获取更多R语言知识，请关注公众号：医学和生信笔记

医学和生信笔记 公众号主要分享：1.医学小知识、肛肠科小知识；2.R语言和Python相关的数据分析、可视化、机器学习等；3.生物信息学学习资料和自己的学习笔记！

mlr3pipelines是一种数据流编程套件，完整的机器学习工作流可被称为Graph/Pipelines，包含数据预处理、建模、多个模型比较等，不同的模型需要不同的数据预处理方法，另外还有集成学习、各种非线性模型等，这些都可以通过mlr3pipelines解决。

数据预处理的R包有很多，比如caret、recipes等，mlr3pipelines创造性的使用了图流的方式。
R语言机器学习mlr3：数据预处理和Pipelines - 图1

pipeops

进行数据预处理的各种方法在mlr3pipelines中被称为pipeops，目前基本涵盖常见的数据预处理方法，比如独热编码、稀疏矩阵、缺失值处理、降维、数据标准化、因子分组等等。

可以用来连接预处理和模型，或者构建复杂的统计建模步骤，例如多种不同的预处理连接多种不同的模型等

查看所有的pipeops

  1. library(mlr3pipelines)
  2. as.data.table(mlr_pipeops) # 目前共有64种
  3. ##                       key                         packages
  4. ##  1:                boxcox      mlr3pipelines,bestNormalize
  5. ##  2:                branch                    mlr3pipelines
  6. ##  3:                 chunk                    mlr3pipelines
  7. ##  4:        classbalancing                    mlr3pipelines
  8. ##  5:            classifavg              mlr3pipelines,stats
  9. ##  6:          classweights                    mlr3pipelines
  10. ##  7:              colapply                    mlr3pipelines
  11. ##  8:       collapsefactors                    mlr3pipelines
  12. ##  9:              colroles                    mlr3pipelines
  13. ## 10:                  copy                    mlr3pipelines
  14. ## 11:          datefeatures                    mlr3pipelines
  15. ## 12:                encode              mlr3pipelines,stats
  16. ## 13:          encodeimpact                    mlr3pipelines
  17. ## 14:            encodelmer        mlr3pipelines,lme4,nloptr
  18. ## 15:          featureunion                    mlr3pipelines
  19. ## 16:                filter                    mlr3pipelines
  20. ## 17:            fixfactors                    mlr3pipelines
  21. ## 18:               histbin           mlr3pipelines,graphics
  22. ## 19:                   ica            mlr3pipelines,fastICA
  23. ## 20:        imputeconstant                    mlr3pipelines
  24. ## 21:            imputehist           mlr3pipelines,graphics
  25. ## 22:         imputelearner                    mlr3pipelines
  26. ## 23:            imputemean                    mlr3pipelines
  27. ## 24:          imputemedian              mlr3pipelines,stats
  28. ## 25:            imputemode                    mlr3pipelines
  29. ## 26:             imputeoor                    mlr3pipelines
  30. ## 27:          imputesample                    mlr3pipelines
  31. ## 28:             kernelpca            mlr3pipelines,kernlab
  32. ## 29:               learner                    mlr3pipelines
  33. ## 30:            learner_cv                    mlr3pipelines
  34. ## 31:               missind                    mlr3pipelines
  35. ## 32:           modelmatrix              mlr3pipelines,stats
  36. ## 33:     multiplicityexply                    mlr3pipelines
  37. ## 34:     multiplicityimply                    mlr3pipelines
  38. ## 35:                mutate                    mlr3pipelines
  39. ## 36:                   nmf           mlr3pipelines,MASS,NMF
  40. ## 37:                   nop                    mlr3pipelines
  41. ## 38:              ovrsplit                    mlr3pipelines
  42. ## 39:              ovrunite                    mlr3pipelines
  43. ## 40:                   pca                    mlr3pipelines
  44. ## 41:                 proxy                    mlr3pipelines
  45. ## 42:           quantilebin              mlr3pipelines,stats
  46. ## 43:      randomprojection                    mlr3pipelines
  47. ## 44:        randomresponse                    mlr3pipelines
  48. ## 45:               regravg                    mlr3pipelines
  49. ## 46:       removeconstants                    mlr3pipelines
  50. ## 47:         renamecolumns                    mlr3pipelines
  51. ## 48:             replicate                    mlr3pipelines
  52. ## 49:                 scale                    mlr3pipelines
  53. ## 50:           scalemaxabs                    mlr3pipelines
  54. ## 51:            scalerange                    mlr3pipelines
  55. ## 52:                select                    mlr3pipelines
  56. ## 53:                 smote        mlr3pipelines,smotefamily
  57. ## 54:           spatialsign                    mlr3pipelines
  58. ## 55:             subsample                    mlr3pipelines
  59. ## 56:          targetinvert                    mlr3pipelines
  60. ## 57:          targetmutate                    mlr3pipelines
  61. ## 58: targettrafoscalerange                    mlr3pipelines
  62. ## 59:        textvectorizer mlr3pipelines,quanteda,stopwords
  63. ## 60:             threshold                    mlr3pipelines
  64. ## 61:         tunethreshold              mlr3pipelines,bbotk
  65. ## 62:              unbranch                    mlr3pipelines
  66. ## 63:                vtreat             mlr3pipelines,vtreat
  67. ## 64:            yeojohnson      mlr3pipelines,bestNormalize
  68. ##                       key                         packages
  69. ##                                 tags
  70. ##  1:                   data transform
  71. ##  2:                             meta
  72. ##  3:                             meta
  73. ##  4:   imbalanced data,data transform
  74. ##  5:                         ensemble
  75. ##  6:   imbalanced data,data transform
  76. ##  7:                   data transform
  77. ##  8:                   data transform
  78. ##  9:                   data transform
  79. ## 10:                             meta
  80. ## 11:                   data transform
  81. ## 12:            encode,data transform
  82. ## 13:            encode,data transform
  83. ## 14:            encode,data transform
  84. ## 15:                         ensemble
  85. ## 16: feature selection,data transform
  86. ## 17:         robustify,data transform
  87. ## 18:                   data transform
  88. ## 19:                   data transform
  89. ## 20:                         missings
  90. ## 21:                         missings
  91. ## 22:                         missings
  92. ## 23:                         missings
  93. ## 24:                         missings
  94. ## 25:                         missings
  95. ## 26:                         missings
  96. ## 27:                         missings
  97. ## 28:                   data transform
  98. ## 29:                          learner
  99. ## 30:  learner,ensemble,data transform
  100. ## 31:          missings,data transform
  101. ## 32:                   data transform
  102. ## 33:                     multiplicity
