Task 4.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Домашнее задание 4."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Выполнил Гончаров Алексей Владимирович, 274 группа МФТИ, ФУПМ\n",
    "Задание:\n",
    "Соревнование Quora. Необходимо определить дубликаты вопросов на сайте Quora. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "import pandas as pd\n",
    "import numpy as np\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline\n",
    "import seaborn as sns\n",
    "pal = sns.color_palette()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "data = pd.read_csv('train.csv')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1. Статистика по предложениям\n",
    "\n",
    "- Посчитаем кол-во объектов, долю повторяющихся объектов\n",
    "- Посчитаем распределение по длине вопросов в символах\n",
    "- Посчитаем распределение по длине вопросов в словах"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Длина выборки: 404290\n",
      "Количество класса 1 (дубликатов в вопросах): 36.92%\n",
      "Всего уникальных вопросов в обучающей выборке: 537933\n",
      "Количество вопросов, которые повторяются в обучении больше одного раза: 111780\n",
      "()\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAtMAAAFKCAYAAADfddMnAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nAMAiwjQAAABgEWEaAAAAsOj/AwuV46Ljsb96AAAAAElFTkSuQmCC\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x104950f0>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "print('Длина выборки: {}'.format(len(data)))\n",
    "print('Количество класса 1 (дубликатов в вопросах): {}%'.format(round(data['is_duplicate'].mean()*100, 2)))\n",
    "\n",
    "qids = pd.Series(data['qid1'].tolist() + data['qid2'].tolist())\n",
    "print('Всего уникальных вопросов в обучающей выборке: {}'.format(len(np.unique(qids))))\n",
    "print('Количество вопросов, которые повторяются в обучении больше одного раза: {}'.format(np.sum(qids.value_counts() > 1)))\n",
    "\n",
    "plt.figure(figsize=(12, 5))\n",
    "plt.hist(qids.value_counts(), bins=50)\n",
    "plt.yscale('log', nonposy='clip')\n",
    "plt.title('Log-Histogram of question appearance counts')\n",
    "plt.xlabel('Number of occurences of question')\n",
    "plt.ylabel('Number of questions')\n",
    "print()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Итак, меньше 100 вопросов повторяются более 30 раз"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "mean-train 59.86 std-train 31.99  max-train 1169.00 \n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAt4AAAFRCAYAAABDmqrPAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x12c0fcc0>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "train_qs = pd.Series(data['question1'].tolist() + data['question2'].tolist()).astype(str)\n",
    "\n",
    "dist_train = train_qs.apply(len)\n",
    "plt.figure(figsize=(12, 5))\n",
    "plt.hist(dist_train, bins=200, range=[0, 200], color=pal[2], normed=True, label='train')\n",
    "plt.title('Normalised histogram of character count in questions', fontsize=15)\n",
    "plt.legend()\n",
    "plt.xlabel('Number of characters', fontsize=15)\n",
    "plt.ylabel('Probability', fontsize=15)\n",
    "\n",
    "print('mean-train {:.2f} std-train {:.2f}  max-train {:.2f} '.format(dist_train.mean(), \n",
    "                          dist_train.std(),   dist_train.max(), ))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "видно, что больше всего коротких вопросов, длина которых 50 знаков. Также у Quora стоит ограничение в 150 знаков. Остальные выбросы - это из времен, когда ограничения не было"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "mean-train 11.06 std-train 5.89 max-train 237.00 \n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAtgAAAFRCAYAAABOhNqIAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\neqsn0QAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x175ba400>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "dist_train = train_qs.apply(lambda x: len(x.split(' ')))\n",
