Mahout源码分析之 -- 文档向量化TF-IDF

2024-06-12 20:18

本文主要是介绍Mahout源码分析之 -- 文档向量化TF-IDF,希望对大家解决编程问题提供一定的参考价值,需要的开发者们随着小编来一起学习吧!

 Mahout之SparseVectorsFromSequenceFiles源码分析

一、原理

TF-IDF是一种统计方法,用以评估一字词对于一个文件集或一个语料库中的其中一份文件的重要程度。字词的重要性随着它在文件中出现的次数成正比增加,但同时会随着它在语料库中出现的频率成反比下降。

TFIDF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。

TFIDF实际上是:TF * IDF,TF词频(Term Frequency),IDF逆向文件频率(Inverse Document Frequency)。

词频 (TF) 指的是某一个给定的词语在文件中出现的次数。这个数字通常会被归一化,以防止它偏向长的文件。(同一个词语在长文件里可能会比短文件有更高的词频,而不管该词语重要与否。)

逆向文件频率(IDF)是一个词语普遍重要性的度量,其主要思想是:如果包含词条t的文档越少,也就是n越小,IDF越大,则说明词条t具有很好的类别区分能力。

对于在某一特定文件里的词语  来说,它的重要性可表示为:


以上式子中  是该词在文件中的出现次数,而分母则是在文件中所有字词的出现次数之和(分母也可以是词出现次数的最大值)。

逆向文件频率(inverse document frequency,IDF)是一个词语普遍重要性的度量。某一特定词语的IDF,可以由总文件数目除以包含该词语之文件的数目,再将得到的商取对数得到:


其中

  • |D|:语料库中的文件总数
  • :包含词语的文件数目(即的文件数目)如果该词语不在语料库中,就会导致分母为零,因此一般情况下使用

然后


某一特定文件内的高词语频率,以及该词语在整个文件集合中的低文件频率,可以产生出高权重的TF-IDF。因此,TF-IDF倾向于过滤掉常见的词语,保留重要的词语。

二、源码分析

目标:将一个给定的sequence文件集合转化为SparseVectors

1、对文档分词

1.1)使用最新的{@link org.apache.lucene.util.Version}创建一个Analyzer,用来下文1.2分词;

复制代码
      Class<? extends Analyzer> analyzerClass = StandardAnalyzer.class;if (cmdLine.hasOption(analyzerNameOpt)) {String className = cmdLine.getValue(analyzerNameOpt).toString();analyzerClass = Class.forName(className).asSubclass(Analyzer.class);// try instantiating it, b/c there isn't any point in setting it if// you can't instantiate it
        AnalyzerUtils.createAnalyzer(analyzerClass);}
复制代码

1.2)使用{@link StringTuple}将input documents转化为token数组(input documents必须是{@link org.apache.hadoop.io.SequenceFile}格式);

DocumentProcessor.tokenizeDocuments(inputDir, analyzerClass, tokenizedPath, conf);

输入:inputDir     输出:tokenizedPath

SequenceFileTokenizerMapper

复制代码
 //将input documents按Analyzer进行分词,并将分得的词放在一个StringTuple中TokenStream stream = analyzer.tokenStream(key.toString(), new StringReader(value.toString()));CharTermAttribute termAtt = stream.addAttribute(CharTermAttribute.class);stream.reset();StringTuple document = new StringTuple();//StringTuple是一个能够被用于Hadoop Map/Reduce Job的String类型有序Listwhile (stream.incrementToken()) {if (termAtt.length() > 0) {document.add(new String(termAtt.buffer(), 0, termAtt.length()));}}
复制代码

2、创建TF向量(Term Frequency Vectors)---多个Map/Reduce Job

复制代码
        DictionaryVectorizer.createTermFrequencyVectors(tokenizedPath,outputDir,tfDirName,conf,minSupport,maxNGramSize,minLLRValue,-1.0f,false,reduceTasks,chunkSize,sequentialAccessOutput,namedVectors);
复制代码

2.1)全局词统计(TF)

startWordCounting(input, dictionaryJobPath, baseConf, minSupport);

使用Map/Reduce并行地统计全局的词频,这里只考虑(maxNGramSize == 1) 