  103. ## 34:                     multiplicity
  104. ## 35:                   data transform
  105. ## 36:                   data transform
  106. ## 37:                             meta
  107. ## 38:    target transform,multiplicity
  108. ## 39:            multiplicity,ensemble
  109. ## 40:                   data transform
  110. ## 41:                             meta
  111. ## 42:                   data transform
  112. ## 43:                   data transform
  113. ## 44:                         abstract
  114. ## 45:                         ensemble
  115. ## 46:         robustify,data transform
  116. ## 47:                   data transform
  117. ## 48:                     multiplicity
  118. ## 49:                   data transform
  119. ## 50:                   data transform
  120. ## 51:                   data transform
  121. ## 52: feature selection,data transform
  122. ## 53:   imbalanced data,data transform
  123. ## 54:                   data transform
  124. ## 55:                   data transform
  125. ## 56:                         abstract
  126. ## 57:                 target transform
  127. ## 58:                 target transform
  128. ## 59:                   data transform
  129. ## 60:                 target transform
  130. ## 61:                 target transform
  131. ## 62:                             meta
  132. ## 63:   encode,missings,data transform
  133. ## 64:                   data transform
  134. ##                                 tags
  135. ##                                            feature_types input.num output.num
  136. ##  1:                                      numeric,integer         1          1
  137. ##  2:                                                   NA         1         NA
  138. ##  3:                                                   NA         1         NA
  139. ##  4: logical,integer,numeric,character,factor,ordered,...         1          1
  140. ##  5:                                                   NA        NA          1
  141. ##  6: logical,integer,numeric,character,factor,ordered,...         1          1
  142. ##  7: logical,integer,numeric,character,factor,ordered,...         1          1
  143. ##  8:                                       factor,ordered         1          1
  144. ##  9: logical,integer,numeric,character,factor,ordered,...         1          1
  145. ## 10:                                                   NA         1         NA
  146. ## 11:                                              POSIXct         1          1
  147. ## 12:                                       factor,ordered         1          1
  148. ## 13:                                       factor,ordered         1          1
  149. ## 14:                                       factor,ordered         1          1
  150. ## 15:                                                   NA        NA          1
  151. ## 16: logical,integer,numeric,character,factor,ordered,...         1          1
  152. ## 17:                                       factor,ordered         1          1
  153. ## 18:                                      numeric,integer         1          1
  154. ## 19:                                      numeric,integer         1          1
  155. ## 20: logical,integer,numeric,character,factor,ordered,...         1          1
  156. ## 21:                                      integer,numeric         1          1
  157. ## 22:                               logical,factor,ordered         1          1
  158. ## 23:                                      numeric,integer         1          1
  159. ## 24:                                      numeric,integer         1          1
  160. ## 25:               factor,integer,logical,numeric,ordered         1          1
  161. ## 26:             character,factor,integer,numeric,ordered         1          1
  162. ## 27:               factor,integer,logical,numeric,ordered         1          1
  163. ## 28:                                      numeric,integer         1          1
  164. ## 29:                                                   NA         1          1
  165. ## 30: logical,integer,numeric,character,factor,ordered,...         1          1
  166. ## 31: logical,integer,numeric,character,factor,ordered,...         1          1
  167. ## 32: logical,integer,numeric,character,factor,ordered,...         1          1
  168. ## 33:                                                   NA         1         NA
  169. ## 34:                                                   NA        NA          1
  170. ## 35: logical,integer,numeric,character,factor,ordered,...         1          1
  171. ## 36:                                      numeric,integer         1          1
  172. ## 37:                                                   NA         1          1
  173. ## 38:                                                   NA         1          1
  174. ## 39:                                                   NA         1          1
  175. ## 40:                                      numeric,integer         1          1
  176. ## 41:                                                   NA        NA          1
  177. ## 42:                                      numeric,integer         1          1
  178. ## 43:                                      numeric,integer         1          1
  179. ## 44:                                                   NA         1          1
  180. ## 45:                                                   NA        NA          1
  181. ## 46: logical,integer,numeric,character,factor,ordered,...         1          1
  182. ## 47: logical,integer,numeric,character,factor,ordered,...         1          1
  183. ## 48:                                                   NA         1          1
  184. ## 49:                                      numeric,integer         1          1
  185. ## 50:                                      numeric,integer         1          1
  186. ## 51:                                      numeric,integer         1          1
  187. ## 52: logical,integer,numeric,character,factor,ordered,...         1          1
  188. ## 53: logical,integer,numeric,character,factor,ordered,...         1          1
  189. ## 54:                                      numeric,integer         1          1
  190. ## 55: logical,integer,numeric,character,factor,ordered,...         1          1
  191. ## 56:                                                   NA         2          1
  192. ## 57:                                                   NA         1          2
