r - 如何使用 MLR3（glmnet 学习器）访问和比较 LASSO 模型系数？

Question

目标

使用 MLR3 创建 LASSO 模型
使用带有内部 CV 或引导程序的嵌套 CV进行超参数 (lambda) 确定，使用外部 CV 进行模型性能评估（而不是只进行一次测试训练），并在不同模型实例中找到不同 LASSO 回归系数的标准偏差。
对尚不可用的测试数据集进行预测。

问题

我不确定在下面的代码中是否正确实现了所描述的嵌套 CV 方法。
我不确定 alpha 是否设置正确 alpha = 1 only。
在 mlr3 中使用重采样时，我不知道如何访问 LASSO lamda 系数。（mlr3learners 中的importance() 尚不支持LASSO）
我不知道如何将可能的模型应用于 mlr3 中不可用的测试集。

代码

library(readr)
library(mlr3)
library(mlr3learners)
library(mlr3pipelines)
library(reprex)

# Data ------
# Prepared according to the Blog post by Julia Silge
# https://juliasilge.com/blog/lasso-the-office/
urlfile = 'https://raw.githubusercontent.com/shudras/office_data/master/office_data.csv'
data = read_csv(url(urlfile))[-1]
#> Warning: Missing column names filled in: 'X1' [1]
#> Parsed with column specification:
#> cols(
#>   .default = col_double()
#> )
#> See spec(...) for full column specifications.

# Add a factor to data
data$factor = as.factor(c(rep('a', 20), rep('b', 50), rep('c', 30), rep('a', 6), rep('c', 10), rep('b', 20)))

# Task creation
task = 
  TaskRegr$new(
    id = 'office', 
    backend = data, 
    target = 'imdb_rating'
  )

# Model creation
graph = 
  po('scale') %>>% 
  po('encode') %>>% # make factors numeric
  # How to normalize predictors, leaving target unchanged?
  lrn('regr.cv_glmnet', # 10-fold CV for inner loop. Is alpha permanently set to 1?
       id = 'rp', alpha = 1, family = 'gaussian'
     ) 

graph_learner = GraphLearner$new(graph)

# Execution (actual modeling)
result = 
  resample(
    task, 
    graph_learner, 
    rsmp('cv', folds = 5) # 5-fold for outer CV 
  )
#> INFO  [13:21:53.485] Applying learner 'scale.encode.regr.cv_glmnet' on task 'office' (iter 3/5) 
#> INFO  [13:21:54.937] Applying learner 'scale.encode.regr.cv_glmnet' on task 'office' (iter 2/5) 
#> INFO  [13:21:55.242] Applying learner 'scale.encode.regr.cv_glmnet' on task 'office' (iter 1/5) 
#> INFO  [13:21:55.500] Applying learner 'scale.encode.regr.cv_glmnet' on task 'office' (iter 4/5) 
#> INFO  [13:21:55.831] Applying learner 'scale.encode.regr.cv_glmnet' on task 'office' (iter 5/5)

# How to access results, i.e. lamda coefficients, 
# and compare them (why no variable importance for glmnet)

# Access prediction
result$prediction()
#> <PredictionRegr> for 136 observations:
#>     row_id truth response
#>          2   8.3 8.373798
#>          6   8.7 8.455151
#>          9   8.4 8.358964
#> ---                      
#>        116   9.7 8.457607
#>        119   8.2 8.130352
#>        128   7.8 8.224150

^{由reprex 包（v0.3.0）于 2020 年 6 月 11 日创建}

编辑 1（LASSO 系数）

根据 misuse LASSO 系数的评论，可以通过result$data$learner[[1]]$model$rp$model$glmnet.fit$beta另外，我发现store_models = TRUE需要在结果中设置以存储模型并依次访问系数。

尽管设置了 alpha = 1，我还是选择了多个 LASSO 系数。我想要“最佳”LASSO 系数（例如源自 lamda = lamda.min 或 lamda.1se）。不同的 s1, s2, s3, ... 是什么意思？这些是不同的lamdas吗？
不同的系数确实似乎源于不同的 lambda 值，表示为 s1, s2 , s3, ... （数字是索引。）我想，可以通过首先找到“最佳”lambda 的索引来访问“最佳”系数index_lamda.1se = which(ft$lambda == ft$lambda.1se)[[1]]; index_lamda.min = which(ft$lambda == ft$lambda.min)[[1]]然后找到系数集。错误使用的评论中给出了一种更简洁的方法来找到“最佳”系数。

library(readr)
library(mlr3)
library(mlr3learners)
library(mlr3pipelines)
library(reprex)
urlfile = 'https://raw.githubusercontent.com/shudras/office_data/master/office_data.csv'
data = read_csv(url(urlfile))[-1]

# Add a factor to data
data$factor = as.factor(c(rep('a', 20), rep('b', 50), rep('c', 30), rep('a', 6), rep('c', 10), rep('b', 20)))

# Task creation
task = 
  TaskRegr$new(
    id = 'office', 
    backend = data, 
    target = 'imdb_rating'
  )

# Model creation
graph = 
  po('scale') %>>% 
  po('encode') %>>% # make factors numeric
  # How to normalize predictors, leaving target unchanged?
  lrn('regr.cv_glmnet', # 10-fold CV for inner loop. Is alpha permanently set to 1?
      id = 'rp', alpha = 1, family = 'gaussian'
  ) 

graph$keep_results = TRUE
graph_learner = GraphLearner$new(graph)

