从Sparklyr中提取和可视化模型树

Question

从Sparklyr中提取和可视化模型树

Rea*_*chy 5 r decision-tree random-forest apache-spark sparklyr

有没有人有任何关于如何将树信息从sparklyr的ml_decision_tree_classifier,ml_gbt_classifier或ml_random_forest_classifier模型转换为.)格式的建议,这种格式可以被其他R树相关的库理解,并且(最终)b.)树的可视化用于非技术消费？这将包括从向量汇编器期间生成的替换字符串索引值转换回实际要素名称的能力.

为了提供一个例子,下面的代码从sparklyr博客文章中大量复制:

library(sparklyr)
library(dplyr)

# If needed, install Spark locally via `spark_install()`
sc <- spark_connect(master = "local")
iris_tbl <- copy_to(sc, iris)

# split the data into train and validation sets
iris_data <- iris_tbl %>%
  sdf_partition(train = 2/3, validation = 1/3, seed = 123)


iris_pipeline <- ml_pipeline(sc) %>%
  ft_dplyr_transformer(
    iris_data$train %>%
      mutate(Sepal_Length = log(Sepal_Length),
             Sepal_Width = Sepal_Width ^ 2)
  ) %>%
  ft_string_indexer("Species", "label")

iris_pipeline_model <- iris_pipeline %>%
  ml_fit(iris_data$train)

iris_vector_assembler <- ft_vector_assembler(
  sc, 
  input_cols = setdiff(colnames(iris_data$train), "Species"), 
  output_col = "features"
)
random_forest <- ml_random_forest_classifier(sc,features_col = "features")

# obtain the labels from the fitted StringIndexerModel
iris_labels <- iris_pipeline_model %>%
  ml_stage("string_indexer") %>%
  ml_labels()

# IndexToString will convert the predicted numeric values back to class labels
iris_index_to_string <- ft_index_to_string(sc, "prediction", "predicted_label", 
                                      labels = iris_labels)

# construct a pipeline with these stages
iris_prediction_pipeline <- ml_pipeline(
  iris_pipeline, # pipeline from previous section
  iris_vector_assembler, 
  random_forest,
  iris_index_to_string
)

# fit to data and make some predictions
iris_prediction_model <- iris_prediction_pipeline %>%
  ml_fit(iris_data$train)
iris_predictions <- iris_prediction_model %>%
  ml_transform(iris_data$validation)
iris_predictions %>%
  select(Species, label:predicted_label) %>%
  glimpse()

Run Code Online (Sandbox Code Playgroud)

在根据此处的建议进行反复试验之后,我能够以"if/else"格式打印出基础决策树的表达式,并将其转换为字符串:

model_stage <- iris_prediction_model$stages[[3]]

