TF-IDF(Term Frequency–InverseDocument Frequency)是一种用于资讯检索与文本挖掘的常用加权技术。TF-IDF的主要思想是:如果某个词或短语在一篇文章中出现的频率TF高,并且在其他文章中很少出现,则认为此词或者短语具有很好的类别区分能力,适合用来分类。
TF-IDF实际是TF*IDF,其中TF(Term Frequency)表示词条 t t t在文档 D i D_i Di中的出现的频率,TF的计算公式如下所示:
T F t , D i = c o u n t ( t ) D i TF_{t,D_i} = \frac{count(t)}{D_i} TFt,Di=Dicount(t)
其中IDF(InverseDocument Frequency)表示总文档与包含词条t的文档的比值求对数,IDF的计算公式如下所示:
I D F t = l o g N ∑ i = 1 N I ( t , D i ) IDF_t = log \frac{N}{ \sum_{i=1}^{N} I(t,D_i) } IDFt=log∑i=1NI(t,Di)N
其中 N N N为所有的文档总数, I ( t , D i ) I(t,D_i) I(t,Di)表示文档 D i D_i Di是否包含词条 t t t,若包含为1,不包含为0。但此处存在一个问题,即当词条 t t t在所有文档中都没有出现的话IDF计算公式的分母为0,此时就需要对IDF做平滑处理,改善后的IDF计算公式如下所示:
I D F t = l o g N 1 + ∑ i = 1 N I ( t , D i ) IDF_t = log \frac{N}{ 1 + \sum_{i=1}^{N} I(t,D_i) } IDFt=log1+∑i=1NI(t,Di)N
那么最终词条 t t t在文档 D i D_i Di中的TF-IDF值为: T F − I D F t , D i = T F t , D i ∗ I D F t TF-IDF_{t,D_i} = TF_{t,D_i} * IDF_t TF−IDFt,Di=TFt,Di∗IDFt 。
从上述的计算词条 t t t在文档 D i D_i Di中的TF-IDF值计算可以看出:当一个词条在文档中出现的频率越高,且新鲜度低(即普遍度低),则其对应的TF-IDF值越高。
比如现在有一个预料库,包含了100篇( N N N)论文,其中涉及包含推荐系统( t t t)这个词条的有20篇,在第一篇论文( D 1 D1 D1)中总共有200个技术词汇,其中推荐系统出现了15次,则词条推荐系统的在第一篇论文( D 1 D1 D1)中的TF-IDF值为:
T F − I D F 推 荐 系 统 = 15 200 ∗ l o g 200 20 + 1 = 0.051 TF-IDF_{推荐系统} = \frac {15} {200} * log \frac{200}{20+1} = 0.051 TF−IDF推荐系统=20015∗log20+1200=0.051
更多详细的关于TFIDF的介绍可以参考
关于TF-IDF的其他实战:
这里需要注意的是在spark2.x中默认不支持dataframe的笛卡尔积操作,需要在创建Spark对象时开启。
创建spark对象,并设置日志等级
// spark.sql.crossJoin.enabled=true spark 2.0 x不支持笛卡尔积操作,需要开启支持
val spark = SparkSession
.builder()
.appName("docSimCalWithTFIDF")
.config("spark.sql.crossJoin.enabled","true")
.master("local[10]")
.enableHiveSupport()
.getOrCreate()
Logger.getRootLogger.setLevel(Level.WARN)
这里以官方样例代码中的三行英文句子为例,创建数据集,并进行分词(spark中的中文分词包有很多,比如jieba,han,ansj,fudannlp等,这里不展开介绍)
val sentenceData = spark.createDataFrame(Seq(
(0, "Hi I heard about Spark"),
(1, "I wish Java could use case classes"),
(2, "Logistic regression models are neat")
)).toDF("label", "sentence")
val tokenizer = new Tokenizer()
.setInputCol("sentence")
.setOutputCol("words")
val wordsData = tokenizer.transform(sentenceData)
wordsData.show(10)
展示的结果为:
+-----+--------------------+--------------------+
|label| sentence| words|
+-----+--------------------+--------------------+
| 0|Hi I heard about ...|[hi, i, heard, ab...|
| 1|I wish Java could...|[i, wish, java, c...|
| 2|Logistic regressi...|[logistic, regres...|
+-----+--------------------+--------------------+
调用官方的tfidf包计算向量:
// setNumFeatures(5)表示将Hash分桶的数量设置为5个,可以根据你的词语数量来调整,一般来说,这个值越大不同的词被计算为一个Hash值的概率就越小,数据也更准确,但需要消耗更大的内存
val hashingTF = new HashingTF()
.setInputCol("words")
.setOutputCol("rawFeatures")
.setNumFeatures(5)
val featurizedData = hashingTF
.transform(wordsData)
featurizedData.show(10)
val idf = new IDF()
.setInputCol("rawFeatures")
.setOutputCol("features")
val idfModel = idf.fit(featurizedData)
val rescaledData = idfModel.transform(featurizedData)
rescaledData.show(10)
rescaledData.select("label", "features").show()
展示的结果为:
+-----+--------------------+--------------------+--------------------+
|label| sentence| words| rawFeatures|
+-----+--------------------+--------------------+--------------------+
| 0|Hi I heard about ...|[hi, i, heard, ab...|(5,[0,2,4],[2.0,2...|
| 1|I wish Java could...|[i, wish, java, c...|(5,[0,2,3,4],[1.0...|
| 2|Logistic regressi...|[logistic, regres...|(5,[0,1,3,4],[1.0...|
+-----+--------------------+--------------------+--------------------+
+-----+--------------------+--------------------+--------------------+--------------------+
|label| sentence| words| rawFeatures| features|
+-----+--------------------+--------------------+--------------------+--------------------+
