数据链接
一、数据预处理# 导入包
import os
import time
from pyspark.sql import SparkSession# 实例化SparkSession对象,以本地模式是运行Spark程序
spark = SparkSession \.builder \.appName("PySpark_ML_Pipeline") \.master("local[4]")\.getOrCreate()print spark
print spark.sparkContext
'''
'''
help(spark.read.csv)
# 读取数据集,
raw_df = spark.read.csv('./datas/train.tsv', header='true', sep='\t',\inferSchema='true')
# 显示条目数
print raw_df.count()
==>7395
raw_df.printSchema()# 由于字段太多,选择某些字段值
raw_df.select('url', 'alchemy_category', 'alchemy_category_score', \'label').show(10)
# 定义函数转换 ?转换为 0
def replace_question_func(x):return '0' if x == '?' else x# 注册函数
from pyspark.sql.functions import udf
replace_question = udf(replace_question_func)# col函数将 一个字符串转换为DataFrame中列, 获取对应DataFrame中此列的值
from pyspark.sql.functions import col# 使用自定义的函数,转换数据
df = raw_df.select(['url', 'alchemy_category'] +\[ replace_question(col(column)).cast('double')\.alias(column) for column in raw_df.columns[4:]])df.printSchema()df.select('url', 'alchemy_category', 'alchemy_category_score', \'label').show(10)
# 将数据集分为 训练集和测试集
train_df, test_df = df.randomSplit([0.7, 0.3])print train_df.cache().count()
print test_df.cache().count()
"""
5216
2179
"""
1、alchemy_category类别特征数据转换第一特征转换器、StringIndexer将文字的类别特征 转换 数字第二特征转换器、OneHotEncoder将数值的 类别特征字段 转换为 多个字段的Vector
2、特征的组合第二特征转换器、VectorAssembler将多个特征整合到一起
网址:http://spark.apache.org/docs/2.2.0/ml-features.html#stringindexer
# 导入模块
from pyspark.ml.feature import StringIndexer
help(StringIndexer)# 创建StringIndexer实例对象
"""参数说明:inputCol -> 要转换的字段名称outputCol -> 转换后的字段名称
"""
categoryIndexer = StringIndexer(inputCol='alchemy_category',\outputCol='alchemy_category_index')print type(categoryIndexer)
"""
==>
"""
调用StringIndexer类中的 fit 方法,获取到转换器Transformer
categoryTransformer = categoryIndexer.fit(df)
print type(categoryTransformer)# 使用 categoryTransformer 转换器 将所有的 train_df 进行转换
df1 = categoryTransformer.transform(train_df)df1.select('alchemy_category', 'alchemy_category_index').show(10)
"""
+------------------+----------------------+
| alchemy_category|alchemy_category_index|
+------------------+----------------------+
| ?| 0.0|
|arts_entertainment| 2.0|
| ?| 0.0|
| business| 3.0|
|arts_entertainment| 2.0|
| ?| 0.0|
| ?| 0.0|
| recreation| 1.0|
| business| 3.0|
|arts_entertainment| 2.0|
+------------------+----------------------+
only showing top 10 rows
"""df1.printSchema() #查看结构数据
OneHotEncoder可以将一个数值的类别特征字段转换为多个字段的Vector向量
from pyspark.ml.feature import OneHotEncoder
# 创建 OneHotEncoder 实例对象
encoder = OneHotEncoder(inputCol='alchemy_category_index', outputCol='alchemy_category_index_vector')print type(encoder)
"""
"""df2 = encoder.transform(df1)df2.printSchema()df2.select('alchemy_category', 'alchemy_category_index',\'alchemy_category_index_vector').show(10)
特征的组合
~~~~~~~~ 第二特征转换器、VectorAssembler,将多个特征整合到一起
from pyspark.ml.feature import VectorAssembler
assembler_inputs = ['alchemy_category_index_vector'] \+ raw_df.columns[4:-1]
print assembler_inputs"""
['alchemy_category_index_vector', 'alchemy_category_score',
'avglinksize', 'commonlinkratio_1', 'commonlinkratio_2',
'commonlinkratio_3', 'commonlinkratio_4', 'compression_ratio','embed_ratio', 'framebased', 'frameTagRatio', 'hasDomainLink',
