In Demongo2009/ZUM_Appliances: What the package does (short line)

TWORZENIE MODELU REGRESJI LASSO

Wczytanie danych

train<-read.csv("../data/energy_data_train.csv")
train <- subset(train, select = -c(lights) )
train_attributes <- subset(train, select = -c(logAppliances) )
test<-read.csv("../data/energy_data_test.csv")
test_attributes <- subset(test, select = -c(lights, logAppliances) )
logAppliances<-train$logAppliances
train <- cbind(train_attributes, logAppliances)

lasso regression - wszystkie kolumny, wyszukiwanie parametru lambda https://www.pluralsight.com/guides/linear-lasso-and-ridge-regression-with-r

library(glmnet)
x = as.matrix(train_attributes)
y_train = train$logAppliances
x_test = as.matrix(test_attributes)
y_test = test$logAppliances
lambdas <- 10^seq(0, -3, by = -.05)
cv_lasso <- cv.glmnet(x, y_train, alpha = 1, lambda = lambdas)
optimal_lambda <- cv_lasso$lambda.min
optimal_lambda
plot(lambdas,cv_lasso$cvm,ylab="Mean-Squared Error",xlab="Lambda",  type="l") 
lasso_reg = glmnet(x, y_train,  alpha = 1, family = 'gaussian', lambda = optimal_lambda,thresh = 1e-07)
summary(lasso_reg)
coef(lasso_reg)

Ocena wyników modelu

eval_results <- function(true, predicted, df) {
  SSE <- sum((predicted - true)^2)
  SST <- sum((true - mean(true))^2)
  R_square <- 1 - SSE / SST
  RMSE = sqrt(SSE/nrow(df))
return(data.frame(
  RMSE = RMSE,
  Rsquare = R_square
))}


# Prediction and evaluation on train data
predictions_train <- predict(lasso_reg, s = optimal_lambda, newx = x)
res<-eval_results(y_train, predictions_train, train)
print(res)

# Prediction and evaluation on test data
predictions_test <- predict(lasso_reg, s = optimal_lambda, newx = x_test)
res<-eval_results(y_test, predictions_test, test)
print(res)

thresh find

thresh_vals = list(10^(-10),10^(-9),10^(-8),10^(-7),10^(-6), 10^(-5), 10^(-4), 10^(-3), 10^(-2), 10^(-1), 10^(-0))
train_rsme=list()
test_rsme=list()
test_r2=list()
train_r2=list()
for(i in 1:length(thresh_vals)){
  lasso_reg = glmnet(x, y_train, nlambda = 25, alpha = 0, family = 'gaussian', lambda = optimal_lambda,thresh = thresh_vals[i])
# Prediction and evaluation on train data
  predictions_train <- predict(lasso_reg, s = optimal_lambda, newx = x)
  ret<-eval_results(y_train, predictions_train, train)
  train_r2=append(train_r2, ret$Rsquare)
  train_rsme=append(train_rsme, ret$RMSE)

# Prediction and evaluation on test data
  predictions_test <- predict(lasso_reg, s = optimal_lambda, newx = x_test)
  ret<-eval_results(y_test, predictions_test, test)
  test_r2=append(test_r2, ret$Rsquare)
  test_rsme=append(test_rsme, ret$RMSE)
}
plot(seq(-10,0,1) ,train_r2 ,pch=19, xlab="log10(thresh)", type="l")
plot(seq(-10,0,1) ,test_r2 ,pch=19, xlab="log10(thresh)", type="l")
plot(seq(-10,0,1) ,test_rsme ,pch=19, xlab="log10(thresh)", type="l")
plot(seq(-10,0,1) ,train_rsme ,pch=19, xlab="log10(thresh)", type="l")

Model bez RH_6, RH_7, RH_out

library(glmnet)

train_attributes <- subset(train, select = -c(logAppliances, RH_6, RH_7, RH_out) )
test<-read.csv("../data/energy_data_test.csv")
test_attributes <- subset(test, select = -c(lights, logAppliances,RH_6, RH_7, RH_out) )
x = as.matrix(train_attributes)
y_train = train$logAppliances
x_test = as.matrix(test_attributes)
y_test = test$logAppliances

lambdas <- 10^seq(0, -3, by = -.05)
cv_lasso <- cv.glmnet(x, y_train, alpha = 1, lambda = lambdas)
optimal_lambda <- cv_lasso$lambda.min
optimal_lambda
plot(lambdas,cv_lasso$cvm,ylab="Mean-Squared Error",xlab="Lambda",  type="l")
lasso_reg = glmnet(x, y_train, nlambda = 25, alpha = 1, family = 'gaussian', lambda = optimal_lambda,thresh = 1e-07)
summary(lasso_reg)
coef(lasso_reg)

# Prediction and evaluation on train data
predictions_train <- predict(lasso_reg, s = optimal_lambda, newx = x)
res<-eval_results(y_train, predictions_train, train)
print(res)

# Prediction and evaluation on test data
predictions_test <- predict(lasso_reg, s = optimal_lambda, newx = x_test)
res<-eval_results(y_test, predictions_test, test)
print(res)