spark_jobj(model_stage) %>% invoke(., "toDebugString") %>% cat()
##print out below##
RandomForestClassificationModel (uid=random_forest_classifier_5c6a1934c8e) with 20 trees
  Tree 0 (weight 1.0):
    If (feature 2 <= 2.5)
     Predict: 1.0
    Else (feature 2 > 2.5)
     If (feature 2 <= 4.95)
      If (feature 3 <= 1.65)
       Predict: 0.0
      Else (feature 3 > 1.65)
       If (feature 0 <= 1.7833559100698644)
        Predict: 0.0
       Else (feature 0 > 1.7833559100698644)
        Predict: 2.0
     Else (feature 2 > 4.95)
      If (feature 2 <= 5.05)
       If (feature 1 <= 6.505000000000001)
        Predict: 2.0
       Else (feature 1 > 6.505000000000001)
        Predict: 0.0
      Else (feature 2 > 5.05)
       Predict: 2.0
  Tree 1 (weight 1.0):
    If (feature 3 <= 0.8)
     Predict: 1.0
    Else (feature 3 > 0.8)
     If (feature 3 <= 1.75)
      If (feature 1 <= 5.0649999999999995)
       If (feature 3 <= 1.05)
        Predict: 0.0
       Else (feature 3 > 1.05)
        If (feature 0 <= 1.8000241202036602)
         Predict: 2.0
        Else (feature 0 > 1.8000241202036602)
         Predict: 0.0
      Else (feature 1 > 5.0649999999999995)
       If (feature 0 <= 1.8000241202036602)
        Predict: 0.0
       Else (feature 0 > 1.8000241202036602)
        If (feature 2 <= 5.05)
         Predict: 0.0
        Else (feature 2 > 5.05)
         Predict: 2.0
     Else (feature 3 > 1.75)
      Predict: 2.0
  Tree 2 (weight 1.0):
    If (feature 3 <= 0.8)
     Predict: 1.0
    Else (feature 3 > 0.8)
     If (feature 0 <= 1.7664051342320237)
      Predict: 0.0
     Else (feature 0 > 1.7664051342320237)
      If (feature 3 <= 1.45)
       If (feature 2 <= 4.85)
        Predict: 0.0
       Else (feature 2 > 4.85)
        Predict: 2.0
      Else (feature 3 > 1.45)
       If (feature 3 <= 1.65)
        If (feature 1 <= 8.125)
         Predict: 2.0
        Else (feature 1 > 8.125)
         Predict: 0.0
       Else (feature 3 > 1.65)
        Predict: 2.0
  Tree 3 (weight 1.0):
    If (feature 0 <= 1.6675287895788053)
     If (feature 2 <= 2.5)
      Predict: 1.0
     Else (feature 2 > 2.5)
      Predict: 0.0
    Else (feature 0 > 1.6675287895788053)
     If (feature 3 <= 1.75)
      If (feature 3 <= 1.55)
       If (feature 1 <= 7.025)
        If (feature 2 <= 4.55)
         Predict: 0.0
        Else (feature 2 > 4.55)
         Predict: 2.0
       Else (feature 1 > 7.025)
        Predict: 0.0
      Else (feature 3 > 1.55)
       If (feature 2 <= 5.05)
        Predict: 0.0
       Else (feature 2 > 5.05)
        Predict: 2.0
     Else (feature 3 > 1.75)
      Predict: 2.0
  Tree 4 (weight 1.0):
    If (feature 2 <= 4.85)
     If (feature 2 <= 2.5)
      Predict: 1.0
     Else (feature 2 > 2.5)
      Predict: 0.0
    Else (feature 2 > 4.85)
     If (feature 2 <= 5.05)
      If (feature 0 <= 1.8484238118815566)
       Predict: 2.0
      Else (feature 0 > 1.8484238118815566)
       Predict: 0.0
     Else (feature 2 > 5.05)
      Predict: 2.0
  Tree 5 (weight 1.0):
    If (feature 2 <= 1.65)
     Predict: 1.0
    Else (feature 2 > 1.65)
     If (feature 3 <= 1.65)
      If (feature 0 <= 1.8325494627242664)
       Predict: 0.0
      Else (feature 0 > 1.8325494627242664)
       If (feature 2 <= 4.95)
        Predict: 0.0
       Else (feature 2 > 4.95)
        Predict: 2.0
     Else (feature 3 > 1.65)
      Predict: 2.0
  Tree 6 (weight 1.0):
    If (feature 2 <= 2.5)
     Predict: 1.0
    Else (feature 2 > 2.5)
     If (feature 2 <= 5.05)
      If (feature 3 <= 1.75)
       Predict: 0.0
      Else (feature 3 > 1.75)
       Predict: 2.0
     Else (feature 2 > 5.05)
      Predict: 2.0
  Tree 7 (weight 1.0):
    If (feature 3 <= 0.55)
     Predict: 1.0
    Else (feature 3 > 0.55)
     If (feature 3 <= 1.65)
      If (feature 2 <= 4.75)
       Predict: 0.0
      Else (feature 2 > 4.75)
       Predict: 2.0
     Else (feature 3 > 1.65)
      If (feature 2 <= 4.85)
       If (feature 0 <= 1.7833559100698644)
        Predict: 0.0
       Else (feature 0 > 1.7833559100698644)
        Predict: 2.0
      Else (feature 2 > 4.85)
       Predict: 2.0
  Tree 8 (weight 1.0):
    If (feature 3 <= 0.8)
     Predict: 1.0
    Else (feature 3 > 0.8)
     If (feature 3 <= 1.85)
      If (feature 2 <= 4.85)
       Predict: 0.0
      Else (feature 2 > 4.85)
       If (feature 0 <= 1.8794359129669855)
        Predict: 2.0
       Else (feature 0 > 1.8794359129669855)
        If (feature 3 <= 1.55)
         Predict: 0.0
        Else (feature 3 > 1.55)
         Predict: 0.0
     Else (feature 3 > 1.85)
      Predict: 2.0
  Tree 9 (weight 1.0):
    If (feature 2 <= 2.5)
     Predict: 1.0
    Else (feature 2 > 2.5)
     If (feature 2 <= 4.95)
      Predict: 0.0
     Else (feature 2 > 4.95)
      Predict: 2.0
  Tree 10 (weight 1.0):