    "plt.figure(figsize=(12, 5))\n",
    "plt.hist(dist_train, bins=50, range=[0, 50], color=pal[2], normed=True, label='train')\n",
    "plt.title('Normalised histogram of word count in questions', fontsize=15)\n",
    "plt.legend()\n",
    "plt.xlabel('Number of words', fontsize=15)\n",
    "plt.ylabel('Probability', fontsize=15)\n",
    "\n",
    "print('mean-train {:.2f} std-train {:.2f} max-train {:.2f} '.format(dist_train.mean(), \n",
    "                          dist_train.std(), dist_train.max()))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "количество слов в вопросе по распределению сильно похоже на количество символов (что объясняется средней длиной слова). Но нет ограничения на количество слов. Распределение более сглажено с правого конца"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 2. Feature generation\n",
    "Признаки, которые я хочу сгенерировать основаны на том, что необходимо выявить различия и схожесть между вопросами.\n",
    "Нельзя пытаться уловить весь смысл всех сообщений, построив модель на одной такой выборке. Нужно ввести критерии схожести и различия между вопросами.\n",
    "\n",
    "Важным фактом является то, что признаки должны быть инвариантны по отношению к перестановке двух вопросов. Например, при обработке двух вопросов они должныы быть упорядочены по определенному критерию, который заранее задан.\n",
    "\n",
    "Предлагаемым критерием служит длина вопроса в символах. Первым вопросом считается более короткий, вторым - более длинный. Тогда для пары вопросов (q1,q2) признаковое описание будет совпадать с признаковым описанием для пары (q2,q1).\n",
    "\n",
    "Таким образом, предлагаются в первом приближении следующие признаки:\n",
    "\n",
    "Для символов:\n",
    "- Длина 1 и 2 вопроса в символах\n",
    "- Доля множества пересекающихся символов для 1 вопроса\n",
    "- Доля множества символов, входящих в 1, за исключением множества пересекающихся (аналогично для 2 вопроса)\n",
    "\n",
    "Для отдельных слов:\n",
    "- Длина 1 и 2 вопроса в словах\n",
    "- Доля множества пересекающихся символов для 1 вопроса\n",
    "- Доля множества слов, входящих в 1, за исключением множества пересекающихся (аналогично для 2 вопроса)\n",
    "\n",
    "Для словаря частот:\n",
    "- Среднее, медианное, минимальное и максимальное значение частот слов из 1 вопроса, не входящих в пересечение\n",
    "\n",
    "Производные функции:\n",
    "- R = (len(w1 not in stops, in w2) + len(w2 not in stops, in w1))/(len(w1) + len(w2))\n",
    "- То же самое для tf_idf "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Сделаем словарь частот слов"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "from collections import Counter\n",
    "\n",
    "# If a word appears only once, we ignore it completely (likely a typo)\n",
    "# Epsilon defines a smoothing constant, which makes the effect of extremely rare words smaller\n",
    "def get_weight(count, eps=10000, min_count=2):\n",
    "    if count < min_count:\n",
    "        return 0.\n",
    "    else:\n",
    "        return 1. / float(count + eps)\n",
    "\n",
    "eps = 5000 \n",
    "words = (\" \".join(train_qs)).lower().split()\n",
    "counts = Counter(words)\n",
    "weights = {word: get_weight(count) for word, count in counts.items()}"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Выведем наиболее популярные в употребления слова, а затем наиболее редкие"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {
    "collapsed": false,
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Most common words and weights: \n",