  193. ## 58:                                                   NA         1          2
  194. ## 59:                                            character         1          1
  195. ## 60:                                                   NA         1          1
  196. ## 61:                                                   NA         1          1
  197. ## 62:                                                   NA        NA          1
  198. ## 63: logical,integer,numeric,character,factor,ordered,...         1          1
  199. ## 64:                                      numeric,integer         1          1
  200. ##                                            feature_types input.num output.num
  201. ##     input.type.train  input.type.predict output.type.train output.type.predict
  202. ##  1:             Task                Task              Task                Task
  203. ##  2:                *                   *                 *                   *
  204. ##  3:             Task                Task              Task                Task
  205. ##  4:      TaskClassif         TaskClassif       TaskClassif         TaskClassif
  206. ##  5:             NULL   PredictionClassif              NULL   PredictionClassif
  207. ##  6:      TaskClassif         TaskClassif       TaskClassif         TaskClassif
  208. ##  7:             Task                Task              Task                Task
  209. ##  8:             Task                Task              Task                Task
  210. ##  9:             Task                Task              Task                Task
  211. ## 10:                *                   *                 *                   *
  212. ## 11:             Task                Task              Task                Task
  213. ## 12:             Task                Task              Task                Task
  214. ## 13:             Task                Task              Task                Task
  215. ## 14:             Task                Task              Task                Task
  216. ## 15:             Task                Task              Task                Task
  217. ## 16:             Task                Task              Task                Task
  218. ## 17:             Task                Task              Task                Task
  219. ## 18:             Task                Task              Task                Task
  220. ## 19:             Task                Task              Task                Task
  221. ## 20:             Task                Task              Task                Task
  222. ## 21:             Task                Task              Task                Task
  223. ## 22:             Task                Task              Task                Task
  224. ## 23:             Task                Task              Task                Task
  225. ## 24:             Task                Task              Task                Task
  226. ## 25:             Task                Task              Task                Task
  227. ## 26:             Task                Task              Task                Task
  228. ## 27:             Task                Task              Task                Task
  229. ## 28:             Task                Task              Task                Task
  230. ## 29:      TaskClassif         TaskClassif              NULL   PredictionClassif
  231. ## 30:      TaskClassif         TaskClassif       TaskClassif         TaskClassif
  232. ## 31:             Task                Task              Task                Task
  233. ## 32:             Task                Task              Task                Task
  234. ## 33:              [*]                 [*]                 *                   *
  235. ## 34:                *                   *               [*]                 [*]
  236. ## 35:             Task                Task              Task                Task
  237. ## 36:             Task                Task              Task                Task
  238. ## 37:                *                   *                 *                   *
  239. ## 38:      TaskClassif         TaskClassif     [TaskClassif]       [TaskClassif]
  240. ## 39:           [NULL] [PredictionClassif]              NULL   PredictionClassif
  241. ## 40:             Task                Task              Task                Task
  242. ## 41:                *                   *                 *                   *
  243. ## 42:             Task                Task              Task                Task
  244. ## 43:             Task                Task              Task                Task
  245. ## 44:             NULL          Prediction              NULL          Prediction
  246. ## 45:             NULL      PredictionRegr              NULL      PredictionRegr
  247. ## 46:             Task                Task              Task                Task
  248. ## 47:             Task                Task              Task                Task
  249. ## 48:                *                   *               [*]                 [*]
  250. ## 49:             Task                Task              Task                Task
  251. ## 50:             Task                Task              Task                Task
  252. ## 51:             Task                Task              Task                Task
  253. ## 52:             Task                Task              Task                Task
  254. ## 53:             Task                Task              Task                Task
  255. ## 54:             Task                Task              Task                Task
  256. ## 55:             Task                Task              Task                Task
  257. ## 56:        NULL,NULL function,Prediction              NULL          Prediction
  258. ## 57:             Task                Task         NULL,Task       function,Task
  259. ## 58:         TaskRegr            TaskRegr     NULL,TaskRegr   function,TaskRegr
  260. ## 59:             Task                Task              Task                Task
  261. ## 60:             NULL   PredictionClassif              NULL   PredictionClassif
  262. ## 61:             Task                Task              NULL          Prediction
  263. ## 62:                *                   *                 *                   *
  264. ## 63:             Task                Task              Task                Task
  265. ## 64:             Task                Task              Task                Task
  266. ##     input.type.train  input.type.predict output.type.train output.type.predict