    If (feature 3 <= 0.8)
     Predict: 1.0
    Else (feature 3 > 0.8)
     If (feature 2 <= 4.95)
      Predict: 0.0
     Else (feature 2 > 4.95)
      If (feature 2 <= 5.05)
       If (feature 3 <= 1.55)
        Predict: 2.0
       Else (feature 3 > 1.55)
        If (feature 3 <= 1.75)
         Predict: 0.0
        Else (feature 3 > 1.75)
         Predict: 2.0
      Else (feature 2 > 5.05)
       Predict: 2.0
  Tree 11 (weight 1.0):
    If (feature 3 <= 0.8)
     Predict: 1.0
    Else (feature 3 > 0.8)
     If (feature 2 <= 5.05)
      If (feature 2 <= 4.75)
       Predict: 0.0
      Else (feature 2 > 4.75)
       If (feature 3 <= 1.75)
        Predict: 0.0
       Else (feature 3 > 1.75)
        Predict: 2.0
     Else (feature 2 > 5.05)
      Predict: 2.0
  Tree 12 (weight 1.0):
    If (feature 3 <= 0.8)
     Predict: 1.0
    Else (feature 3 > 0.8)
     If (feature 3 <= 1.75)
      If (feature 3 <= 1.35)
       Predict: 0.0
      Else (feature 3 > 1.35)
       If (feature 0 <= 1.695573522904327)
        Predict: 0.0
       Else (feature 0 > 1.695573522904327)
        If (feature 1 <= 8.125)
         Predict: 2.0
        Else (feature 1 > 8.125)
         Predict: 0.0
     Else (feature 3 > 1.75)
      If (feature 0 <= 1.7833559100698644)
       Predict: 0.0
      Else (feature 0 > 1.7833559100698644)
       Predict: 2.0
  Tree 13 (weight 1.0):
    If (feature 3 <= 0.55)
     Predict: 1.0
    Else (feature 3 > 0.55)
     If (feature 2 <= 4.95)
      If (feature 2 <= 4.75)
       Predict: 0.0
      Else (feature 2 > 4.75)
       If (feature 0 <= 1.8000241202036602)
        If (feature 1 <= 9.305)
         Predict: 2.0
        Else (feature 1 > 9.305)
         Predict: 0.0
       Else (feature 0 > 1.8000241202036602)
        Predict: 0.0
     Else (feature 2 > 4.95)
      Predict: 2.0
  Tree 14 (weight 1.0):
    If (feature 2 <= 2.5)
     Predict: 1.0
    Else (feature 2 > 2.5)
     If (feature 3 <= 1.65)
      If (feature 3 <= 1.45)
       Predict: 0.0
      Else (feature 3 > 1.45)
       If (feature 2 <= 4.95)
        Predict: 0.0
       Else (feature 2 > 4.95)
        Predict: 2.0
     Else (feature 3 > 1.65)
      If (feature 0 <= 1.7833559100698644)
       If (feature 0 <= 1.7664051342320237)
        Predict: 2.0
       Else (feature 0 > 1.7664051342320237)
        Predict: 0.0
      Else (feature 0 > 1.7833559100698644)
       Predict: 2.0
  Tree 15 (weight 1.0):
    If (feature 2 <= 2.5)
     Predict: 1.0
    Else (feature 2 > 2.5)
     If (feature 3 <= 1.75)
      If (feature 2 <= 4.95)
       Predict: 0.0
      Else (feature 2 > 4.95)
       If (feature 1 <= 8.125)
        Predict: 2.0
       Else (feature 1 > 8.125)
        If (feature 0 <= 1.9095150692894909)
         Predict: 0.0
        Else (feature 0 > 1.9095150692894909)
         Predict: 2.0
     Else (feature 3 > 1.75)
      Predict: 2.0
  Tree 16 (weight 1.0):
    If (feature 3 <= 0.8)
     Predict: 1.0
    Else (feature 3 > 0.8)
     If (feature 0 <= 1.7491620461964392)
      Predict: 0.0
     Else (feature 0 > 1.7491620461964392)
      If (feature 3 <= 1.75)
       If (feature 2 <= 4.75)
        Predict: 0.0
       Else (feature 2 > 4.75)
        If (feature 0 <= 1.8164190316151556)
         Predict: 2.0
        Else (feature 0 > 1.8164190316151556)
         Predict: 0.0
      Else (feature 3 > 1.75)
       Predict: 2.0
  Tree 17 (weight 1.0):
    If (feature 0 <= 1.695573522904327)
     If (feature 2 <= 1.65)
      Predict: 1.0
     Else (feature 2 > 1.65)
      Predict: 0.0
    Else (feature 0 > 1.695573522904327)
     If (feature 2 <= 4.75)
      If (feature 2 <= 2.5)
       Predict: 1.0
      Else (feature 2 > 2.5)
       Predict: 0.0
     Else (feature 2 > 4.75)
      If (feature 3 <= 1.75)
       If (feature 1 <= 5.0649999999999995)
        Predict: 2.0
       Else (feature 1 > 5.0649999999999995)
        If (feature 3 <= 1.65)
         Predict: 0.0
        Else (feature 3 > 1.65)
         Predict: 0.0
      Else (feature 3 > 1.75)
       Predict: 2.0
  Tree 18 (weight 1.0):
    If (feature 3 <= 0.8)
     Predict: 1.0
    Else (feature 3 > 0.8)
     If (feature 3 <= 1.65)
      Predict: 0.0
     Else (feature 3 > 1.65)
      If (feature 0 <= 1.7833559100698644)
       Predict: 0.0
      Else (feature 0 > 1.7833559100698644)
       Predict: 2.0
  Tree 19 (weight 1.0):
    If (feature 2 <= 2.5)
     Predict: 1.0
    Else (feature 2 > 2.5)
     If (feature 2 <= 4.95)
      If (feature 1 <= 8.705)
       Predict: 0.0
      Else (feature 1 > 8.705)
       If (feature 2 <= 4.85)
        Predict: 0.0
       Else (feature 2 > 4.85)
        If (feature 0 <= 1.8164190316151556)
         Predict: 2.0
        Else (feature 0 > 1.8164190316151556)
         Predict: 0.0
     Else (feature 2 > 4.95)
      Predict: 2.0