      "\n",
      "[('the', 2.5891040146646852e-06), ('what', 3.115623919267953e-06), ('is', 3.5861702928825277e-06), ('how', 4.366449945201053e-06), ('i', 4.4805878531263305e-06), ('a', 4.540645588989843e-06), ('to', 4.671434644293609e-06), ('in', 4.884625153865692e-06), ('of', 5.920242493132519e-06), ('do', 6.070908207867897e-06)]\n",
      "\n",
      "Least common words and weights: \n",
      "\n",
      "[('kernel:', 9.998000399920016e-05), ('chameleons', 9.998000399920016e-05), ('275', 9.998000399920016e-05), ('276', 9.998000399920016e-05), ('kid:', 9.998000399920016e-05), ('usenet', 9.998000399920016e-05), ('(football', 9.998000399920016e-05), ('sneha', 9.998000399920016e-05), ('population)', 9.998000399920016e-05), (',delhi', 9.998000399920016e-05)]\n"
     ]
    }
   ],
   "source": [
    "print('Most common words and weights: \\n')\n",
    "print(sorted(weights.items(), key=lambda x: x[1] if x[1] > 0 else 9999)[:10])\n",
    "print('\\nLeast common words and weights: \\n')\n",
    "print(sorted(weights.items(), key=lambda x: x[1], reverse=True)[:10])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Видно, что наиболее редкие слова - это слова, которые смерджились со знаками препинания, опечатками, и цифры, то есть слова, употребленные случайным образом один раз во всей выборке"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Выбор признаков по символам:\n",
    "- Длина 1 и 2 вопроса в символах\n",
    "- Доля множества пересекающихся символов для 1 вопроса\n",
    "- Доля множества символов, входящих в 1, за исключением множества пересекающихся (аналогично для 2 вопроса)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "from collections import Counter as mset\n",
    "\n",
    "def get_features_based_on_symbols(row):\n",
    "    len_1 = len(row['question1'])\n",
    "    len_2 = len(row['question2'])\n",
    "    if len_1 <= len_2:\n",
    "        row_1 = row['question1'].lower()\n",
    "        row_2 = row['question2'].lower()\n",
    "    else:\n",
    "        row_1 = row['question2'].lower()\n",
    "        row_2 = row['question1'].lower()\n",
    "    \n",
    "    symbols_1 = list(row_1)\n",
    "    symbols_2 = list(row_2)\n",
    "    symbols_intersection = mset(symbols_1) & mset(symbols_2)\n",
    "    intersection = list(symbols_intersection.elements())\n",
    "    \n",
    "    c = mset(list(intersection))\n",
    "    c1 = mset(symbols_1)\n",
    "    c2 = mset(symbols_2)\n",
    "    \n",
    "    diff_1 = c1 - c\n",
    "    diff_2 = c2 - c\n",
    "    \n",
    "    uniq_1 = len(list(diff_1.elements()))/float(len_1)\n",
    "    uniq_2 = len(list(diff_2.elements()))/float(len_2)\n",
    "    \n",
    "    uniq_symb_1_not_intercept = sum([x not in intersection for x in list(symbols_1)])/float(len_1)\n",
    "    uniq_symb_2_not_intercept = sum([x not in intersection for x in list(symbols_2)])/float(len_2)\n",
    "    \n",
    "    return pd.Series((len_1, len_2, uniq_1, uniq_2, uniq_symb_1_not_intercept, uniq_symb_2_not_intercept))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>0</th>\n",
       "      <th>1</th>\n",
       "      <th>2</th>\n",
       "      <th>3</th>\n",
       "      <th>4</th>\n",
       "      <th>5</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>66.0</td>\n",
       "      <td>57.0</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.157895</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>51.0</td>\n",
       "      <td>88.0</td>\n",
       "      <td>0.039216</td>\n",
       "      <td>0.443182</td>\n",
       "      <td>0.019608</td>\n",
       "      <td>0.102273</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>73.0</td>\n",
       "      <td>59.0</td>\n",
       "      <td>0.109589</td>\n",