看到有很多数据预处理方法了，但其实常用的也就10来种左右。

创建预处理步骤可通过以下方法：

  1. pca <- mlr_pipeops$get("pca")
  3. # 或者用简便写法
  4. pca <- po("pca")

非常重要的一点是，不仅能创建预处理步骤，也可以用这种方法选择算法，选择特征选择方法等：

  1. # 选择学习器/算法
  2. library(mlr3)
  3. learner <- po("learner" ,lrn("classif.rpart"))
  5. # 选择特征选择的方法并设置参数
  6. filter <- po("filter",
  7.              filter = mlr3filters::flt("variance"),
  8.              filter.frac = 0.5
  9.              )

mlr3pipelines中的管道符: %>>%

这是mlr3团队发明的专用管道符，可用于连接不同的预处理步骤、预处理和模型等操作：

  1. gr <- po("scale") %>>% po("pca")
  2. gr$plot(html = F)

R语言机器学习mlr3：数据预处理和Pipelines - 图2

很多强大的操作都是基于此管道符运行的。

建立模型

一个简单的例子，先预处理数据，再训练

  1. # 连接预处理和模型，有点类似tidymodels的workflow
  2. mutate <- po("mutate")
  3. filter <- po("filter",
  4.              filter = mlr3filters::flt("variance"),
  5.              param_vals = list(filter.frac = 0.5))
  7. graph <- mutate %>>%
  8.   filter %>>%
  9.   po("learner", learner = lrn("classif.rpart"))

