rank SNPs after elastic net modeling

Question

[rheab1230]

Hello everyone,

I have a question related to ranking of features after modeling step.

I had total 1204 features in original predictor file. After doing elastic net modeling(using cv.glment package), out of 1204 128 features got selected.

I want to see the ranking of these features. Like out the selected features, which one has higher rank in explaining the variance?

I tried to do backward algorithm, wherein I removed one feature at a time and then do remodeling with n_1 feature and looked at correlation value.

But in my case for certain feature its not giving good correlation.

For example: if feature 1 in model has effect size of 1.573 its removal during model and then doing correlation should be lower as compared to correlation with its presence.

But in my case its removal doesn't affect the correlation at all.

Does anyone know of any method/algorithm that can be used to rank features after modeling and then do statistics to see its performance?

Thank you

[Shred]

It depends on the task you're trying to accomplish. Is a classification, a regression, or something else?

[rheab1230]

Its a regression as well.

[Shred]

So you're doing a correlation (Pearson?) between the regressed obtained from the full model and a model with n-1 features? To answer what?

The feature importance with an elastic net could be "derived" from the absolute value of the coefficient of each feature. The elastic net uses a combination of L1 and L2 penalties to already do a feature selection.. which question do you want to answer here?

[rheab1230]

I want to see how the correlation between predicted and residual expression change after removal of n_1 snps, in my case I want to rank the SNPs obtained in model file. In doing so, i am using correlation, an unimportant SNP its removal shouldn't affect correlation a lot. whereas a important snps in the model file for a particular gene removal should affect the correlation a lot.

[Shred]

You're trying to doing what the elastic net has already do but in a wrong way. The fitted model already has information about how much a feature is important in your model.

Check the StatQuest video about