       "      <td>0.372881</td>\n",
       "      <td>0.041096</td>\n",
       "      <td>0.067797</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>50.0</td>\n",
       "      <td>65.0</td>\n",
       "      <td>0.340000</td>\n",
       "      <td>0.492308</td>\n",
       "      <td>0.180000</td>\n",
       "      <td>0.369231</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>76.0</td>\n",
       "      <td>39.0</td>\n",
       "      <td>0.039474</td>\n",
       "      <td>1.025641</td>\n",
       "      <td>0.013158</td>\n",
       "      <td>0.230769</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>5</th>\n",
       "      <td>86.0</td>\n",
       "      <td>90.0</td>\n",
       "      <td>0.116279</td>\n",
       "      <td>0.155556</td>\n",
       "      <td>0.069767</td>\n",
       "      <td>0.044444</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>6</th>\n",
       "      <td>19.0</td>\n",
       "      <td>62.0</td>\n",
       "      <td>0.210526</td>\n",
       "      <td>0.758065</td>\n",
       "      <td>0.210526</td>\n",
       "      <td>0.419355</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>7</th>\n",
       "      <td>30.0</td>\n",
       "      <td>41.0</td>\n",
       "      <td>0.066667</td>\n",
       "      <td>0.317073</td>\n",
       "      <td>0.066667</td>\n",
       "      <td>0.048780</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>8</th>\n",
       "      <td>35.0</td>\n",
       "      <td>37.0</td>\n",
       "      <td>0.171429</td>\n",
       "      <td>0.216216</td>\n",
       "      <td>0.171429</td>\n",
       "      <td>0.135135</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>9</th>\n",
       "      <td>60.0</td>\n",
       "      <td>49.0</td>\n",
       "      <td>0.133333</td>\n",
       "      <td>0.387755</td>\n",
       "      <td>0.050000</td>\n",
       "      <td>0.122449</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "      0     1         2         3         4         5\n",
       "0  66.0  57.0  0.000000  0.157895  0.000000  0.000000\n",
       "1  51.0  88.0  0.039216  0.443182  0.019608  0.102273\n",
       "2  73.0  59.0  0.109589  0.372881  0.041096  0.067797\n",
       "3  50.0  65.0  0.340000  0.492308  0.180000  0.369231\n",
       "4  76.0  39.0  0.039474  1.025641  0.013158  0.230769\n",
       "5  86.0  90.0  0.116279  0.155556  0.069767  0.044444\n",
       "6  19.0  62.0  0.210526  0.758065  0.210526  0.419355\n",
       "7  30.0  41.0  0.066667  0.317073  0.066667  0.048780\n",
       "8  35.0  37.0  0.171429  0.216216  0.171429  0.135135\n",
       "9  60.0  49.0  0.133333  0.387755  0.050000  0.122449"
      ]
     },
     "execution_count": 11,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "data.iloc[0:10,:].apply(get_features_based_on_symbols,axis=1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "выбор признаков по отдельным словам:\n",
    "- Длина 1 и 2 вопроса в словах\n",
    "- Доля множества пересекающихся символов для 1 вопроса\n",
    "- Доля множества слов, входящих в 1, за исключением множества пересекающихся (аналогично для 2 вопроса)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def get_features_based_on_words(row):\n",
    "    len_1 = len(row['question1'])\n",
    "    len_2 = len(row['question2'])\n",
    "    if len_1 <= len_2:\n",
    "        row_1 = row['question1'].lower()\n",
    "        row_2 = row['question2'].lower()\n",
    "    else:\n",
    "        row_1 = row['question2'].lower()\n",
    "        row_2 = row['question1'].lower()\n",
    "    \n",
    "    words_1 = row_1.split()\n",
    "    words_2 = row_2.split()\n",
    "    words_intersection = mset(words_1) & mset(words_2)\n",
    "    intersection = list(words_intersection.elements())\n",
    "    \n",