现在这个graph就变成了一个含有预处理步骤的学习器（learner），可以像前面介绍的那样直接用于训练、预测：

  1. task <- tsk("iris")
  2. graph$train(task)
  3. ## $classif.rpart.output
  4. ## NULL

预测

  1. graph$predict(task)
  2. ## $classif.rpart.output
  3. ## <PredictionClassif> for 150 observations:
  4. ##     row_ids     truth  response
  5. ##           1    setosa    setosa
  6. ##           2    setosa    setosa
  7. ##           3    setosa    setosa
  8. ## ---                            
  9. ##         148 virginica virginica
  10. ##         149 virginica virginica
  11. ##         150 virginica virginica

除此之外，还可以把graph变成一个GraphLearner对象，用于resample和benchmark等

  1. glrn <- as_learner(graph) # 变成graphlearner
  3. cv3 <- rsmp("cv", folds = 5)
  4. resample(task, glrn, cv3)
  5. ## INFO  [21:01:53.145] [mlr3] Applying learner 'mutate.variance.classif.rpart' on task 'iris' (iter 2/5) 
  6. ## INFO  [21:01:53.230] [mlr3] Applying learner 'mutate.variance.classif.rpart' on task 'iris' (iter 5/5) 
  7. ## INFO  [21:01:53.291] [mlr3] Applying learner 'mutate.variance.classif.rpart' on task 'iris' (iter 1/5) 
  8. ## INFO  [21:01:53.350] [mlr3] Applying learner 'mutate.variance.classif.rpart' on task 'iris' (iter 4/5) 
  9. ## INFO  [21:01:53.407] [mlr3] Applying learner 'mutate.variance.classif.rpart' on task 'iris' (iter 3/5)
  10. ## <ResampleResult> of 5 iterations
  11. ## * Task: iris
  12. ## * Learner: mutate.variance.classif.rpart
  13. ## * Warnings: 0 in 0 iterations
  14. ## * Errors: 0 in 0 iterations

在很多数据预处理步骤中也是有参数需要调整的，mlr3pipelines不仅可以用于调整算法的超参数，还可以调整预处理中的参数。

  1. library(paradox)
  3. ps <- ps(
  4.   classif.rpart.cp = p_dbl(0,0.05), # 算法中的参数
  5.   variance.filter.frac = p_dbl(0.25,1) # 特征选择方法中的参数
  6. )
  8. library(mlr3tuning)
  9. instance <- TuningInstanceSingleCrit$new(
  10.   task = task,
  11.   learner = glrn,
  12.   resampling = rsmp("holdout", ratio = 0.7),
  13.   measure = msr("classif.acc"),
  14.   search_space = ps,
  15.   terminator = trm("evals", n_evals = 20)
  16. )
  18. tuner <- tnr("random_search")
  20. lgr::get_logger("mlr3")$set_threshold("warn")
  21. lgr::get_logger("bbotk")$set_threshold("warn")
  23. tuner$optimize(instance)
  24. ##    classif.rpart.cp variance.filter.frac learner_param_vals  x_domain
  25. ## 1:       0.02162802            0.3852356          <list[5]> <list[2]>
  26. ##    classif.acc
  27. ## 1:   0.9777778
  29. instance$result_y
  30. ## classif.acc 
  31. ##   0.9777778
  32. instance$result_learner_param_vals
  33. ## $mutate.mutation
  34. ## list()
  35. ## 
  36. ## $mutate.delete_originals
  37. ## [1] FALSE
  38. ## 
  39. ## $variance.filter.frac
  40. ## [1] 0.3852356
  41. ## 
  42. ## $classif.rpart.xval
  43. ## [1] 0
  44. ## 
  45. ## $classif.rpart.cp
  46. ## [1] 0.02162802

可以看到结果直接给出了算法的超参数和特征选择中的参数。

非线性graph

  • Branching: 一个点通往多个分支，例如在比较多个特征选择方法时很有用。只有一条路会被执行。
  • Copying: 一个点通往多个分支，所有的分支都会被执行，但是只能1次执行1个分支，并行计算目前还不支持。
  • Stacking: 单个图被彼此堆叠，一个图的输出是另一个图的输入。

branching & copying

使用PipeOpBranch和PipeOpUnbranch实现分支操作，分支操作的概念如下图所示：
R语言机器学习mlr3：数据预处理和Pipelines - 图3

