Spark 机器学习 ---CountVectorizer

文本特征提取->> CountVectorizer:基于词频数的文档向量

package Spark_MLlib

import org.apache.spark.ml.feature.CountVectorizer
import org.apache.spark.sql.SparkSession


object 特征抽取_CountVectorizer {
      val spark=SparkSession.builder().master("local").appName("CountVectorizer").getOrCreate()
      import spark.implicits._
  def main(args: Array[String]): Unit = {

       val df= spark.createDataFrame(Seq(
         (0,Array("soyo","spark","soyo2","soyo","8")),
         (1,Array("soyo","hadoop","soyo","hadoop","xiaozhou","soyo2","spark","8","8")),
         (2,Array("soyo","spark","soyo2","hadoop","soyo3","8")),
         (3,Array("soyo","spark","soyo20","hadoop","soyo2","8","8")),
         (4,Array("soyo","8","spark","8","spark","spark","8"))
       )).toDF("id","words")
       val CountVectorizer_Model=new CountVectorizer().setInputCol("words").setOutputCol("features").setVocabSize(3).setMinDF(5).fit(df)//设置词汇表的最大个数为3,在5个文档中出现
    //将根据语料库(所有文档)中的词频排序从高到低进行选择
    CountVectorizer_Model.vocabulary.foreach(println)
        CountVectorizer_Model.transform(df).show(false)
  }

}

结果：

8
spark
soyo
+---+----------------------------------------------------------+-------------------------+
|id |words                                                     |features                 |
+---+----------------------------------------------------------+-------------------------+
|0  |[soyo, spark, soyo2, soyo, 8]                             |(3,[0,1,2],[1.0,1.0,2.0])|
|1  |[soyo, hadoop, soyo, hadoop, xiaozhou, soyo2, spark, 8, 8]|(3,[0,1,2],[2.0,1.0,2.0])|
|2  |[soyo, spark, soyo2, hadoop, soyo3, 8]                    |(3,[0,1,2],[1.0,1.0,1.0])|
|3  |[soyo, spark, soyo20, hadoop, soyo2, 8, 8]                |(3,[0,1,2],[2.0,1.0,1.0])|
|4  |[soyo, 8, spark, 8, spark, spark, 8]                      |(3,[0,1,2],[3.0,3.0,1.0])|
+---+----------------------------------------------------------+-------------------------+
将5篇文档中的词去重后就组成了一个字典，这个字典中有3个词：8,spark,soyo,分别建立索引为0,1,2.
在第三列的文档向量，是由基于字典的索引向量，与对应索引的词频向量所组成的。
文档向量是稀疏的表征，例子中只有3个词可能感觉不出，在实际业务中，字典的长度是上万，而文章中出现的词可能是几百或几千，故很多索引对应的位置词频都是0.