输入:tokenizedPath   输出:wordCountPath

TermCountMapper

复制代码
  //统计一个文本文档中的词频OpenObjectLongHashMap<String> wordCount = new OpenObjectLongHashMap<String>();for (String word : value.getEntries()) {if (wordCount.containsKey(word)) {wordCount.put(word, wordCount.get(word) + 1);} else {wordCount.put(word, 1);}}wordCount.forEachPair(new ObjectLongProcedure<String>() {@Overridepublic boolean apply(String first, long second) {try {context.write(new Text(first), new LongWritable(second));} catch (IOException e) {context.getCounter("Exception", "Output IO Exception").increment(1);} catch (InterruptedException e) {context.getCounter("Exception", "Interrupted Exception").increment(1);}return true;}});
复制代码

TermCountCombiner:( 同 TermCountReducer)

TermCountReducer

复制代码
//汇总所有的words和单词的weights,并将同一word的权重sumlong sum = 0;for (LongWritable value : values) {sum += value.get();}if (sum >= minSupport) {//TermCountCombiner没有这个过滤)context.write(key, new LongWritable(sum));}
复制代码

2.2)创建词典

 List<Path> dictionaryChunks;dictionaryChunks =createDictionaryChunks(dictionaryJobPath, output, baseConf, chunkSizeInMegabytes, maxTermDimension);

读取2.1词频Job的feature frequency List,并给它们指定id

输入:wordCountPath   输出:dictionaryJobPath

复制代码
 /*** Read the feature frequency List which is built at the end of the Word Count Job and assign ids to them.* This will use constant memory and will run at the speed of your disk read*/private static List<Path> createDictionaryChunks(Path wordCountPath,Path dictionaryPathBase,Configuration baseConf,int chunkSizeInMegabytes,int[] maxTermDimension) throws IOException {List<Path> chunkPaths = Lists.newArrayList();Configuration conf = new Configuration(baseConf);FileSystem fs = FileSystem.get(wordCountPath.toUri(), conf);long chunkSizeLimit = chunkSizeInMegabytes * 1024L * 1024L;//默认64Mint chunkIndex = 0;Path chunkPath = new Path(dictionaryPathBase, DICTIONARY_FILE + chunkIndex);chunkPaths.add(chunkPath);SequenceFile.Writer dictWriter = new SequenceFile.Writer(fs, conf, chunkPath, Text.class, IntWritable.class);try {long currentChunkSize = 0;Path filesPattern = new Path(wordCountPath, OUTPUT_FILES_PATTERN);int i = 0;for (Pair<Writable,Writable> record: new SequenceFileDirIterable<Writable,Writable>(filesPattern, PathType.GLOB, null, null, true, conf)) {if (currentChunkSize > chunkSizeLimit) {//生成新的词典文件Closeables.close(dictWriter, false);chunkIndex++;chunkPath = new Path(dictionaryPathBase, DICTIONARY_FILE + chunkIndex);chunkPaths.add(chunkPath);dictWriter = new SequenceFile.Writer(fs, conf, chunkPath, Text.class, IntWritable.class);currentChunkSize = 0;}Writable key = record.getFirst();int fieldSize = DICTIONARY_BYTE_OVERHEAD + key.toString().length() * 2 + Integer.SIZE / 8;currentChunkSize += fieldSize;dictWriter.append(key, new IntWritable(i++));//指定id}maxTermDimension[0] = i;//记录最大word数目} finally {Closeables.close(dictWriter, false);}return chunkPaths;}
复制代码

2.3)构造PartialVectors(TF)

复制代码
int partialVectorIndex = 0;Collection<Path> partialVectorPaths = Lists.newArrayList();for (Path dictionaryChunk : dictionaryChunks) {Path partialVectorOutputPath = new Path(output, VECTOR_OUTPUT_FOLDER + partialVectorIndex++);partialVectorPaths.add(partialVectorOutputPath);makePartialVectors(input, baseConf, maxNGramSize, dictionaryChunk, partialVectorOutputPath,maxTermDimension[0], sequentialAccess, namedVectors, numReducers);}
复制代码

将input documents使用a chunk of features创建a partial vector

(这是由于词典文件被分成了多个文件,每个文件只能构造总的vector的一部分,其中每一部分叫一个partial vector)

输入:tokenizedPath   输出:partialVectorPaths

Mapper:(Mapper)

