R: GLMNET odd behavior when model is reran

问题:

I am trying to use LASSO for variable selection, and attempted the implementation in R using the glmnet package. This is the code I wrote so far:

 set.seed(1)
 library(glmnet)
 return =  matrix(ret.ff.zoo[which(index(ret.ff.zoo) == beta.df$date[1]),])
 data = matrix(unlist(beta.df[which(beta.df$date == beta.df$date[1]),][,-1]), ncol = num.factors)
 dimnames(data)[[2]] <- names(beta.df)[-1]
 model <- cv.glmnet(data, return, standardize = TRUE)
 coef(model)

This is what I obtain when I run it the first time:

 > coef(model)
 15 x 1 sparse Matrix of class "dgCMatrix"
                       1
 (Intercept) 0.009159452
 VAL         .          
 EQ          .          
 EFF         .          
 SIZE        0.018479078
 MOM         .          
 FSCR        .          
 MSCR        .          
 SY          .          
 URP         .          
 UMP         .          
 UNIF        .          
 OIL         .          
 DEI         .          
 PROD        .             

BUT, this is what I obtain when I run the SAME code once more:

 > coef(model)
 15 x 1 sparse Matrix of class "dgCMatrix"
                       1
 (Intercept) 0.008031915
 VAL         .          
 EQ          .          
 EFF         .          
 SIZE        0.021250778
 MOM         .          
 FSCR        .          
 MSCR        .          
 SY          .          
 URP         .          
 UMP         .          
 UNIF        .          
 OIL         .          
 DEI         .          
 PROD        .          

I am not sure why the model behaves this way. How would I be able to choose a final model if the coefficients change at every run? Does it use a different tuning parameter $\lambda$ at every run? I thought that cv.glmnet uses model$lambda.1se by default?!

I have just started learning about this package, and would appreciate any help I can get!

Thank you!

回答1:

The model isn't deterministic. Run set.seed(1) before your model fit to produce deterministic results.

回答2:

You need to feed the same nfolds and foldid to both models. Check help(cv.glmnet) for more details. This will make the cross-validation is identical and you should get the same model if you run the models on the same data-set.

回答3:

Just a supplement to the answer of @nograpes. Each time before fitting the model, the same seed should be set. In short, one seed is only available for one model. For example,

set.seed(1)
model1 = cv.glmnet(x, y, alpha = 0, family = 'binomial')
model2 = cv.glmnet(x, y, alpha = 0, family = 'binomial')

For the code above, the coefficients of model1 and model2 could be different.

set.seed(1)
model1 = cv.glmnet(x, y, alpha = 0, family = 'binomial')
set.seed(1)
model2 = cv.glmnet(x, y, alpha = 0, family = 'binomial')

Only after you set the same seed before fitting the model, the result are exactly the same.