    "    c = mset(list(intersection))\n",
    "    c1 = mset(list(words_1))\n",
    "    c2 = mset(list(words_2))\n",
    "    \n",
    "    diff_1 = c1 - c\n",
    "    diff_2 = c2 - c\n",
    "    \n",
    "    uniq_1 = len(list(diff_1.elements()))/float(len(words_1))\n",
    "    uniq_2 = len(list(diff_2.elements()))/float(len(words_2))\n",
    "    \n",
    "    uniq_w_1_not_intercept = sum([x not in intersection for x in list(words_1)])/float(len(words_1))\n",
    "    uniq_w_2_not_intercept = sum([x not in intersection for x in list(words_2)])/float(len(words_2))\n",
    "    \n",
    "    return pd.Series((len(words_1), len(words_2), uniq_1, uniq_2, uniq_w_1_not_intercept, uniq_w_2_not_intercept))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>0</th>\n",
       "      <th>1</th>\n",
       "      <th>2</th>\n",
       "      <th>3</th>\n",
       "      <th>4</th>\n",
       "      <th>5</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>12.0</td>\n",
       "      <td>14.0</td>\n",
       "      <td>0.083333</td>\n",
       "      <td>0.214286</td>\n",
       "      <td>0.083333</td>\n",
       "      <td>0.142857</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>8.0</td>\n",
       "      <td>13.0</td>\n",
       "      <td>0.500000</td>\n",
       "      <td>0.692308</td>\n",
       "      <td>0.500000</td>\n",
       "      <td>0.615385</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>10.0</td>\n",
       "      <td>14.0</td>\n",
       "      <td>0.600000</td>\n",
       "      <td>0.714286</td>\n",
       "      <td>0.600000</td>\n",
       "      <td>0.714286</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>11.0</td>\n",
       "      <td>9.0</td>\n",
       "      <td>1.000000</td>\n",
       "      <td>1.000000</td>\n",
       "      <td>1.000000</td>\n",
       "      <td>1.000000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>7.0</td>\n",
       "      <td>13.0</td>\n",
       "      <td>0.714286</td>\n",
       "      <td>0.846154</td>\n",
       "      <td>0.714286</td>\n",
       "      <td>0.846154</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>5</th>\n",
       "      <td>16.0</td>\n",
       "      <td>16.0</td>\n",
       "      <td>0.500000</td>\n",
       "      <td>0.500000</td>\n",
       "      <td>0.500000</td>\n",
       "      <td>0.500000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>6</th>\n",
       "      <td>4.0</td>\n",
       "      <td>11.0</td>\n",
       "      <td>1.000000</td>\n",
       "      <td>1.000000</td>\n",
       "      <td>1.000000</td>\n",
       "      <td>1.000000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>7</th>\n",
       "      <td>7.0</td>\n",
       "      <td>9.0</td>\n",
       "      <td>0.428571</td>\n",
       "      <td>0.555556</td>\n",
       "      <td>0.428571</td>\n",
       "      <td>0.555556</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>8</th>\n",
       "      <td>8.0</td>\n",
       "      <td>8.0</td>\n",
       "      <td>0.250000</td>\n",
       "      <td>0.250000</td>\n",
       "      <td>0.250000</td>\n",
       "      <td>0.250000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>9</th>\n",
       "      <td>9.0</td>\n",
       "      <td>9.0</td>\n",
       "      <td>0.666667</td>\n",
       "      <td>0.666667</td>\n",
       "      <td>0.666667</td>\n",
       "      <td>0.666667</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "      0     1         2         3         4         5\n",
       "0  12.0  14.0  0.083333  0.214286  0.083333  0.142857\n",
       "1   8.0  13.0  0.500000  0.692308  0.500000  0.615385\n",
       "2  10.0  14.0  0.600000  0.714286  0.600000  0.714286\n",
       "3  11.0   9.0  1.000000  1.000000  1.000000  1.000000\n",