| 0|Hi I heard about ...|[hi, i, heard, ab...|(5,[0,2,4],[2.0,2...|(5,[0,2,4],[0.0,0...|
| 1|I wish Java could...|[i, wish, java, c...|(5,[0,2,3,4],[1.0...|(5,[0,2,3,4],[0.0...|
| 2|Logistic regressi...|[logistic, regres...|(5,[0,1,3,4],[1.0...|(5,[0,1,3,4],[0.0...|
+-----+--------------------+--------------------+--------------------+--------------------+
+-----+--------------------+
|label| features|
+-----+--------------------+
| 0|(5,[0,2,4],[0.0,0...|
| 1|(5,[0,2,3,4],[0.0...|
| 2|(5,[0,1,3,4],[0.0...|
+-----+--------------------+
其中 ( 5 , [ 0 , 2 , 4 ] , [ 0.0 , 0... (5,[0,2,4],[0.0,0... (5,[0,2,4],[0.0,0... 是向量的一种压缩表示格式,例如 ( 3 , [ 0 , 1 ] , [ 0.1 , 0.2 ] ) (3,[0,1],[0.1,0.2]) (3,[0,1],[0.1,0.2])表示的是 向量的长度为3,其中第 1位和第2位的值为0.1 和0.3,第3位的值为0。
这里需要将其转化为向量的形式,方便后续进行计算,可以直接通过dataframe进行转化,也可以先将dataframe转化为rdd,再进行转化。
datafram通过自定义UDF进行转化如下:
import spark.implicits._
// 解析数据 转化为denseVector格式 datafra
val sparseVectorToDenseVector = udf {
features: SV => features.toDense
}
val df = rescaledData
.select($"label", sparseVectorToDenseVector($"features"))
.withColumn("tag",lit(1))
df.show(10)
展示结果为:
+------+--------------------+---+
|label1| features1|tag|
+------+--------------------+---+
| 0|[0.0,0.0,0.575364...| 1|
| 1|[0.0,0.0,0.575364...| 1|
| 2|[0.0,0.6931471805...| 1|
+------+--------------------+---+
先转化为RDD,再进行转化如下:
val selectedRDD = rescaledData.select("label", "features").rdd
.map( l=>( l.get(0).toString, l.getAs[SV](1).toDense))
selectedRDD.take(10).foreach(println)
展示结果为:
(0,[0.0,0.0,0.5753641449035617,0.0,0.0])
(1,[0.0,0.0,0.5753641449035617,0.28768207245178085,0.0])
(2,[0.0,0.6931471805599453,0.0,0.5753641449035617,0.0])
当然也可以在进行相似度计算时进行转化,实现代码如下:
// 定义相似度计算udf
import spark.implicits._
val df1 = rescaledData
.select($"label".alias("id1"), $"features".alias("f1"))
.withColumn("tag",lit(1))
val df2 = rescaledData
.select($"label".alias("id2"), $"features".alias("f2"))
.withColumn("tag",lit(1))
val simTwoDoc = udf{
(f1: SV, f2: SV) => calTwoDocSim(f1,f2)
}
val df = df1.join(df2, Seq("tag"), "inner")
.where("id1 != id2")
.withColumn("simscore",simTwoDoc(col("f1"), col("f2")))
.where("simscore > 0.0")
.select("id1","id2","simscore")
df.printSchema()
df.show(20)
其中calTwoDocSim 函数实现如下:
/**
* @Author: GaoYangtuan
* @Description: 自定义计算两个文本的距离
* @Thinkgamer: 《推荐系统开发实战》作者,「搜索与推荐Wiki」公号负责人,算法工程师
* @Param: [f1, f2]
* @return: double
**/
def calTwoDocSim(f1: SV, f2: SV): Double = {
val breeze1 =new SparseVector(f1.indices,f1.values, f1.size)
val breeze2 =new SparseVector(f2.indices,f2.values, f2.size)
val cosineSim = breeze1.dot(breeze2) / (norm(breeze1) * norm(breeze2))
cosineSim
}
打印结果如下:
root
|-- id1: integer (nullable = false)
|-- id2: integer (nullable = false)
|-- simscore: double (nullable = false)
+---+---+------------------+
|id1|id2| simscore|
+---+---+------------------+
| 0| 1|0.8944271909999159|
| 1| 0|0.8944271909999159|
| 1| 2|0.2856369296406274|
| 2| 1|0.2856369296406274|
+---+---+------------------+
最后进行排序和保存,代码如下:
val sortAndSlice = udf { simids: Seq[Row] =>
simids.map{
case Row(id2: Int, simscore: Double) => (id2,simscore)
}
.sortBy(_._2)
.reverse
.slice(0,100)
.map(e => e._1 + ":" + e._2.formatted("%.3f"))
.mkString(",")
}
val result = df
.groupBy($"id1")
.agg(collect_list(struct($"id2", $"simscore")).as("simids"))
.withColumn("simids", sortAndSlice(sort_array($"simids", asc = false)))
result.show(10)
result.coalesce(1).write.format("parquet").mode("overwrite").save("data/tfidf")
打印结果如下:
+---+---------------+
|id1| simids|
+---+---------------+
| 1|0:0.894,2:0.286|
| 2| 1:0.286|
| 0| 1:0.894|
+---+---------------+