绘制决策树分类器时出现交互错误，获取值数组.. 使树很难可视化

Posted 2023-03-12

【中文标题】绘制决策树分类器时出现交互错误，获取值数组.. 使树很难可视化

【英文标题】：Interactive error when plotting the decision tree classifier, get an array of values.. makes the tree very hard to visualize

【发布时间】：2021-08-04 09:20:00

【问题描述】：

这是重现决策树分类器树所需的代码，它给出了太多的值来解释图形，如果可能的话，我想避免使用更简单的值数组的公开值数组。在尝试绘制树之前处理数据集需要大部分代码。

import numpy as np
import pandas as pd
df = pd.read_csv("https://raw.githubusercontent.com/Joseph-Villegas/CST383/main/Building_and_Safety_Customer_Service_Request__Closed_.csv")

df.drop([ 'CSR Number', 'Address House Number', 'Address Street Direction', 
          'Address Street Name', 'Address Street Suffix',  
          'Parcel Identification Number (PIN)','Address House Fraction Number', 
          'Address Street Suffix Direction', 'Case Number Related to CSR'], 
           axis=1, inplace=True)


# Drop any row found with an NA value
df.dropna(axis=0, how='any', inplace=True)
# Observed date columns
date_columns = ['Date Received', 'Date Closed', 'Due Date']

# Function to reformat date string
str_2_date = lambda date: f"date[:6]2re.split('/', date)[2][1:]"

# Apply said function to the date columns in the dataframe
df = df.apply(lambda column: df[column.name].apply(str_2_date) if column.name in date_columns else column)



for column in date_columns:
  original_dtype = str(df[column].dtypes)
  df[column] = pd.to_datetime(df[column])
  new_dtype = str(df[column].dtypes)
  print(":<20 :<20 :<20".format(column, original_dtype, new_dtype))

for column in date_columns:
  df[f"column Day of Week"] = df[column].dt.dayofweek # Monday=0, Sunday=6.
  df[f"column Month"] = df[column].dt.month
  df[f"column Year"] = df[column].dt.year
# Remove original date columns
df.drop(date_columns, axis=1, inplace=True)
df['Lat.'] = [literal_eval(x)[0] for x in df['Latitude/Longitude']]
df['Lon.'] = [literal_eval(x)[1] for x in df['Latitude/Longitude']]
df.drop('Latitude/Longitude', axis=1, inplace=True)
# Encode the rest of the columns having dtype 'object' using ordinal encoding
object_columns = df.dtypes[(df.dtypes == "object")].index.tolist()
for column in object_columns:
  values_list = df[column].value_counts(ascending=True).index.tolist()
  ordinal_map = value:(index + 1) for index, value in enumerate(values_list)
  df[column] = df[column].map(ordinal_map)

def sincos(x, period):
  radians = (2 * np.pi * x) / period
  return np.column_stack((np.sin(radians), np.cos(radians)))
# Encode the day of week columns
day_of_week_columns = df.filter(like='Day of Week', axis=1).columns.tolist()
for column in day_of_week_columns:
  day_sc = sincos(df[column], 7)
  df[f"column Sin"] = day_sc[:,0]
  df[f"column Cosine"] = day_sc[:,1]

# Encode the month columns
month_columns = df.filter(like='Month', axis=1).columns.tolist()
for column in month_columns:
  month_sc = sincos(df[column], 12)
  df[f"column Sin"] = day_sc[:,0]
  df[f"column Cosine"] = day_sc[:,1]

date_info_columns = day_of_week_columns + month_columns
df.drop(date_info_columns, axis=1, inplace=True)


num_na = df.isna().sum().sum()
num_rows, num_cols = df.shape
# Below is the decision tree plot that gives unwanted array of values, is there a way to avoid this???
#-----------------------------------------------------------------------------------------
from sklearn.tree import export_graphviz 
import graphviz # needed for the graph
predictors = ['LADBS Inspection District', 'Address Street Zip', 'Date Received Year', 
              'Date Closed Year', 'Due Date Year', 'Case Flag', 'CSR Priority', 
              'Lat.', 'Lon.'] # features to predict from
# we must pass np arrays into our decision tree
X = df[predictors].values  # numpy array for predictor variables
y = df['Response Days'].values  # numpy array for target variable
X_train, X_test, y_train, y_test = train_test_split(X,y,test_size = 0.30, random_state=0)
clf = DecisionTreeClassifier(max_depth=3, random_state=0).fit(X_train, y_train) 
# Using Decision Tree classifier model, and fit the model with the training data

dot_data = export_graphviz(clf, precision=3, 
                    feature_names=predictors,  
                    proportion=True,
                    class_names=predictors,  
                    filled=False, rounded=True,  
                    special_characters=True)
# plot it
graph = graphviz.Source(dot_data)  
graph

【问题讨论】：

您能否更清楚地解释您的问题是什么？可视化一棵树到底有什么问题？

@pavel 阅读树非常困难，每个节点都会显示一组值，我想简化一下。

奇怪，通常export_graphviz对于每个节点只绘制节点的分裂特征名称、分裂值和杂质值。尝试设置proportion=False

@pavel 设置 proportion = False 不幸的是没有工作

【参考方案1】：

您的目标变量Response Days 具有许多唯一值，因此使用分类器意味着每个叶子都会跟踪每个叶子有多少样本，因此列表很长。您可能更愿意使用回归模型，如果这样做，每片叶子的报告值只是样本中的（单个！）平均目标值。