分词-TFIDF-特征降维（信息增益）

2019-11-15 00:35:45

字体：大中小

来源：转载

供稿：网友

分词-TFIDF-特征降维（信息增益）

前提：首先说明一下TFIDF的部分是借用

http://www.VEVb.com/ywl925/archive/2013/08/26/3275878.html

这篇博文写的代码，因为工作需要在后面加上了使用信息增益的方法进行特征降维。

TFIDF的介绍在此就不赘述了，直接将公式摆出来。

TF公式：

$/mathrm{tf_{i,j}} = /frac{n_{i,j}}{/sum_k n_{k,j}}$

以上式子中 $n_{i,j}$ 是该词在文件 $d_{j}$ 中的出现次数，而分母则是在文件 $d_{j}$ 中所有字词的出现次数之和。

IDF公式：

$/mathrm{idf_{i}} = /log /frac{|D|}{|/{j: t_{i} /in d_{j}/}|}$

|D|：语料库中的文件总数
$|/{ j: t_{i} /in d_{j}/}|$ ：包含词语 $t_{i}$ 的文件数目（即 $n_{i,j} /neq 0$ 的文件数目）如果该词语不在语料库中，就会导致被除数为零，因此一般情况下使用 $1 + |/{j: t_{i} /in d_{j}/}|$

然后

$/mathrm{tf{}idf_{i,j}} = /mathrm{tf_{i,j}} /times /mathrm{idf_{i}}$

信息增益

其公式为：

假如有变量X，其可能的取值有n种，每一种取到的概率为Pi，那么X的熵就定义为

熵公式

也就是说X可能的变化越多，X所携带的信息量越大，熵也就越大。对于文本分类或聚类而言，就是说文档属于哪个类别的变化越多，类别的信息量就越大。所以特征T给聚类C或分类C带来的信息增益为

IG(T)=H(C)-H(C|T)

H(C|T)包含两种情况：一种是特征T出现，标记为t，一种是特征T不出现，标记为t'。所以

H(C|T)=P(t)H(C|t)+P(t')H(C|t‘)

