我一直在尝试使用LASSO进行无监督特征选择(通过删除类列)。数据集包括分类(因子)和连续(数值)变量。这里是链接。我使用
如果您检查
有没有办法在
感谢您提供的任何帮助。提前致谢。
model.matrix()
构建了一个设计矩阵,为每个分类变量的级别创建虚拟变量。dataset <- read.xlsx("./hepatitis.data.xlsx", sheet = "hepatitis", na.strings = "")
names_df <- names(dataset)
formula_LASSO <- as.formula(paste("~ 0 +", paste(names_df, collapse = " + ")))
LASSO_df <- model.matrix(object = formula_LASSO, data = dataset, contrasts.arg = lapply(dataset[ ,sapply(dataset, is.factor)], contrasts, contrasts = FALSE ))
### Group LASSO using gglasso package
gglasso_group <- c(1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 15, 16, 17, 17)
fit <- gglasso(x = LASSO_df, y = y_k, group = gglasso_group, loss = "ls", intercept = FALSE, nlambda = 100)
# Cross validation
fit.cv <- cv.gglasso(x = LASSO_df, y = y_k, group = gglasso_group, nfolds = 10)
# Best lambda
best_lambda_fit.cv <- fit.cv$lambda.1se
# Final coefficients of variables
coefs = coef.gglasso(object = fit, s = best_lambda_fit.cv)
### Group LASSO with grpreg package
group_lasso <- grpreg(X = LASSO_df, y = y_k, group = gglasso_group, penalty = "grLasso")
plot(group_lasso)
cv_group_lasso <- cv.grpreg(X = LASSO_df, y = y_k, group = gglasso_group, penalty = "grLasso", se = "quick")
# Best lambda
best_lambda_group_lasso <- cv_group_lasso$lambda.min
coef_mat_group_lasso <- as.matrix(coef(cv_group_lasso))
如果您检查
coefs
和 coef_mat_group_lasso
,您会发现它们并不相同。此外,最佳的 lambda 值也不相同。我不确定哪一个适合用于特征选择。有没有办法在
grpreg()
函数中去除截距? intercept = FALSE
没有生效。感谢您提供的任何帮助。提前致谢。