       "4   7.0  13.0  0.714286  0.846154  0.714286  0.846154\n",
       "5  16.0  16.0  0.500000  0.500000  0.500000  0.500000\n",
       "6   4.0  11.0  1.000000  1.000000  1.000000  1.000000\n",
       "7   7.0   9.0  0.428571  0.555556  0.428571  0.555556\n",
       "8   8.0   8.0  0.250000  0.250000  0.250000  0.250000\n",
       "9   9.0   9.0  0.666667  0.666667  0.666667  0.666667"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "data.iloc[0:10,:].apply(get_features_based_on_words,axis=1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Работа над признаками, содержащими частоты слов\n",
    "\n",
    "- Среднее, медианное, минимальное и максимальное значение частот слов из 1 вопроса, не входящих в пересечение"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "def get_features_based_on_tf_idf_words(row):\n",
    "    len_1 = len(row['question1'])\n",
    "    len_2 = len(row['question2'])\n",
    "    if len_1 <= len_2:\n",
    "        row_1 = row['question1'].lower()\n",
    "        row_2 = row['question2'].lower()\n",
    "    else:\n",
    "        row_1 = row['question2'].lower()\n",
    "        row_2 = row['question1'].lower()\n",
    "    \n",
    "    words_1 = row_1.split()\n",
    "    words_2 = row_2.split()\n",
    "    words_intersection = mset(words_1) & mset(words_2)\n",
    "    intersection = list(words_intersection.elements())\n",
    "    \n",
    "    c = mset(list(intersection))\n",
    "    c1 = mset(list(words_1))\n",
    "    c2 = mset(list(words_2))\n",
    "    \n",
    "    diff_1 = c1 - c\n",
    "    diff_2 = c2 - c\n",
    "    \n",
    "    uniq_1 = len(list(diff_1.elements()))/float(len(words_1))\n",
    "    uniq_2 = len(list(diff_2.elements()))/float(len(words_2))\n",
    "    \n",
    "    uniq_w_1_not_intercept = sum([x not in intersection for x in list(words_1)])/float(len(words_1))\n",
    "    uniq_w_2_not_intercept = sum([x not in intersection for x in list(words_2)])/float(len(words_2))\n",
    "    \n",
    "    if len([x for x in list(words_2) if x not in intersection ])!=0:\n",
    "        weights_1 = [weights[w] for w in [x for x in list(words_1) if x not in intersection ]]\n",
    "    else:\n",
    "        weights_1 = [0]\n",
    "    if len([x for x in list(words_2) if x not in intersection ])!=0:\n",
    "        weights_2 = [weights[w] for w in [x for x in list(words_2) if x not in intersection ]]\n",
    "    else:\n",
    "        weights_2 = [0]\n",
    "    try:\n",
    "        mean_1 = np.mean(weights_1)\n",
    "        median_1 = np.median(weights_1)\n",
    "        max_1 = np.max(weights_1)\n",
    "        min_1 = np.min(weights_1)\n",
    "    except:\n",
    "        mean_1 = 0\n",
    "        median_1 = 0\n",
    "        max_1 = 0\n",
    "        min_1 = 0\n",
    "    try:\n",
    "        mean_2 = np.mean(weights_2)\n",
    "        median_2 = np.median(weights_2)\n",
    "        max_2 = np.max(weights_2)\n",
    "        min_2 = np.min(weights_2)\n",
    "    except:\n",
    "        mean_2 = 0\n",
    "        median_2 = 0\n",
    "        max_2 = 0\n",
    "        min_2 = 0\n",
    "    return pd.Series((mean_1, median_1, max_1, min_1, mean_2, median_2, max_2, min_2))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>0</th>\n",
       "      <th>1</th>\n",
       "      <th>2</th>\n",
       "      <th>3</th>\n",
       "      <th>4</th>\n",
       "      <th>5</th>\n",
       "      <th>6</th>\n",
       "      <th>7</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>0.000091</td>\n",
       "      <td>0.000091</td>\n",
       "      <td>0.000091</td>\n",
       "      <td>0.000091</td>\n",
       "      <td>0.000063</td>\n",
       "      <td>0.000063</td>\n",
       "      <td>0.000088</td>\n",