Run Code Online (Sandbox Code Playgroud)

正如您所看到的,这种格式不是最佳的,可以用于我所见过的许多可视化决策树图形的漂亮方法之一(例如革命分析或statmethods)

Answer 1

zer*_*323 8

截至今天（Spark 2.4.0 版本已经批准并等待官方公告），您最好的选择*，在不涉及复杂的第 3 方工具（例如您可以看看 MLeap）的情况下，可能是保存模型并读回规格：

ml_stage(iris_prediction_model, "random_forest") %>% 
  ml_save("/tmp/model")

rf_spec <- spark_read_parquet(sc, "rf", "/tmp/model/data/")

Run Code Online (Sandbox Code Playgroud)

结果将是DataFrame具有以下架构的 Spark：

rf_spec %>% 
  spark_dataframe() %>% 
  invoke("schema") %>% invoke("treeString") %>% 
  cat(sep = "\n")

Run Code Online (Sandbox Code Playgroud)

ml_stage(iris_prediction_model, "random_forest") %>% 
  ml_save("/tmp/model")

rf_spec <- spark_read_parquet(sc, "rf", "/tmp/model/data/")

Run Code Online (Sandbox Code Playgroud)

提供有关所有节点和分片的信息。

可以使用列元数据检索特征映射：

meta <- iris_predictions %>% 
    select(features) %>% 
    spark_dataframe() %>% 
    invoke("schema") %>% invoke("apply", 0L) %>% 
    invoke("metadata") %>% 
    invoke("getMetadata", "ml_attr") %>% 
    invoke("getMetadata", "attrs") %>% 
    invoke("json") %>%
    jsonlite::fromJSON() %>% 
    dplyr::bind_rows() %>% 
    copy_to(sc, .) %>%
    rename(featureIndex = idx)

meta

Run Code Online (Sandbox Code Playgroud)

rf_spec %>% 
  spark_dataframe() %>% 
  invoke("schema") %>% invoke("treeString") %>% 
  cat(sep = "\n")

Run Code Online (Sandbox Code Playgroud)

以及您已经检索到的标签映射：

labels <- tibble(prediction = seq_along(iris_labels) - 1, label = iris_labels) %>%
  copy_to(sc, .)