TFPartialVectorReducer

复制代码
    //读取词典文件
//MAHOUT-1247Path dictionaryFile = HadoopUtil.getSingleCachedFile(conf);// key is word value is idfor (Pair<Writable, IntWritable> record: new SequenceFileIterable<Writable, IntWritable>(dictionaryFile, true, conf)) {dictionary.put(record.getFirst().toString(), record.getSecond().get());}
复制代码
复制代码
//转化a document为a sparse vectorStringTuple value = it.next();Vector vector = new RandomAccessSparseVector(dimension, value.length()); // guess at initial sizefor (String term : value.getEntries()) {if (!term.isEmpty() && dictionary.containsKey(term)) { // unigramint termId = dictionary.get(term);vector.setQuick(termId, vector.getQuick(termId) + 1);}}
复制代码

2.4)合并PartialVectors(TF)

    Configuration conf = new Configuration(baseConf);Path outputDir = new Path(output, tfVectorsFolderName);PartialVectorMerger.mergePartialVectors(partialVectorPaths, outputDir, conf, normPower, logNormalize,maxTermDimension[0], sequentialAccess, namedVectors, numReducers);

合并所有的partial {@link org.apache.mahout.math.RandomAccessSparseVector}s为完整的{@link org.apache.mahout.math.RandomAccessSparseVector}

输入:partialVectorPaths   输出:tfVectorsFolder

Mapper:(Mapper)

PartialVectorMergeReducer:

//合并partial向量为完整的TF向量Vector vector = new RandomAccessSparseVector(dimension, 10);for (VectorWritable value : values) {vector.assign(value.get(), Functions.PLUS);//将包含不同word的向量合并为一个}

 3、创建IDF向量(document frequency Vectors)---多个Map/Reduce Job

复制代码
      Pair<Long[], List<Path>> docFrequenciesFeatures = null;// Should document frequency features be processedif (shouldPrune || processIdf) {log.info("Calculating IDF");docFrequenciesFeatures =TFIDFConverter.calculateDF(new Path(outputDir, tfDirName), outputDir, conf, chunkSize);}
复制代码

3.1)统计DF词频

Path wordCountPath = new Path(output, WORDCOUNT_OUTPUT_FOLDER);

startDFCounting(input, wordCountPath, baseConf);

输入:tfDir  输出:featureCountPath

 TermDocumentCountMapper

复制代码
 //为一个文档中的每个word计数1、文档数1Vector vector = value.get();for (Vector.Element e : vector.nonZeroes()) {out.set(e.index());context.write(out, ONE);}context.write(TOTAL_COUNT, ONE);
复制代码

Combiner:(TermDocumentCountReducer)

TermDocumentCountReducer

   //将每个word的文档频率和文档总数sumlong sum = 0;for (LongWritable value : values) {sum += value.get();}

3.2)df词频分块

 return createDictionaryChunks(wordCountPath, output, baseConf, chunkSizeInMegabytes);

将df词频分块存放到多个文件,记录word总数、文档总数

输入:featureCountPath    输出:dictionaryPathBase

复制代码
  /*** Read the document frequency List which is built at the end of the DF Count Job. This will use constant* memory and will run at the speed of your disk read*/private static Pair<Long[], List<Path>> createDictionaryChunks(Path featureCountPath,Path dictionaryPathBase,Configuration baseConf,int chunkSizeInMegabytes) throws IOException {List<Path> chunkPaths = Lists.newArrayList();Configuration conf = new Configuration(baseConf);FileSystem fs = FileSystem.get(featureCountPath.toUri(), conf);long chunkSizeLimit = chunkSizeInMegabytes * 1024L * 1024L;int chunkIndex = 0;Path chunkPath = new Path(dictionaryPathBase, FREQUENCY_FILE + chunkIndex);chunkPaths.add(chunkPath);SequenceFile.Writer freqWriter =new SequenceFile.Writer(fs, conf, chunkPath, IntWritable.class, LongWritable.class);try {long currentChunkSize = 0;long featureCount = 0;long vectorCount = Long.MAX_VALUE;Path filesPattern = new Path(featureCountPath, OUTPUT_FILES_PATTERN);for (Pair<IntWritable,LongWritable> record: new SequenceFileDirIterable<IntWritable,LongWritable>(filesPattern,PathType.GLOB,null,null,true,conf)) {if (currentChunkSize > chunkSizeLimit) {Closeables.close(freqWriter, false);chunkIndex++;chunkPath = new Path(dictionaryPathBase, FREQUENCY_FILE + chunkIndex);chunkPaths.add(chunkPath);freqWriter = new SequenceFile.Writer(fs, conf, chunkPath, IntWritable.class, LongWritable.class);currentChunkSize = 0;}int fieldSize = SEQUENCEFILE_BYTE_OVERHEAD + Integer.SIZE / 8 + Long.SIZE / 8;currentChunkSize += fieldSize;IntWritable key = record.getFirst();LongWritable value = record.getSecond();if (key.get() >= 0) {freqWriter.append(key, value);} else if (key.get() == -1) {//文档数目vectorCount = value.get();}featureCount = Math.max(key.get(), featureCount);}featureCount++;Long[] counts = {featureCount, vectorCount};//word数目、文档数目return new Pair<Long[], List<Path>>(counts, chunkPaths);} finally {Closeables.close(freqWriter, false);}}
复制代码