本例属于文本分类其p(t)为该词在所有分类中出现的概率，H(C|t)该词出现的条件下分类的熵。

本例的数据为自行搜索的不良信息中的两类，暴力和反动。提供两种筛选方式一种设立阈值另一种是进行排序后选取前多少个为特征值。

涉及的文件

停用词表和分词jar包：http://files.VEVb.com/files/mansiisnam/%E6%96%87%E4%BB%B6.zip

代码如下

package TIDF;import java.io.*;import java.util.*;import org.wltea.analyzer.lucene.IKAnalyzer;/** * 分词-TFIDF-信息增益 * @author LJ *  * @datetime 2015-6-15  */public class TestTfIdf {    public static final String stopWordTable = "C:/Users/zzw/Desktop/sc_ot-tingyongzhongwen_hc/stopWordTable.txt"; // 加载停用词库    PRivate static ArrayList<String> FileList = new ArrayList<String>(); // 文件列表    // 递归读取该路径下文件返回文件列表    public static List<String> readDirs(String filepath)            throws FileNotFoundException, IOException {        try {            File file = new File(filepath);            if (!file.isDirectory()) {                System.out.println("输入的[]");                System.out.println("filepath:" + file.getAbsolutePath());            } else {                String[] flist = file.list();                for (int i = 0; i < flist.length; i++) {                    File newfile = new File(filepath + "//" + flist[i]);                    if (!newfile.isDirectory()) {                        FileList.add(newfile.getAbsolutePath());                    } else if (newfile.isDirectory()) {                        readDirs(filepath + "//" + flist[i]);                    }                }            }        } catch (FileNotFoundException e) {            System.out.println(e.getMessage());        }        return FileList;    }    // 读入文件    public static String readFile(String file) throws FileNotFoundException,            IOException {        StringBuffer strSb = new StringBuffer();        InputStreamReader inStrR = new InputStreamReader(new FileInputStream(                file), "gbk");        BufferedReader br = new BufferedReader(inStrR);        String line = br.readLine();        while (line != null) {            strSb.append(line).append("/r/n");            line = br.readLine();        }        return strSb.toString();    }    // 分词处理    public static ArrayList<String> cutWords(String file) throws IOException {        ArrayList<String> fenci = new ArrayList<String>();        ArrayList<String> words = new ArrayList<String>();        String text = TestTfIdf.readFile(file);        IKAnalyzer analyzer = new IKAnalyzer();        fenci = analyzer.split(text); // 分词处理        BufferedReader StopWordFileBr = new BufferedReader(                new InputStreamReader(new FileInputStream(new File(                        stopWordTable))));        // 用来存放停用词的集合        Set<String> stopWordSet = new HashSet<String>();        // 初如化停用词集        String stopWord = null;        for (; (stopWord = StopWordFileBr.readLine()) != null;) {            stopWordSet.add(stopWord);        }        for (String word : fenci) {            if (stopWordSet.contains(word)) {                continue;            }            words.add(word);        }        System.out.println(words);        return words;    }    // 统计一个文件中每个词出现的次数    public static HashMap<String, Integer> normalTF(ArrayList<String> cutwords) {        HashMap<String, Integer> resTF = new HashMap<String, Integer>();        for (String word : cutwords) {            if (resTF.get(word) == null) {                resTF.put(word, 1);                System.out.println(word);            } else {                resTF.put(word, resTF.get(word) + 1);                System.out.println(word.toString());            }        }        System.out.println(resTF);        return resTF;    }    // 计算一个文件每个词tf值    @SuppressWarnings("unchecked")    public static HashMap<String, Float> tf(ArrayList<String> cutwords) {        HashMap<String, Float> resTF = new HashMap<String, Float>();        int wordLen = cutwords.size();        HashMap<String, Integer> intTF = TestTfIdf.normalTF(cutwords);        Iterator iter = intTF.entrySet().iterator();        while (iter.hasNext()) {            Map.Entry entry = (Map.Entry) iter.next();            resTF.put(entry.getKey().toString(), Float.parseFloat(entry                    .getValue().toString())                    / wordLen);            System.out.println(entry.getKey().toString() + " = "                    + Float.parseFloat(entry.getValue().toString()) / wordLen);        }        return resTF;    }    // tf times for file 。。。。。。。    public static HashMap<String, HashMap<String, Integer>> normalTFAllFiles(            String dirc) throws IOException {        FileList.clear();        HashMap<String, HashMap<String, Integer>> allNormalTF = new HashMap<String, HashMap<String, Integer>>();        List<String> filelist = TestTfIdf.readDirs(dirc);        for (String file : filelist) {            HashMap<String, Integer> dict = new HashMap<String, Integer>();            ArrayList<String> cutwords = TestTfIdf.cutWords(file);            dict = TestTfIdf.normalTF(cutwords);            allNormalTF.put(file, dict);        }        return allNormalTF;    }    // 返回所有文件tf值    public static HashMap<String, HashMap<String, Float>> tfAllFiles(String dirc)            throws IOException {        FileList.clear();        HashMap<String, HashMap<String, Float>> allTF = new HashMap<String, HashMap<String, Float>>();        