'linkwordscore', 'news_front_page', 'non_markup_alphanum_characters',
'numberOfLinks', 'numwords_in_url', 'parametrizedLinkRatio',
'spelling_errors_ratio']
"""
######创建 VectorAssembler 实例对象,传递参数,指定合并哪些字段,输出的字段名称
assembler = VectorAssembler(inputCols=assembler_inputs, outputCol='features')
df3 = assembler.transform(df2)df3.printSchema()"""
+--------------------+-----+
| features|label|
+--------------------+-----+
|(35,[0,14,15,16,1...| 1.0|
|(35,[2,13,14,15,1...| 1.0|
|(35,[0,14,15,19,2...| 0.0|
|(35,[3,13,14,15,1...| 1.0|
|(35,[2,13,14,15,1...| 0.0|
+--------------------+-----+
only showing top 5 rows
"""df3.select('features').take(1)
"""
[Row(features=SparseVector(35,
{0: 1.0, 14: 2.1446, 15: 0.7969, 16: 0.3945, 17: 0.332,
18: 0.3203, 19: 0.5022, 22: 0.028, 24: 0.1898, 25: 0.2354,26: 1.0, 27: 1.0, 28: 17.0, 30: 10588.0, 31: 256.0, 32: 5.0, 33: 0.3828, 34: 0.1368}))]
"""
二、建模
from pyspark.ml.classification import DecisionTreeClassifier# 使用决策树分类算法
dtc = DecisionTreeClassifier(featuresCol='features', labelCol='label',impurity='gini', maxDepth=5, maxBins=32)# 将 训练数据 应用到 算法
dtc_model = dtc.fit(df3)# 使用模型预测
df4 = dtc_model.transform(df3)
df4.select('label', 'prediction', 'rawPrediction', 'probability').show(20, truncate=False)
label | prediction | rawPrediction | probability |
---|---|---|---|
1.0 | 1.0 | [909.0,1104.0] | [0.45156482861400893,0.5484351713859911] |
1.0 | 1.0 | [909.0,1104.0] | [0.45156482861400893,0.5484351713859911] |
0.0 | 0.0 | [38.0,1.0] | [0.9743589743589743,0.02564102564102564] |
1.0 | 1.0 | [27.0,177.0] | [0.1323529411764706,0.8676470588235294] |
0.0 | 0.0 | [95.0,28.0] | [0.7723577235772358,0.22764227642276422] |
1.0 | 1.0 | [909.0,1104.0] | [0.45156482861400893,0.5484351713859911] |
1.0 | 1.0 | [909.0,1104.0] | [0.45156482861400893,0.5484351713859911] |
1.0 | 0.0 | [144.0,95.0] | [0.602510460251046,0.39748953974895396] |
0.0 | 0.0 | [363.0,146.0] | [0.7131630648330058,0.2868369351669941] |
0.0 | 0.0 | [86.0,23.0] | [0.7889908256880734,0.21100917431192662] |
0.0 | 0.0 | [144.0,95.0] | [0.602510460251046,0.39748953974895396] |
0.0 | 0.0 | [144.0,95.0] | [0.602510460251046,0.39748953974895396] |
0.0 | 0.0 | [43.0,1.0] | [0.9772727272727273,0.022727272727272728] |
1.0 | 1.0 | [909.0,1104.0] | [0.45156482861400893,0.5484351713859911] |
1.0 | 1.0 | [909.0,1104.0] | [0.45156482861400893,0.5484351713859911] |
1.0 | 1.0 | [27.0,177.0] | [0.1323529411764706,0.8676470588235294] |
1.0 | 1.0 | [129.0,417.0] | [0.23626373626373626,0.7637362637362637] |
1.0 | 1.0 | [909.0,1104.0] | [0.45156482861400893,0.5484351713859911] |
0.0 | 1.0 | [909.0,1104.0] | [0.45156482861400893,0.5484351713859911] |
1.0 | 1.0 | [909.0,1104.0] | [0.45156482861400893,0.5484351713859911] |
only showing top 20 rows
三、评估(ROC曲线)from pyspark.ml.evaluation import BinaryClassificationEvaluator
# 创建 实例对象, 传递参数值
evaluator = BinaryClassificationEvaluator(labelCol='label',rawPredictionCol='rawPrediction')
# 计算指标 metricName="areaUnderROC"
auc = evaluator.evaluate(df4)
print auc
"""
0.6087142511
"""
总结上述开发流程:1、从原始数据 提取特征数据2、特征数据应用到算法,得到模型3、使用模型预测数据4、评估模型Pipeline:相当于一个“算法” -> 模型学习器包含两部分内容;-a. Estimator 模型学习器fit()-b. transformers 转换器transformer()
pipeline = Pipeline(Stages(.....))pipeline.fit().....
model.transfor().....