       "      <td>0.000038</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>0.000050</td>\n",
       "      <td>0.000048</td>\n",
       "      <td>0.000100</td>\n",
       "      <td>0.000004</td>\n",
       "      <td>0.000067</td>\n",
       "      <td>0.000073</td>\n",
       "      <td>0.000100</td>\n",
       "      <td>0.000019</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>0.000070</td>\n",
       "      <td>0.000084</td>\n",
       "      <td>0.000100</td>\n",
       "      <td>0.000019</td>\n",
       "      <td>0.000043</td>\n",
       "      <td>0.000037</td>\n",
       "      <td>0.000099</td>\n",
       "      <td>0.000003</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>0.000045</td>\n",
       "      <td>0.000044</td>\n",
       "      <td>0.000099</td>\n",
       "      <td>0.000004</td>\n",
       "      <td>0.000035</td>\n",
       "      <td>0.000028</td>\n",
       "      <td>0.000099</td>\n",
       "      <td>0.000000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>0.000082</td>\n",
       "      <td>0.000094</td>\n",
       "      <td>0.000097</td>\n",
       "      <td>0.000030</td>\n",
       "      <td>0.000083</td>\n",
       "      <td>0.000100</td>\n",
       "      <td>0.000100</td>\n",
       "      <td>0.000007</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>5</th>\n",
       "      <td>0.000058</td>\n",
       "      <td>0.000069</td>\n",
       "      <td>0.000100</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000052</td>\n",
       "      <td>0.000053</td>\n",
       "      <td>0.000100</td>\n",
       "      <td>0.000000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>6</th>\n",
       "      <td>0.000022</td>\n",
       "      <td>0.000012</td>\n",
       "      <td>0.000061</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000052</td>\n",
       "      <td>0.000067</td>\n",
       "      <td>0.000097</td>\n",
       "      <td>0.000000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>7</th>\n",
       "      <td>0.000014</td>\n",
       "      <td>0.000008</td>\n",
       "      <td>0.000030</td>\n",
       "      <td>0.000004</td>\n",
       "      <td>0.000023</td>\n",
       "      <td>0.000006</td>\n",
       "      <td>0.000080</td>\n",
       "      <td>0.000003</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>8</th>\n",
       "      <td>0.000100</td>\n",
       "      <td>0.000100</td>\n",
       "      <td>0.000100</td>\n",
       "      <td>0.000100</td>\n",
       "      <td>0.000100</td>\n",
       "      <td>0.000100</td>\n",
       "      <td>0.000100</td>\n",
       "      <td>0.000100</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>9</th>\n",
       "      <td>0.000030</td>\n",
       "      <td>0.000008</td>\n",
       "      <td>0.000092</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000037</td>\n",
       "      <td>0.000010</td>\n",
       "      <td>0.000100</td>\n",
       "      <td>0.000000</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "          0         1         2         3         4         5         6  \\\n",
       "0  0.000091  0.000091  0.000091  0.000091  0.000063  0.000063  0.000088   \n",
       "1  0.000050  0.000048  0.000100  0.000004  0.000067  0.000073  0.000100   \n",
       "2  0.000070  0.000084  0.000100  0.000019  0.000043  0.000037  0.000099   \n",
       "3  0.000045  0.000044  0.000099  0.000004  0.000035  0.000028  0.000099   \n",
       "4  0.000082  0.000094  0.000097  0.000030  0.000083  0.000100  0.000100   \n",
       "5  0.000058  0.000069  0.000100  0.000000  0.000052  0.000053  0.000100   \n",
       "6  0.000022  0.000012  0.000061  0.000000  0.000052  0.000067  0.000097   \n",
       "7  0.000014  0.000008  0.000030  0.000004  0.000023  0.000006  0.000080   \n",