下面一个例子演示了分支操作，分支之后一定要解除分支：

  1. graph <- po("branch", c("nop","pca","scale")) %>>% # 开始分支
  2.   gunion(list(
  3.     po("nop", id = "null1"), # 分支1，并且取了个名字null1
  4.     po("pca"),               # 分支2
  5.     po("scale")              # 分支3
  6.   )) %>>%
  7.   po("unbranch",c("nop","pca","scale")) # 结束分支
  9. graph$plot(html = F)

R语言机器学习mlr3：数据预处理和Pipelines - 图4

bagging

属于集成学习的一种，概念不做介绍，感兴趣的可自行学习，其概念可查看下图：
R语言机器学习mlr3：数据预处理和Pipelines - 图5

下面演示基本使用方法。

  1. single_pred <- po("subsample", frac = 0.7) %>>%
  2.   po("learner", lrn("classif.rpart")) # 建立一个模型
  4. pred_set <- ppl("greplicate", single_pred, 10L) # 复制10次
  6. bagging <- pred_set %>>%
  7.   po("classifavg", innum = 10)
  9. bagging$plot(html = FALSE)

R语言机器学习mlr3：数据预处理和Pipelines - 图6

把上面的对象变成一个GraphLearner，然后就可以进行训练和预测了：

  1. task <- tsk("iris")
  2. split <- partition(task, ratio = 0.7, stratify = T)
  5. baglrn <- as_learner(bagging)
  6. baglrn$train(task, row_ids = split$train)
  7. baglrn$predict(task, row_ids = split$test)
  8. ## <PredictionClassif> for 45 observations:
  9. ##     row_ids     truth  response prob.setosa prob.versicolor prob.virginica
  10. ##           4    setosa    setosa           1               0              0
  11. ##           6    setosa    setosa           1               0              0
  12. ##           8    setosa    setosa           1               0              0
  13. ## ---                                                                       
  14. ##         141 virginica virginica           0               0              1
  15. ##         147 virginica virginica           0               0              1
  16. ##         150 virginica virginica           0               0              1

stacking

另一种提高模型性能的方法，概念可看下图：
R语言机器学习mlr3：数据预处理和Pipelines - 图7

这里为了防止过拟合，使用PipeOpLearnerCV预测袋外数据，它可以在数据内部自动执行嵌套重抽样。

首先创建level 0学习器，然后复制一份，并取一个名字：

  1. lrn <- lrn("classif.rpart")
  2. lrn_0 <- po("learner_cv", lrn$clone())
  3. lrn_0$id<- "rpart_cv"

然后联合使用gunion和PipeOpNOP，把没动过的task传到下一个level，这样经过决策树的task和没处理过的task就能一起传到下一个level了。

  1. level_0 <- gunion(list(lrn_0, po("nop")))

把上面传下来的东西联合到一起：

  1. combined <- level_0 %>>% po("featureunion", 2)
  2. stack <- combined %>>% po("learner", lrn$clone())
  3. stack$plot(html = FALSE)

R语言机器学习mlr3：数据预处理和Pipelines - 图8

然后就可以进行训练、预测了：

  1. stacklrn <- as_learner(stack)
  2. stacklrn$train(task, split$train)
  3. stacklrn$predict(task, split$test)
  4. ## <PredictionClassif> for 45 observations:
  5. ##     row_ids     truth  response
  6. ##           4    setosa    setosa
  7. ##           6    setosa    setosa
  8. ##           8    setosa    setosa
  9. ## ---                            
  10. ##         141 virginica virginica
  11. ##         147 virginica virginica
  12. ##         150 virginica virginica