4、创建TFIDF(Term Frequency-Inverse Document Frequency (Tf-Idf) Vectors)

        TFIDFConverter.processTfIdf(new Path(outputDir, DictionaryVectorizer.DOCUMENT_VECTOR_OUTPUT_FOLDER),outputDir, conf, docFrequenciesFeatures, minDf, maxDF, norm, logNormalize,sequentialAccessOutput, namedVectors, reduceTasks);

4.1)生成PartialVectors(TFIDF)

复制代码
  int partialVectorIndex = 0;List<Path> partialVectorPaths = Lists.newArrayList();List<Path> dictionaryChunks = datasetFeatures.getSecond();for (Path dictionaryChunk : dictionaryChunks) {Path partialVectorOutputPath = new Path(output, VECTOR_OUTPUT_FOLDER + partialVectorIndex++);partialVectorPaths.add(partialVectorOutputPath);makePartialVectors(input,baseConf,datasetFeatures.getFirst()[0],datasetFeatures.getFirst()[1],minDf,maxDF,dictionaryChunk,partialVectorOutputPath,sequentialAccessOutput,namedVector);}
复制代码

使用a chunk of features创建a partial tfidf vector

输入:tfVectorsFolder   输出:partialVectorOutputPath

    DistributedCache.setCacheFiles(new URI[] {dictionaryFilePath.toUri()}, conf);//缓存df分块文件

Mapper:(Mapper)

TFIDFPartialVectorReducer

复制代码
  //计算每个文档中每个word的TFIDF值
Vector value = it.next().get();Vector vector = new RandomAccessSparseVector((int) featureCount, value.getNumNondefaultElements());for (Vector.Element e : value.nonZeroes()) {if (!dictionary.containsKey(e.index())) {continue;}long df = dictionary.get(e.index());if (maxDf > -1 && (100.0 * df) / vectorCount > maxDf) {continue;}if (df < minDf) {df = minDf;}vector.setQuick(e.index(), tfidf.calculate((int) e.get(), (int) df, (int) featureCount, (int) vectorCount));}
复制代码

 4.2)合并partial向量(TFIDF)

复制代码
    Configuration conf = new Configuration(baseConf);Path outputDir = new Path(output, DOCUMENT_VECTOR_OUTPUT_FOLDER);PartialVectorMerger.mergePartialVectors(partialVectorPaths,outputDir,baseConf,normPower,logNormalize,datasetFeatures.getFirst()[0].intValue(),sequentialAccessOutput,namedVector,numReducers);
复制代码

合并所有的partial向量为一个完整的文档向量

 输入:partialVectorOutputPath   输出:outputDir

 Mapper:Mapper

 PartialVectorMergeReducer

    //汇总TFIDF向量Vector vector = new RandomAccessSparseVector(dimension, 10);for (VectorWritable value : values) {vector.assign(value.get(), Functions.PLUS);}

 

这篇关于Mahout源码分析之 -- 文档向量化TF-IDF的文章就介绍到这儿,希望我们推荐的文章对编程师们有所帮助!


