Statistical-Learning-Method_Code/PLSA/PLSA.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 55,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Original Topics:\n",
      "['tech', 'business', 'sport', 'entertainment', 'politics']\n",
      "1/10\n",
      "2/10\n",
      "3/10\n",
      "4/10\n",
      "5/10\n",
      "6/10\n",
      "7/10\n",
      "8/10\n",
      "9/10\n",
      "10/10\n",
      "Topic 1: said year government people mobile last number growth phone market\n",
      "Topic 2: said people film could would also technology made make government\n",
      "Topic 3: said would could best music also world election labour people\n",
      "Topic 4: said first england also time game players wales would team\n",
      "Topic 5: said also would company year world sales firm market last\n",
      "Time: 531.1292963027954\n"
     ]
    }
   ],
   "source": [
    "import numpy as np\n",
    "import pandas as pd\n",
    "import string\n",
    "from nltk.corpus import stopwords\n",
    "import time\n",
    "\n",
    "\n",
    "#定义加载数据的函数\n",
    "def load_data(file):\n",
    "    '''\n",
    "    INPUT:\n",
    "    file - (str) 数据文件的路径\n",
    "    \n",
    "    OUTPUT:\n",
    "    org_topics - (list) 原始话题标签列表\n",
    "    text - (list) 文本列表\n",
    "    words - (list) 单词列表\n",
    "    \n",
    "    '''\n",
    "    df = pd.read_csv(file)  #读取文件\n",
    "    org_topics = df['category'].unique().tolist()  #保存文本原始的话题标签\n",
    "    df.drop('category', axis=1, inplace=True)\n",
    "    n = df.shape[0]  #n为文本数量\n",
    "    text = []\n",
    "    words = []\n",
    "    for i in df['text'].values:\n",
    "        t = i.translate(str.maketrans('', '', string.punctuation))  #去除文本中的标点符号\n",
    "        t = [j for j in t.split() if j not in stopwords.words('english')]  #去除文本中的停止词\n",
    "        t = [j for j in t if len(j) > 3]  #长度小于等于3的单词大多是无意义的，直接去除\n",
    "        text.append(t)  #将处理后的文本保存到文本列表中\n",
    "        words.extend(set(t))  #将文本中所包含的单词保存到单词列表中\n",
    "    words = list(set(words))  #去除单词列表中的重复单词\n",
    "    return org_topics, text, words\n",
    "\n",
    "\n",
    "#定义构建单词-文本矩阵的函数，这里矩阵的每一项表示单词在文本中的出现频次，也可以用TF-IDF来表示\n",
    "def frequency_counter(text, words):\n",
    "    '''\n",
    "    INPUT:\n",
    "    text - (list) 文本列表\n",
    "    words - (list) 单词列表\n",
    "    \n",
    "    OUTPUT:\n",
    "    words - (list) 出现频次为前1000的单词列表\n",
    "    X - (array) 单词-文本矩阵\n",
    "    \n",
    "    '''\n",
    "    words_cnt = np.zeros(len(words))  #用来保存单词的出现频次\n",
    "    X = np.zeros((1000, len(text)))  #定义m*n的矩阵，其中m为单词列表中的单词个数，为避免运行时间过长，这里只取了出现频次为前1000的单词，因此m为1000，n为文本个数\n",
    "    #循环计算words列表中各单词出现的词频\n",
    "    for i in range(len(text)):\n",
    "        t = text[i]  #取出第i条文本\n",
    "        for w in t:\n",
    "            ind = words.index(w)  #取出第i条文本中的第t个单词在单词列表中的索引\n",
    "            words_cnt[ind] += 1  #对应位置的单词出现频次加一\n",
    "    sort_inds = np.argsort(words_cnt)[::-1]  #对单词出现频次降序排列后取出其索引值\n",
    "    words = [words[ind] for ind in sort_inds[:1000]]  #将出现频次前1000的单词保存到words列表\n",
    "    #构建单词-文本矩阵\n",
    "    for i in range(len(text)):\n",
    "        t = text[i]  #取出第i条文本\n",
    "        for w in t:\n",
    "            if w in words:  #如果文本t中的单词w在单词列表中，则将X矩阵中对应位置加一\n",
    "                ind = words.index(w)\n",
    "                X[ind, i] += 1\n",
    "    return words, X\n",
    "\n",
    "\n",
    "#定义概率潜在语义分析函数，采用EM算法进行PLSA模型的参数估计\n",
    "def do_plsa(X, K, words, iters = 10):\n",
    "    '''\n",
    "    INPUT:\n",
    "    X - (array) 单词-文本矩阵\n",
    "    K - (int) 设定的话题数\n",
    "    words - (list) 出现频次为前1000的单词列表\n",