Run Code Online (Sandbox Code Playgroud)

最后你可以结合所有这些：

full_rf_spec <- rf_spec %>% 
  spark_dataframe() %>% 
  invoke("selectExpr", list("treeID", "nodeData.*", "nodeData.split.*")) %>% 
  sdf_register() %>% 
  select(-split, -impurityStats) %>% 
  left_join(meta, by = "featureIndex") %>% 
  left_join(labels, by = "prediction")

full_rf_spec

Run Code Online (Sandbox Code Playgroud)

root
 |-- treeID: integer (nullable = true)
 |-- nodeData: struct (nullable = true)
 |    |-- id: integer (nullable = true)
 |    |-- prediction: double (nullable = true)
 |    |-- impurity: double (nullable = true)
 |    |-- impurityStats: array (nullable = true)
 |    |    |-- element: double (containsNull = true)
 |    |-- gain: double (nullable = true)
 |    |-- leftChild: integer (nullable = true)
 |    |-- rightChild: integer (nullable = true)
 |    |-- split: struct (nullable = true)
 |    |    |-- featureIndex: integer (nullable = true)
 |    |    |-- leftCategoriesOrThreshold: array (nullable = true)
 |    |    |    |-- element: double (containsNull = true)
 |    |    |-- numCategories: integer (nullable = true)

Run Code Online (Sandbox Code Playgroud)

由收集和分隔treeID，应该提供足够的信息**来模拟树状对象（您可以通过检查rpart::rpart.object文档和/或unclass模型来很好地理解所需的结构rpart。tree::tree需要较少的工作，但其绘图实用程序是远非令人印象深刻），并构建了一个像样的情节。

另一种方法是使用Sparklyr2PMML将数据导出到 PMML并使用此表示形式。

您还可以查看如何在 Apache Spark (PySpark 1.4.1) 中可视化/绘制决策树？这建议第三方Python包来解决同样的问题。

如果您不需要任何花哨的东西，您可以使用以下命令创建一个粗略的绘图igraph：

library(igraph)

gframe <- full_rf_spec %>% 
  filter(treeID == 0) %>%   # Take the first tree
  mutate(
    leftCategoriesOrThreshold = ifelse(
      size(leftCategoriesOrThreshold) == 1,
      # Continuous variable case
      concat("<= ", round(concat_ws("", leftCategoriesOrThreshold), 3)),
      # Categorical variable case. Decoding variables might be involved
      # but can be achieved if needed, using column metadata or indexer labels
      concat("in {", concat_ws(",", leftCategoriesOrThreshold), "}")
    ),
    name = coalesce(name, label)) %>% 
 select(
   id, label, impurity, gain, 
   leftChild, rightChild, leftCategoriesOrThreshold, name) %>%
 collect()

vertices <- gframe %>% rename(label = name, name = id)

edges <- gframe %>%
  transmute(from = id, to = leftChild, label = leftCategoriesOrThreshold) %>% 
  union_all(gframe %>% select(from = id, to = rightChild)) %>% 
  filter(to != -1)

g <- igraph::graph_from_data_frame(edges, vertices = vertices)

plot(
  g, layout = layout_as_tree(g, root = c(1)),
  vertex.shape = "rectangle",  vertex.size = 45)

Run Code Online (Sandbox Code Playgroud)

* 它应该在不久的将来得到改进，通过新引入的与格式无关的 ML 编写器 API（它已经支持选定模型的 PMML 编写器。希望新的模型和格式将随之而来）。

** 如果您使用分类特征，您可能希望映射leftCategoriesOrThreshold到相应的索引级别。

如果特征向量包含分类变量，则的输出jsonlite::fromJSON()将包含nominal组。例如，如果您有foo三个级别的索引列，在第一个位置组装，它将是这样的：

$nominal
     vals idx      name
1 a, b, c   1       foo

Run Code Online (Sandbox Code Playgroud)

其中valscolumn是可变长度向量的列表。

length(meta$nominal$vals[[1]])

Run Code Online (Sandbox Code Playgroud)

[1] 3

Run Code Online (Sandbox Code Playgroud)

标签对应于该结构的索引，因此在示例中：

a有标签0.0（不是说标签是双精度浮点数，编号从0.0开始）
b有标签 1.0

依此类推，如果您使用leftCategoriesOrThresholdequal to 进行 split ，那么c(0.0, 2.0)这意味着 split 在 labels 上{"a", "c"}。

另请注意，如果存在分类数据，您可能必须在调用之前对其进行处理copy_to- 目前看来它不支持复杂字段。

在 Spark <= 2.3 中，您将必须使用 R 代码进行映射（在本地结构上，有些purrr应该可以很好）。在 Spark 2.4 中（据sparklyr我所知尚不支持），使用 Spark 的 JSON 读取器直接读取元数据并使用其高阶函数进行映射可能会更容易。

归档时间：	7 年，10 月前
查看次数：	482 次
最近记录：	7 年，10 月前