原文地址:
本文来自互联网用户投稿,该文观点仅代表作者本人,不代表本站立场。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。如若转载,请注明出处:http://www.chinasem.cn/article/1055225

相关文章

Apache 高级配置实战之从连接保持到日志分析的完整指南

《Apache高级配置实战之从连接保持到日志分析的完整指南》本文带你从连接保持优化开始,一路走到访问控制和日志管理,最后用AWStats来分析网站数据,对Apache配置日志分析相关知识感兴趣的朋友... 目录Apache 高级配置实战:从连接保持到日志分析的完整指南前言 一、Apache 连接保持 - 性

Maven项目中集成数据库文档生成工具的操作步骤

《Maven项目中集成数据库文档生成工具的操作步骤》在Maven项目中,可以通过集成数据库文档生成工具来自动生成数据库文档,本文为大家整理了使用screw-maven-plugin(推荐)的完... 目录1. 添加插件配置到 pom.XML2. 配置数据库信息3. 执行生成命令4. 高级配置选项5. 注意事

Linux中的more 和 less区别对比分析

《Linux中的more和less区别对比分析》在Linux/Unix系统中,more和less都是用于分页查看文本文件的命令,但less是more的增强版,功能更强大,:本文主要介绍Linu... 目录1. 基础功能对比2. 常用操作对比less 的操作3. 实际使用示例4. 为什么推荐 less?5.

spring-gateway filters添加自定义过滤器实现流程分析(可插拔)

《spring-gatewayfilters添加自定义过滤器实现流程分析(可插拔)》:本文主要介绍spring-gatewayfilters添加自定义过滤器实现流程分析(可插拔),本文通过实例图... 目录需求背景需求拆解设计流程及作用域逻辑处理代码逻辑需求背景公司要求,通过公司网络代理访问的请求需要做请

Java集成Onlyoffice的示例代码及场景分析

《Java集成Onlyoffice的示例代码及场景分析》:本文主要介绍Java集成Onlyoffice的示例代码及场景分析,本文给大家介绍的非常详细,对大家的学习或工作具有一定的参考借鉴价值,需要... 需求场景:实现文档的在线编辑,团队协作总结:两个接口 + 前端页面 + 配置项接口1:一个接口,将o

Python使用python-docx实现自动化处理Word文档

《Python使用python-docx实现自动化处理Word文档》这篇文章主要为大家展示了Python如何通过代码实现段落样式复制,HTML表格转Word表格以及动态生成可定制化模板的功能,感兴趣的... 目录一、引言二、核心功能模块解析1. 段落样式与图片复制2. html表格转Word表格3. 模板生

IDEA下"File is read-only"可能原因分析及"找不到或无法加载主类"的问题

《IDEA下Fileisread-only可能原因分析及找不到或无法加载主类的问题》:本文主要介绍IDEA下Fileisread-only可能原因分析及找不到或无法加载主类的问题,具有很好的参... 目录1.File is read-only”可能原因2.“找不到或无法加载主类”问题的解决总结1.File

8种快速易用的Python Matplotlib数据可视化方法汇总(附源码)

《8种快速易用的PythonMatplotlib数据可视化方法汇总(附源码)》你是否曾经面对一堆复杂的数据,却不知道如何让它们变得直观易懂?别慌,Python的Matplotlib库是你数据可视化的... 目录引言1. 折线图(Line Plot)——趋势分析2. 柱状图(Bar Chart)——对比分析3

Dubbo之SPI机制的实现原理和优势分析

《Dubbo之SPI机制的实现原理和优势分析》:本文主要介绍Dubbo之SPI机制的实现原理和优势,具有很好的参考价值,希望对大家有所帮助,如有错误或未考虑完全的地方,望不吝赐教... 目录Dubbo中SPI机制的实现原理和优势JDK 中的 SPI 机制解析Dubbo 中的 SPI 机制解析总结Dubbo中

C#继承之里氏替换原则分析

《C#继承之里氏替换原则分析》:本文主要介绍C#继承之里氏替换原则,具有很好的参考价值,希望对大家有所帮助,如有错误或未考虑完全的地方,望不吝赐教... 目录C#里氏替换原则一.概念二.语法表现三.类型检查与转换总结C#里氏替换原则一.概念里氏替换原则是面向对象设计的基本原则之一:核心思想:所有引py