四、打包(ML Pipeline)
# 1. 导入全部需要 模块
from pyspark.ml import Pipeline
from pyspark.ml.feature import StringIndexer, OneHotEncoder, VectorAssembler
from pyspark.ml.classification import DecisionTreeClassifier
# a. StringIndexer
string_indexer = StringIndexer(inputCol='alchemy_category',\outputCol='alchemy_category_index')# b. OneHotEncoding
one_hot_encoder = OneHotEncoder(inputCol='alchemy_category_index',\outputCol='alchemy_category_index_vector')# c. VectorAessmbler
assembler_inputs = ['alchemy_category_index_vector'] \+ raw_df.columns[4:-1]
vector_assembler = VectorAssembler(inputCols=assembler_inputs,\outputCol='features')# d. DecisionTreeClassifier 模型学习器
dt = DecisionTreeClassifier(featuresCol='features', labelCol='label',\impurity='gini', maxDepth=5, maxBins=32)
# 按照数据处理顺序
pipeline = Pipeline(stages=[string_indexer,one_hot_encoder, vector_assembler, dt])
pipeline.getStages()"""
[StringIndexer_43e8b50676a58dad4d05,OneHotEncoder_4bf2a31a6b4b12aebd78,VectorAssembler_4429bf16ed1cc6c14207,DecisionTreeClassifier_451682088ef8fcaa79ae]"""
# 调用fit方法学,
pipleline_model = pipeline.fit(train_df)type(pipleline_model) #pyspark.ml.pipeline.PipelineModel
pipleline_model.stages[3]
predict_df = pipleline_model.transform(test_df)
# 保存 模型
pipleline_model.save('./datas/dtc-model')
# 加载模型
from pyspark.ml.pipeline import PipelineModelload_pipeline_model = PipelineModel.load('./datas/dtc-model')
load_pipeline_model.stages[3]# 预测
load_pipeline_model.transform(test_df) \.select('label', 'prediction', 'rawPrediction',\'probability').show(20, truncate=False)
label | prediction | rawPrediction | probability |
---|---|---|---|
0.0 | 0.0 | [361.0,300.0] | [0.546142208774584,0.45385779122541603] |
1.0 | 0.0 | [144.0,95.0] | [0.602510460251046,0.39748953974895396] |
0.0 | 1.0 | [0.0,8.0] | [0.0,1.0] |
1.0 | 1.0 | [129.0,417.0] | [0.23626373626373626,0.7637362637362637] |
0.0 | 0.0 | [363.0,146.0] | [0.7131630648330058,0.2868369351669941] |
0.0 | 0.0 | [363.0,146.0] | [0.7131630648330058,0.2868369351669941] |
1.0 | 1.0 | [909.0,1104.0] | [0.45156482861400893,0.5484351713859911] |
1.0 | 1.0 | [129.0,417.0] | [0.23626373626373626,0.7637362637362637] |
1.0 | 1.0 | [27.0,177.0] | [0.1323529411764706,0.8676470588235294] |
1.0 | 1.0 | [27.0,177.0] | [0.1323529411764706,0.8676470588235294] |
1.0 | 1.0 | [27.0,177.0] | [0.1323529411764706,0.8676470588235294] |
1.0 | 1.0 | [27.0,177.0] | [0.1323529411764706,0.8676470588235294] |
1.0 | 1.0 | [27.0,177.0] | [0.1323529411764706,0.8676470588235294] |
1.0 | 1.0 | [909.0,1104.0] | [0.45156482861400893,0.5484351713859911] |
0.0 | 0.0 | [363.0,146.0] | [0.7131630648330058,0.2868369351669941] |
1.0 | 1.0 | [909.0,1104.0] | [0.45156482861400893,0.5484351713859911] |
1.0 | 1.0 | [909.0,1104.0] | [0.45156482861400893,0.5484351713859911] |
1.0 | 1.0 | [909.0,1104.0] | [0.45156482861400893,0.5484351713859911] |
1.0 | 0.0 | [361.0,300.0] | [0.546142208774584,0.45385779122541603] |
0.0 | 0.0 | [86.0,23.0] | [0.7889908256880734,0.21100917431192662] |
only showing top 20 rows
五、验证选择最优模型(训练检验分离选择最优)