# Execution (actual modeling)
result = 
  resample(
    task, 
    graph_learner, 
    rsmp('cv', folds = 5), # 5-fold for outer CV 
    store_models = TRUE # Store model needed to acces coefficients
  )

# LASSO coefficients 
# Why more than one coefficient per predictor? 
# What are s1, s2 etc.? Shouldn't 'lrn' fix alpha = 1?
# How to obtain the best coefficient (for lamda 1se or min) if multiple?
as.matrix(result$data$learner[[1]]$model$rp$model$glmnet.fit$beta)
#>                  s0          s1         s2         s3          s4         s5
#> andy              0 0.000000000 0.00000000 0.00000000 0.000000000 0.00000000
#> angela            0 0.000000000 0.00000000 0.00000000 0.000000000 0.00000000
#> b_j_novak         0 0.000000000 0.00000000 0.00000000 0.000000000 0.00000000
#> brent_forrester   0 0.000000000 0.00000000 0.00000000 0.000000000 0.00000000
#> darryl            0 0.000000000 0.00000000 0.00000000 0.000000000 0.00000000
#> dwight            0 0.000000000 0.00000000 0.00000000 0.000000000 0.00000000
#> episode           0 0.000000000 0.00000000 0.00000000 0.010297763 0.02170423
#> erin              0 0.000000000 0.00000000 0.00000000 0.000000000 0.00000000
#> gene_stupnitsky   0 0.000000000 0.00000000 0.00000000 0.000000000 0.00000000
#> greg_daniels      0 0.000000000 0.00000000 0.00000000 0.001845101 0.01309437
#> jan               0 0.000000000 0.00000000 0.00000000 0.005663699 0.01357832
#> jeffrey_blitz     0 0.000000000 0.00000000 0.00000000 0.000000000 0.00000000
#> jennifer_celotta  0 0.000000000 0.00000000 0.00000000 0.000000000 0.00000000
#> jim               0 0.006331732 0.01761548 0.02789682 0.036853510 0.04590513
#> justin_spitzer    0 0.000000000 0.00000000 0.00000000 0.000000000 0.00000000
#> [...]
#>                           s6           s7           s8          s9         s10
#> andy             0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> angela           0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> b_j_novak        0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> brent_forrester  0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> darryl           0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> dwight           0.002554576  0.007006995  0.011336058  0.01526851  0.01887180
#> episode          0.031963475  0.040864492  0.047487987  0.05356482  0.05910066
#> erin             0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> gene_stupnitsky  0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> greg_daniels     0.023040791  0.031866343  0.040170917  0.04779004  0.05472702
#> jan              0.021030152  0.028094541  0.035062678  0.04143812  0.04725379
#> jeffrey_blitz    0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> jennifer_celotta 0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> jim              0.053013058  0.058503984  0.062897112  0.06683734  0.07041964
#> justin_spitzer   0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> kelly            0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> ken_kwapis       0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> kevin            0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> lee_eisenberg    0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> michael          0.057190859  0.062963830  0.068766981  0.07394472  0.07865977
#> mindy_kaling     0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> oscar            0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> pam              0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> paul_feig        0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> paul_lieberstein 0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> phyllis          0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> randall_einhorn  0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> ryan             0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> season           0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> toby             0.000000000  0.000000000  0.005637169  0.01202893  0.01785309
#> factor.a         0.000000000 -0.003390125 -0.022365768 -0.03947047 -0.05505681
#> factor.b         0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#> factor.c         0.000000000  0.000000000  0.000000000  0.00000000  0.00000000
#>                           s11          s12          s13           s14
#> andy              0.000000000  0.000000000  0.000000000  0.0000000000
#> angela            0.000000000  0.000000000  0.000000000  0.0000000000
#> b_j_novak         0.000000000  0.000000000  0.000000000  0.0000000000
#> brent_forrester   0.000000000  0.000000000  0.000000000  0.0000000000
#> darryl            0.000000000  0.000000000  0.000000000  0.0017042281
#> dwight            0.022170870  0.025326337  0.027880703  0.0303865693
#> episode           0.064126846  0.069018240  0.074399623  0.0794693480
#> [...]

^{由reprex 包（v0.3.0）于 2020 年 6 月 15 日创建}

编辑 2（可选的后续问题）

嵌套 CV 提供了多个模型之间的差异评估。差异可以表示为外部 CV 获得的误差（例如 RMSE）。虽然该误差可能很小，但来自模型（由外部 CV 实例化）的单个 LASSO 系数（预测变量的重要性）可能会有很大差异。

mlr3 是否提供描述预测变量定量重要性一致性的功能，即外部 CV 创建的模型中 LASSO 系数的 RMSE？还是应该创建一个自定义函数，使用result$data$learner[[i]]$model$rp$model$glmnet.fit$beta（由误用建议）检索 LASSO 系数，其中 i = 1、2、3、4、5 是外部 CV 的折叠，然后取匹配系数的 RMSE？

r - 如何使用 MLR3（glmnet 学习器）访问和比较 LASSO 模型系数？

目标

问题

代码

编辑 1（LASSO 系数）

编辑 2（可选的后续问题）

0 回答 0

Related

Reference