一个超级复杂的例子

这个例子有多个不同的预处理步骤，使用多个不同的算法。

  1. library("magrittr")
  2. library("mlr3learners") 
  4. rprt = lrn("classif.rpart", predict_type = "prob")
  5. glmn = lrn("classif.glmnet", predict_type = "prob")
  7. # 创建学习器
  8. lrn_0 = po("learner_cv", rprt, id = "rpart_cv_1")
  9. lrn_0$param_set$values$maxdepth = 5L
  10. lrn_1 = po("pca", id = "pca1") %>>% po("learner_cv", rprt, id = "rpart_cv_2")
  11. lrn_1$param_set$values$rpart_cv_2.maxdepth = 1L
  12. lrn_2 = po("pca", id = "pca2") %>>% po("learner_cv", glmn)
  14. # 第0层
  15. level_0 = gunion(list(lrn_0, lrn_1, lrn_2, po("nop", id = "NOP1")))
  17. # 第1层
  18. level_1 = level_0 %>>%
  19.   po("featureunion", 4) %>>%
  20.   po("copy", 3) %>>%
  21.   gunion(list(
  22.     po("learner_cv", rprt, id = "rpart_cv_l1"),
  23.     po("learner_cv", glmn, id = "glmnt_cv_l1"),
  24.     po("nop", id = "NOP_l1")
  25.   ))
  27. # 第2层
  28. level_2 = level_1 %>>%
  29.   po("featureunion", 3, id = "u2") %>>%
  30.   po("learner", rprt, id = "rpart_l2")
  33. level_2$plot(html = FALSE)

R语言机器学习mlr3：数据预处理和Pipelines - 图9

下面就可以进行训练、预测：

  1. task = tsk("iris")
  2. lrn = as_learner(level_2)
  4. lrn$
  5.   train(task, split$train)$
  6.   predict(task, split$test)$
  7.   score()
  8. ## classif.ce 
  9. ## 0.08888889

一些特殊预处理步骤

其实是一些很常用的步骤…

缺失值处理：PipeOpImpute

缺失值处理实在是太常见了，mlr3pipelines对于数值型和因子型都能处理。

  1. pom <- po("missind")
  2. pon <- po("imputehist", # 条形图插补数值型 
  3.           id  = "impute_num", # 取个名字
  4.           affect_columns  = is.numeric # 设置处理哪些列
  5.           )
  7. pof = po("imputeoor", id = "imputer_fct", affect_columns = is.factor) # 处理因子
  9. imputer = pom %>>% pon %>>% pof

连接学习器：

  1. polrn <- po("learner", lrn("classif.rpart"))
  2. lrn <- as_learner(imputer %>>% polrn)

创建新的变量：PipeOpMutate

  1. pom <- po("mutate",
  2.           mutation = list(
  3.             Sepal.Sum = ~ Sepal.Length + Sepal.Width,
  4.             Petal.Sum = ~ Petal.Length + Petal.Width,
  5.             Sepal.Petal.Ratio = ~ (Sepal.Length / Petal.Length)
  6.             )
  7.           )

使用子集训练：PipeOpChunk

有时候数据集太大，把数据分割成小块进行分块训练是很好的办法。

  1. chks = po("chunk", 4)
  2. lrns = ppl("greplicate", po("learner", lrn("classif.rpart")), 4)
  4. mjv = po("classifavg", 4)
  6. pipeline = chks %>>% lrns %>>% mjv
  7. pipeline$plot(html = FALSE)

R语言机器学习mlr3：数据预处理和Pipelines - 图10

  1. task = tsk("iris")
  2. train.idx = sample(seq_len(task$nrow), 120)
  3. test.idx = setdiff(seq_len(task$nrow), train.idx)
  5. pipelrn = as_learner(pipeline)
  6. pipelrn$train(task, train.idx)$
  7.   predict(task, train.idx)$
  8.   score()
  9. ## classif.ce 
  10. ##  0.3333333

特征选择：PipeOpFilter和PipeOpSelect

可以使用PipeOpFilter对象把mlr3filters里面的变量选择方法放进mlr3pipelines中。

  1. po("filter", mlr3filters::flt("information_gain"))
  2. ## PipeOp: <information_gain> (not trained)
  3. ## values: <list()>
  4. ## Input channels <name [train type, predict type]>:
  5. ##   input [Task,Task]
  6. ## Output channels <name [train type, predict type]>:
  7. ##   output [Task,Task]

可使用filter_nfeat/filter_frac/filter_cutoff决定保留哪些变量/特征。

获取更多R语言知识，请关注公众号：医学和生信笔记

医学和生信笔记 公众号主要分享：1.医学小知识、肛肠科小知识；2.R语言和Python相关的数据分析、可视化、机器学习等；3.生物信息学学习资料和自己的学习笔记！