导入模块
from pyspark.ml.tuning import TrainValidationSplit, ParamGridBuilder
构建一个 决策树分类算法 网格参数
"""调整三个参数:-1. 不纯度度量-2. 最多深度-3. 最大分支数
"""
param_grid = ParamGridBuilder() \.addGrid(dt.impurity, ['gini', 'entropy']) \.addGrid(dt.maxDepth, [5, 10, 20]) \.addGrid(dt.maxBins, [8, 16, 32]) \.build()print type(param_grid)
for param in param_grid:print param
针对二分类创建模型评估器
binary_class_evaluator = BinaryClassificationEvaluator(labelCol='label',\rawPredictionCol='rawPrediction')
创建 TrainValidationSplit 实例对象
"""__init__(self, estimator=None, estimatorParamMaps=None, evaluator=None, trainRatio=0.75, seed=None)参数解释:estimator:模型学习器,针对哪个算法进行调整超参数,这里是DTestimatorParamMaps:算法调整的参数组合evaluator:评估模型的评估器,比如二分类的话,使用auc面积trainRatio:训练集与验证集 所占的比例,此处的值表示的是 训练集比例
"""train_validataion_split = TrainValidationSplit(estimator=dt,evaluator=binary_class_evaluator, estimatorParamMaps=param_grid, trainRatio=0.8)type(train_validataion_split)
#pyspark.ml.tuning.TrainValidationSplit
建立新的Pipeline实例对象
#使用 train_validataion_split 代替 原先 dt
tvs_pipeline = Pipeline(stages=[string_indexer, \one_hot_encoder, vector_assembler, \train_validataion_split])
# tvs_pipeline 进行数据处理、模型训练(找到最佳模型)
tvs_pipeline_model = tvs_pipeline.fit(train_df)best_model = tvs_pipeline_model.stages[3].bestModel
"""
DecisionTreeClassificationModel (uid=DecisionTreeClassifier_\
451682088ef8fcaa79ae) of depth 20 with 1851 nodes
"""
评估最佳模型
predictions_df = tvs_pipeline_model.transform(test_df)model_auc = binary_class_evaluator.evaluate(predictions_df)
print model_auc0.649609702764
"""__init__(self, estimator=None, estimatorParamMaps=None, \evaluator=None, numFolds=3, seed=None)假设 K-Fold的CrossValidation交叉验证 K = 3,将数据分为3个部分:1、A + B作为训练,C作为验证2、B + C作为训练,A作为验证3、A + C最为训练,B作为验证"""# 导入模块
from pyspark.ml.tuning import CrossValidator
# 构建 CrossValidator实例对象,设置相关参数
cross_validator = CrossValidator(estimator=dt, \evaluator=binary_class_evaluator,\estimatorParamMaps=param_grid, numFolds=3)# 创建Pipeline
cv_pipeline = Pipeline(stages=[string_indexer, one_hot_encoder, \vector_assembler, cross_validator])
使用 cv_pipeline 进行训练与验证(交叉)
cv_pipeline_model = cv_pipeline.fit(train_df)
查看最佳模型
best_model = cv_pipeline_model.stages[3].bestModel
"""
DecisionTreeClassificationModel (uid=DecisionTreeClassifier_ \
451682088ef8fcaa79ae) of depth 10 with 527 nodes
"""
使用测试集评估最佳模型
cv_predictions = cv_pipeline_model.transform(test_df)
cv_model_auc = binary_class_evaluator.evaluate(cv_predictions)
print cv_model_auc
六、提升:随即森林(RF算法)
# 导入随机森林分类算法模块
from pyspark.ml.classification import RandomForestClassifier# 创建RFC实例对象
rfc = RandomForestClassifier(labelCol='label', \featuresCol='features',\numTrees=10, \featureSubsetStrategy="auto",\maxDepth=5, \maxBins=32, \impurity="gini")# 创建Pipeline实例对象
rfc_pipeline = Pipeline(stages=[string_indexer, one_hot_encoder, \vector_assembler, rfc])# 使用训练数据训练模型
rfc_pipeline_model = rfc_pipeline.fit(train_df)# 预测
rfc_predictions = rfc_pipeline_model.transform(test_df)rfc_model_auc = binary_class_evaluator.evaluate(rfc_predictions)
print rfc_model_auc
"""
0.716242043615
"""