    "    iters - (int) 设定的迭代次数\n",
    "    \n",
    "    OUTPUT:\n",
    "    P_wi_zk - (array) 话题zk条件下产生单词wi的概率数组\n",
    "    P_zk_dj - (array) 文本dj条件下属于话题zk的概率数组\n",
    "    \n",
    "    '''\n",
    "    M, N = X.shape  #M为单词数，N为文本数\n",
    "    #P_wi_zk表示P(wi|zk)，是一个K*M的数组，其中每个值表示第k个话题zk条件下产生第i个单词wi的概率，这里将每个值随机初始化为0-1之间的浮点数\n",
    "    P_wi_zk = np.random.rand(K, M)\n",
    "    #对于每个话题zk，保证产生单词wi的概率的总和为1\n",
    "    for k in range(K):\n",
    "        P_wi_zk[k] /= np.sum(P_wi_zk[k])\n",
    "    #P_zk_dj表示P(zk|dj)，是一个N*K的数组，其中每个值表示第j个文本dj条件下产生第k个话题zk的概率，这里将每个值随机初始化为0-1之间的浮点数\n",
    "    P_zk_dj = np.random.rand(N, K)\n",
    "    #对于每个文本dj，属于话题zk的概率的总和为1\n",
    "    for n in range(N):\n",
    "        P_zk_dj[n] /= np.sum(P_zk_dj[n])\n",
    "    #P_zk_wi_dj表示P(zk|wi,dj)，是一个M*N*K的数组，其中每个值表示在单词-文本对(wi,dj)的条件下属于第k个话题zk的概率，这里设置初始值为0\n",
    "    P_zk_wi_dj = np.zeros((M, N, K))\n",
    "    #迭代执行E步和M步\n",
    "    for i in range(iters):\n",
    "        print('{}/{}'.format(i+1, iters))  \n",
    "        #执行E步\n",
    "        for m in range(M):\n",
    "            for n in range(N):\n",
    "                sums = 0\n",
    "                for k in range(K):\n",
    "                    P_zk_wi_dj[m, n, k] = P_wi_zk[k, m] * P_zk_dj[n, k]  #计算P(zk|wi,dj)的分子部分，即P(wi|zk)*P(zk|dj)\n",
    "                    sums += P_zk_wi_dj[m, n, k]  #计算P(zk|wi,dj)的分母部分，即P(wi|zk)*P(zk|dj)在K个话题上的总和\n",
    "                P_zk_wi_dj[m, n, :] = P_zk_wi_dj[m, n, :] / sums  #得到单词-文本对(wi,dj)条件下的P(zk|wi,dj)\n",
    "        #执行M步，计算P(wi|zk)\n",
    "        for k in range(K):\n",
    "            s1 = 0\n",
    "            for m in range(M):\n",
    "                P_wi_zk[k, m] = 0\n",
    "                for n in range(N):\n",
    "                    P_wi_zk[k, m] += X[m, n] * P_zk_wi_dj[m, n, k]  #计算P(wi|zk)的分子部分，即n(wi,dj)*P(zk|wi,dj)在N个文本上的总和，其中n(wi,dj)为单词-文本矩阵X在文本对(wi,dj)处的频次\n",
    "                s1 += P_wi_zk[k, m]  #计算P(wi|zk)的分母部分，即n(wi,dj)*P(zk|wi,dj)在N个文本和M个单词上的总和\n",
    "            P_wi_zk[k, :] = P_wi_zk[k, :] / s1  #得到话题zk条件下的P(wi|zk)\n",
    "        #执行M步，计算P(zk|dj)\n",
    "        for n in range(N):\n",
    "            for k in range(K):\n",
    "                P_zk_dj[n, k] = 0\n",
    "                for m in range(M):\n",
    "                    P_zk_dj[n, k] += X[m, n] * P_zk_wi_dj[m, n, k]  #同理计算P(zk|dj)的分子部分，即n(wi,dj)*P(zk|wi,dj)在N个文本上的总和\n",
    "                P_zk_dj[n, k] = P_zk_dj[n, k] / np.sum(X[:, n])  #得到文本dj条件下的P(zk|dj)，其中n(dj)为文本dj中的单词个数，由于我们只取了出现频次前1000的单词，所以这里n(dj)计算的是文本dj中在单词列表中的单词数\n",
    "    return P_wi_zk, P_zk_dj\n",
    "\n",
    "\n",
    "if __name__ == \"__main__\":\n",
    "    org_topics, text, words = load_data('bbc_text.csv')  #加载数据\n",
    "    print('Original Topics:')\n",
    "    print(org_topics)  #打印原始的话题标签列表\n",
    "    start = time.time()  #保存开始时间\n",
    "    words, X = frequency_counter(text, words)  #取频次前1000的单词重新构建单词列表，并构建单词-文本矩阵\n",
    "    K = 5  #设定话题数为5\n",
    "    P_wi_zk, P_zk_dj = do_plsa(X, K, words, iters = 10)  #采用EM算法对PLSA模型进行参数估计\n",
    "    #打印出每个话题zk条件下出现概率最大的前10个单词，即P(wi|zk)在话题zk中最大的10个值对应的单词，作为对话题zk的文本描述\n",
    "    for k in range(K):\n",
    "        sort_inds = np.argsort(P_wi_zk[k])[::-1]  #对话题zk条件下的P(wi|zk)的值进行降序排列后取出对应的索引值\n",
    "        topic = []  #定义一个空列表用于保存话题zk概率最大的前10个单词\n",
    "        for i in range(10):\n",
    "            topic.append(words[sort_inds[i]])  \n",
    "        topic = ' '.join(topic)  #将10个单词以空格分隔，构成对话题zk的文本表述\n",
    "        print('Topic {}: {}'.format(k+1, topic))  #打印话题zk\n",
    "    end = time.time()\n",
    "    print('Time:', end-start)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}