of datapoints, sampled randomly), use this for building the model, then test it on the rest of the dataset. If this still gives a good fit, you probably have a usable model.
Why don't you compare to a "static" tree (e.g. use rpart() from package "rpart") using the 23 variables and see if result for mtry=1 holds? In that case may be you did a very good preselection... or all 23 vars are highly correlated (?!)
Hello, thank you for your replies and suggestions.
As you said, I realized that most of the variables are part of common pathways and are coexpresed and heavily interconnected one to another. This is likely why creating decision trees from this small subset is no more informative than looking at the single gene information.
- Badges
- Science topic
- Similar topics
- Bioinformatics
- Bioinformatics and Computational Biology