List<String> filelist = TestTfIdf.readDirs(dirc);        for (String file : filelist) {            HashMap<String, Float> dict = new HashMap<String, Float>();            ArrayList<String> cutwords = TestTfIdf.cutWords(file);            dict = TestTfIdf.tf(cutwords);            allTF.put(file, dict);        }        return allTF;    }    // 计算该目录下所有词的idf    @SuppressWarnings("unchecked")    public static HashMap<String, Float> idf(            HashMap<String, HashMap<String, Float>> all_tf, String file)            throws IOException {        FileList.clear();        HashMap<String, Float> resIdf = new HashMap<String, Float>();        HashMap<String, Integer> dict = new HashMap<String, Integer>();        int docNum = readDirs(file).size();        for (int i = 0; i < docNum; i++) {            HashMap<String, Float> temp = all_tf.get(FileList.get(i));            Iterator iter = temp.entrySet().iterator();            while (iter.hasNext()) {                Map.Entry entry = (Map.Entry) iter.next();                String word = entry.getKey().toString();                if (dict.get(word) == null) {                    dict.put(word, 1);                } else {                    dict.put(word, dict.get(word) + 1);                }            }        }        // 生成文件记录所有词和包含该词的文件数        StringBuilder sb1 = new StringBuilder();        Iterator iter1 = dict.entrySet().iterator();        while (iter1.hasNext()) {            Map.Entry entry = (Map.Entry) iter1.next();            if (entry.getKey().toString() != null) {                sb1.append(entry.getKey().toString() + " "                        + dict.get(entry.getKey()) + "/r/n");            }        }        File filewriter = new File("E:/allCount.txt");        FileWriter fw = new FileWriter(filewriter.getAbsoluteFile());        BufferedWriter bb = new BufferedWriter(fw);        bb.write(sb1.toString());        bb.close();        System.out.println(dict);        // 计算idf        System.out.println("IDF for every word is:");        Iterator iter_dict = dict.entrySet().iterator();        while (iter_dict.hasNext()) {            Map.Entry entry = (Map.Entry) iter_dict.next();            float value = (float) Math.log(docNum                    / Float.parseFloat(entry.getValue().toString()));            resIdf.put(entry.getKey().toString(), value);            System.out.println(entry.getKey().toString() + " = " + value);        }        return resIdf;    }    // 返回该目录下所有词以及包含词的文件数    @SuppressWarnings("unchecked")    public static HashMap<String, Integer> idf_dict(            HashMap<String, HashMap<String, Float>> all_tf, String file)            throws IOException {        FileList.clear();        HashMap<String, Integer> dict = new HashMap<String, Integer>();        List<String> filelist = readDirs(file);        int docNum = filelist.size();        for (int i = 0; i < docNum; i++) {            HashMap<String, Float> temp = all_tf.get(filelist.get(i));            Iterator iter = temp.entrySet().iterator();            while (iter.hasNext()) {                Map.Entry entry = (Map.Entry) iter.next();                String word = entry.getKey().toString();                if (dict.get(word) == null) {                    dict.put(word, 1);                } else {                    dict.put(word, dict.get(word) + 1);                }            }        }        System.out.println(dict);        return dict;    }    // 计算TFIDF值    @SuppressWarnings("unchecked")    public static void tf_idf(HashMap<String, HashMap<String, Float>> all_tf,            HashMap<String, Float> idfs, String file) throws IOException {        HashMap<String, HashMap<String, Float>> resTfIdf = new HashMap<String, HashMap<String, Float>>();        FileList.clear();        int docNum = readDirs(file).size();        for (int i = 0; i < docNum; i++) {            String filepath = FileList.get(i);            HashMap<String, Float> tfidf = new HashMap<String, Float>();            HashMap<String, Float> temp = all_tf.get(filepath);            Iterator iter = temp.entrySet().iterator();            while (iter.hasNext()) {                Map.Entry entry = (Map.Entry) iter.next();                String word = entry.getKey().toString();                Float value = (float) Float.parseFloat(entry.getValue()                        .toString())                        * idfs.get(word);                tfidf.put(word, value);            }            resTfIdf.put(filepath, tfidf);        }        System.out.println("TF-IDF for Every file is :");        DisTfIdf(resTfIdf); // 显示TFIDF    }    // 返回计算的TFIDF值    @SuppressWarnings("unchecked")    public static HashMap<String, HashMap<String, Float>> tf_idf_return(            HashMap<String, HashMap<String, Float>> all_tf,            HashMap<String, Float> idfs, String file) throws IOException {        FileList.clear();        HashMap<String, HashMap<String, Float>> resTfIdf = new HashMap<String, HashMap<String, Float>>();        int docNum = readDirs(file).size();        for (int i = 0; i < docNum; i++) {            @SuppressWarnings("unused")            HashMap<String, Float> tfidf_reduce = new HashMap<String, Float>();            String filepath = FileList.get(i);            HashMap<String, Float> tfidf = new HashMap<String, Float>();            HashMap<String, Float> temp = all_tf.get(filepath);            Iterator iter = temp.entrySet().iterator();            while (iter.hasNext()) {                Map.Entry entry = (Map.Entry) iter.next();                String word = entry.getKey().toString();                Float value = (float) Float.parseFloat(entry.getValue()                        .toString())                        * idfs.get(word);                tfidf.put(word, value);            }            resTfIdf.put(filepath, tfidf);        }        return resTfIdf;    }    // TFIDF显示输出 并建立文件存储该信息    @SuppressWarnings("unchecked")    public static void DisTfIdf(HashMap<String, HashMap<String, Float>> tfidf)            throws IOException {        StringBuilder stall = new StringBuilder();        Iterator iter1 = tfidf.entrySet().iterator();        while (iter1.hasNext()) {            Map.Entry entrys = (Map.Entry) iter1.next();            System.out.println("FileName: " + entrys.getKey().toString());            System.out.print("{");            HashMap<String, Float> temp = (HashMap<String, Float>) entrys                    .getValue();            Iterator iter2 = temp.entrySet().iterator();            while (iter2.hasNext()) {                Map.Entry entry = (Map.Entry) iter2.next();                System.out.print(entry.getKey().toString() + " = "                        + entry.getValue().toString() + ", ");                stall.append(entrys.getKey().toString() + " "                        + entry.getKey().toString() + " "                        + entry.getValue().toString() + "/r/n");            }            System.out.println("}");        }        File filewriter = new File("E:/allTFIDF.txt");        FileWriter fw = new FileWriter(filewriter.getAbsoluteFile());        BufferedWriter bz = new BufferedWriter(fw);        bz.write(stall.toString());        bz.close();    }    // 单属性熵    public static double Entropy(double[] p, double tot) {        double entropy = 0.0;        for (int i = 0; i < p.length; i++) {            if (p[i] > 0.0) {                entropy += -p[i] / tot * Math.log(p[i] / tot) / Math.log(2.0);            }        }        return entropy;    }    // 信息增益特征降维    @SuppressWarnings("unchecked")    private static void Total(int N,            HashMap<String, HashMap<String, Float>> result,            HashMap<String, Integer> idfs_dict_neg,            HashMap<String, Integer> idfs_dict_pos, String file)            throws IOException {        FileList.clear();        double[] classCnt = new double[N]; // 类别数组        double totalCnt = 0.0; // 总文件数        for (int c = 0; c < N; c++) {            classCnt[c] = 125; // 每个类别的文件数目            totalCnt += classCnt[c];        }        int docNum = readDirs(file).size();        int num = 0; // 词f的编号        int numb = 0; // 词f的编号        double totalEntroy = Entropy(classCnt, totalCnt); // 总的熵        HashMap<String, Integer> count = new HashMap<String, Integer>();// 存储词及其编号        HashMap<String, Integer> countG = new HashMap<String, Integer>();// 存储特征降维后word和其编号        HashMap<String, Double> countG1 = new HashMap<String, Double>();// 存储特征降维后word和其信息增益        HashMap<String, Double> infogains = new HashMap<String, Double>();// 存储词和该词的信息增益        StringBuilder st = new StringBuilder();// 缓存文件名,词,信息增益,TFIDF        StringBuilder ss = new StringBuilder();// 缓存未特征处理的类别,单词的编号,单词的TFIDF值        StringBuilder sr = new StringBuilder();// 缓存经过特征处理后的类别,单词的编号,单词的TFIDF值        for (int i = 0; i < docNum; i++) {            String filepath = FileList.get(i);            HashMap<String, Float> temp = result.get(filepath);            Iterator iter = temp.entrySet().iterator();            if (filepath.contains("dubo")) {                ss.append(1 + "  "); // 将赌博类定义为类别1            } else if (filepath.contains("fangdong")) {                ss.append(2 + "  "); // 将反动类定义为类别2            }            while (iter.hasNext()) {                Map.Entry entry = (Map.Entry) iter.next();                String f = entry.getKey().toString();                double[] featureCntWithF = new double[N]; // 包括词F的分布（类别1,2分别包含该词的文件数）                double[] featureCntWithoutF = new double[N]; // 不包括词F的分布                double totalCntWithF = 0.0; // 所有类别中包括词F的文件数                double totalCntWithoutF = 0.0; // 所有类别中不包括词F的文件数                for (int c = 0; c < N; c++) {                    Iterator iter_dict = null;                    switch (c) {                    case 0:                        iter_dict = idfs_dict_neg.entrySet().iterator();                        break;                    case 1:                        iter_dict = idfs_dict_pos.entrySet().iterator();                        break;                    }                    while (iter_dict.hasNext()) {                        Map.Entry entry_neg = (Map.Entry) iter_dict.next();                        if (f.equals(entry_neg.getKey().toString())) { // 该词在该类别中出现                            featureCntWithF[c] = Double.parseDouble(entry_neg                                    .getValue().toString()); // 将该出现该词的文件数赋值给数组                            break;                        } else {                            featureCntWithF[c] = 0.0;                        }                    }                    featureCntWithoutF[c] = classCnt[c] - featureCntWithF[c]; // 不包括词F的文件数等于该类别总数减去包含该词的文件数                    totalCntWithF += featureCntWithF[c];                    totalCntWithoutF += featureCntWithoutF[c];                }                double entropyWithF = Entropy(featureCntWithF, totalCntWithF);                double entropyWithoutF = Entropy(featureCntWithoutF,                        totalCntWithoutF);                double wf = totalCntWithF / totalCnt;                double infoGain = totalEntroy - wf * entropyWithF - (1.0 - wf) // 信息增益的公式                        * entropyWithoutF;                infogains.put(f, infoGain);                st.append(filepath + " " + f + " " + "信息增益" + "="                        + infoGain // 缓存格式                        + " " + "tfidf" + "=" + entry.getValue().toString()                        + "/r/n");                // }                // 方式一：直接用阈值选取特征值可以省去下面再次遍历的过程                // if(infogains.get(f)>0.004011587943125061){                // 给词f编号                if (count.get(f) == null) {                    num++;                    count.put(f, num);                }                ss.append(count.get(f) + ":" + entry.getValue() + " "); // 缓存格式                // }            }            ss.append("/r/n");        }        File fileprepare = new File("E:/test.txt");        FileWriter fz = new FileWriter(fileprepare.getAbsoluteFile());        BufferedWriter bz = new BufferedWriter(fz);        bz.write(ss.toString());        bz.close();        File filewriter = new File("E:/jieguo.txt");        FileWriter fw = new FileWriter(filewriter.getAbsoluteFile());        BufferedWriter bw = new BufferedWriter(fw);        bw.write(st.toString());        bw.close();        // 方式二：将信息增益从大到小排列,选取前特定数的词为特征词        // 对信息增益排序（从大到小）        ArrayList<Map.Entry<String, Double>> infoIds = new ArrayList<Map.Entry<String, Double>>(                infogains.entrySet());        Collections.sort(infoIds, new Comparator<Map.Entry<String, Double>>() {            public int compare(Map.Entry<String, Double> o1,                    Map.Entry<String, Double> o2) {                if (o2.getValue() - o1.getValue() > 0) {                    return 1; // 降序排列                } else {                    return -1;                }            }        });        // 选取信息增益为前2000的词做特征词        for (int c = 0; c < 2000; c++) {            countG1.put(infoIds.get(c).getKey(), infoIds.get(c).getValue()); // 将处理后的数据存储到countG1中        }        // 再次遍历        for (int i = 0; i < docNum; i++) {            String filepath = FileList.get(i);            HashMap<String, Float> temp = result.get(filepath);            Iterator iter = temp.entrySet().iterator();            if (filepath.contains("dubo")) {                sr.append(1 + "  ");            } else if (filepath.contains("fangdong")) {                sr.append(2 + "  ");            }            while (iter.hasNext()) {                Map.Entry entry = (Map.Entry) iter.next();                // for(Iterator<Feature>                // i=index.featureIterator();i.hasNext();){                String f = entry.getKey().toString();                // 判断该词是特征降维后的那些词                if (countG1.get(f) != null) {                    // 给该词编号                    if (countG.get(f) == null) {                        numb++;                        countG.put(f, numb);                    }                    sr.append(countG.get(f) + ":" + entry.getValue() + " ");                }            }            sr.append("/r/n");        }        File fileprepare1 = new File("E:/testt.txt");        FileWriter fr = new FileWriter(fileprepare1.getAbsoluteFile());        BufferedWriter br = new BufferedWriter(fr);        br.write(sr.toString());        br.close();    }    public static void main(String[] args) throws IOException {        // TODO Auto-generated method stub        String file = "C:/Users/zzw/Desktop/项目管理/语料/test"; // 总的数据路径        String file1 = "C:/Users/zzw/Desktop/项目管理/语料/test/赌博"; // 类1数据路径        String file2 = "C:/Users/zzw/Desktop/项目管理/语料/test/反动"; // 类2数据路径        HashMap<String, HashMap<String, Float>> all_tf = tfAllFiles(file);        HashMap<String, HashMap<String, Float>> all_tf_neg = tfAllFiles(file1); // file1文件的tf值和路径        HashMap<String, HashMap<String, Float>> all_tf_pos = tfAllFiles(file2); // file2文件的tf值和路径        System.out.println();        HashMap<String, Integer> idfs_dict_neg = idf_dict(all_tf_neg, file1); // 返回file1下所有词以及包含词的文件数        HashMap<String, Integer> idfs_dict_pos = idf_dict(all_tf_pos, file2); // 返回file2下所有词以及包含词的文件数        HashMap<String, Float> idfs = idf(all_tf, file);        System.out.println();        tf_idf(all_tf, idfs, file);        HashMap<String, HashMap<String, Float>> result = tf_idf_return(all_tf,                idfs, file);        int N = 2; // 输入类别数        /*         * 信息增益公式 IG(T)=H(C)-H(C|T) H(C|T)=P(t)H(C|t)+P(t')H(C|t‘)         */        Total(N, result, idfs_dict_neg, idfs_dict_pos, file); // 按信息增益进行特征降维    }}

各个文件结果如下：

allCount.txt

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