I'm using R for machine learning. The objective is to classify the onset of disease (Two-class).
Before conducting a machine learning algorithm, I ran the glmnet (to utilize elastic net) to reduce the number of total variables.
My dataset has a lot of categorical variables such as gender, educational level, marital status, living regions, ...
Whereas some variables are binary (i.e., male/female in gender), others are more than two (i.e., elementary/middle/high/college/professional in educational level).
To enter those variables in the glmnet, I converted them into the dummy variables. For example, 'gender' was split in two ('male' (1 for male and 0 for female) and 'female' (1 for female and 0 for male), and educational level was split in 5 ('elementary', 'middle', 'high', 'college', 'professional'). This way, all the categorical variables were split in by the number of their levels.
After the glmnet was conducted, I got the coefs for each dummy variable as following.
male0 0.159 male1 0.000 female0 0.000 female1 0.004 elementary0 -0.127 elementary1 0.000 middle0 . middle1 . high0 . high1 . college0 0.719 college1 . professional0 1.360 professional1 -0.084 ......
My question is two.
(1) Can I drop only dummy variables with coef = . in the final machine learning model?
(2) How can I interpret the importance of the 'gender' and 'education' but not the splitted dummy variables (male, female, elementary, ...) ?