LASSO stands for Least Absolute Shrinkage and Selection Operator. The algorithm is another variation of linear regression like ridge regression. We use lasso regression when we have large number of predictor variables.
Things You Will Master
- Overview - Lasso Regression
- Training and Predicting Lasso Regression Model
- Getting the list of important variables
Overview - Lasso Regression
Lasso regression is a parsimonious model which performs L1 regularization. The L1 regularization adds a penality equivalent to the absolute of the maginitude of regression coefficients and tries to minimize them. The equation of lasso is similar to ridge regression and looks like as given below.
LS Obj + λ (sum of the absolute values of coefficients)
Here the objective is as follows: If λ = 0, We get same coefficients as linear regression If λ = vary large, All coefficients are shriked towards zero
The two models, lasso and ridge regression are almost similar to each other. However, in lasso the coefficients which are responsible for large variance are converted to zero. On the other hand, coefficients are only shrinked but are never made zero.
Lasso regression analysis is also used for variable selection as the model imposes coefficients of some variables to shrink towards zero.
What does large number of variables mean?
- Large number here means that model tend to over-fit. Theoratically, minimum ten variables can cause overfitting problem.
- When you face computational challenges due to presence of n number of variables. Given todays processing power of systems, this situation arises rarely.
The following diagram is the visual interpretation comparing OLS and lasso regression.
The LASSO is not very good at handling variables which show correlation between them and thus can sometimes show very wild behaviors.
Training Lasso Regression Model
The training of lasso regression model is exactly same as that of ridge regression. We need to identify the optimal lambda value and then use that value to train the model. To achieve this, we can use the same
glmnet function by passing
alpha = 1 argument. When we pass
alpha = 0,
glmnet() runs a ridge regression and when we pass
alpha = 0.5 the glmnet runs another kind of model which is called as elastic net and is a combination of ridge and lasso regression.
- We use
cv.glmnet()function to identify the optimal lambda value
- Extract the best lambda and best model
- Rebuild the model using
- Use predict function to predict the values on future data
For this example we will be using
swiss dataset to predict the fertility based upon Socioeconomic Indicators for the year 1888.
# Loaging the library library(glmnet) # Loading the data data(swiss) x_vars <- model.matrix(Fertility~. , swiss)[,-1] y_var <- swiss$Fertility lambda_seq <- 10^seq(2, -2, by = -.1) # Splitting the data into test and train set.seed(86) train = sample(1:nrow(x_var), nrow(x_var)/2) x_test = (-train) y_test = y_var[test] cv_output <- cv.glmnet(x_vars[train,], y_var[train], alpha = 1, lambda = lambda_seq) # identifying best lamda best_lam <- cv_output$lambda.min
# Output  1.995262
Using this value, let us train the lasso model again.
# Rebuilding the model with best lamda value identified lasso_best <- glmnet(x_vars[train,], y_var[train], alpha = 1, lambda = best_lam) pred <- predict(lasso_best, s = best_lam, newx = x_vars[test,])
Finally, we combine the predicted values and actual values to see the two values side by side and then you can use the R-Squared formula to check the model performance. Note - you must calculate the R-Squared values for both train and test dataset.
final <- cbind(y_var[test], pred) # Checking the first six obs head(final)
# Output Actual Pred Courtelary 80.2 69.92666 Delemont 83.1 76.15793 Franches-Mnt 92.5 75.16697 Moutier 85.8 70.33981 Glane 92.4 76.61480 Veveyse 87.1 76.34404
Sharing the R Squared formula
The function provided below is just indicative and you must provide the actual and predicted values based upon your dataset.
actual <- test$actual preds <- test$predicted rss <- sum((preds - actual) ^ 2) tss <- sum((actual - mean(actual)) ^ 2) rsq <- 1 - rss/tss rsq
Getting the list of important variables
To get the list of important variables we just need to investigate the beta coefficients of final best model.
# Inspecting beta coefficients coef(lasso_best)
# Output 6 x 1 sparse Matrix of class "dgCMatrix" s0 (Intercept) 55.16706057 Agriculture . Examination -0.30124968 Education . Catholic 0.04700893 Infant.Mortality 0.84730322
The model indicates that the coefficients of Agriculture and Education have been shriked to zero. Thus we are left with three variables namely; Examination, Catholic, and Infant.Mortality