       "8  0.000100  0.000100  0.000100  0.000100  0.000100  0.000100  0.000100   \n",
       "9  0.000030  0.000008  0.000092  0.000000  0.000037  0.000010  0.000100   \n",
       "\n",
       "          7  \n",
       "0  0.000038  \n",
       "1  0.000019  \n",
       "2  0.000003  \n",
       "3  0.000000  \n",
       "4  0.000007  \n",
       "5  0.000000  \n",
       "6  0.000000  \n",
       "7  0.000003  \n",
       "8  0.000100  \n",
       "9  0.000000  "
      ]
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "data.iloc[0:10,:].apply(get_features_based_on_tf_idf_words, axis=1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Теперь заполним все пропущенные значения нулями и применим полученные преобразования"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "data_2 = data.fillna('zero')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "C:\\Users\\agoncharov\\Anaconda2\\lib\\site-packages\\numpy\\core\\_methods.py:59: RuntimeWarning: Mean of empty slice.\n",
      "  warnings.warn(\"Mean of empty slice.\", RuntimeWarning)\n",
      "C:\\Users\\agoncharov\\Anaconda2\\lib\\site-packages\\numpy\\core\\_methods.py:70: RuntimeWarning: invalid value encountered in double_scalars\n",
      "  ret = ret.dtype.type(ret / rcount)\n"
     ]
    }
   ],
   "source": [
    "X = pd.concat((\n",
    "    data_2.apply(get_features_based_on_symbols,axis=1),\n",
    "    data_2.apply(get_features_based_on_words,axis=1),\n",
    "    data_2.apply(get_features_based_on_tf_idf_words, axis=1)\n",
    "        ),axis=1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Переименуем все значения в интерпретируемые"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "X.columns = ['len_1', 'len_2', 'uniq_1_s', 'uniq_2_s', 'uniq_symb_1_not_intercept', 'uniq_symb_2_not_intercept', \n",
    "            'len(words_1)', 'len(words_2)', 'uniq_1_w', 'uniq_2_w', 'uniq_w_1_not_intercept', 'uniq_w_2_not_intercept',\n",
    "            'mean_1', 'median_1', 'max_1', 'min_1', 'mean_2', 'median_2', 'max_2', 'min_2']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>len_1</th>\n",
       "      <th>len_2</th>\n",
       "      <th>uniq_1_s</th>\n",
       "      <th>uniq_2_s</th>\n",
       "      <th>uniq_symb_1_not_intercept</th>\n",
       "      <th>uniq_symb_2_not_intercept</th>\n",
       "      <th>len(words_1)</th>\n",
       "      <th>len(words_2)</th>\n",
       "      <th>uniq_1_w</th>\n",
       "      <th>uniq_2_w</th>\n",
       "      <th>uniq_w_1_not_intercept</th>\n",
       "      <th>uniq_w_2_not_intercept</th>\n",
       "      <th>mean_1</th>\n",
       "      <th>median_1</th>\n",
       "      <th>max_1</th>\n",
       "      <th>min_1</th>\n",
       "      <th>mean_2</th>\n",
       "      <th>median_2</th>\n",
       "      <th>max_2</th>\n",
       "      <th>min_2</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>66.0</td>\n",
       "      <td>57.0</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.157895</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>0.000000</td>\n",
       "      <td>12.0</td>\n",
       "      <td>14.0</td>\n",
       "      <td>0.083333</td>\n",
       "      <td>0.214286</td>\n",
       "      <td>0.083333</td>\n",
       "      <td>0.142857</td>\n",
       "      <td>0.000091</td>\n",
       "      <td>0.000091</td>\n",
       "      <td>0.000091</td>\n",
       "      <td>0.000091</td>\n",
       "      <td>0.000063</td>\n",
       "      <td>0.000063</td>\n",
       "      <td>0.000088</td>\n",
       "      <td>0.000038</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>51.0</td>\n",
       "      <td>88.0</td>\n",
       "      <td>0.039216</td>\n",
       "      <td>0.443182</td>\n",
       "      <td>0.019608</td>\n",
       "      <td>0.102273</td>\n",
       "      <td>8.0</td>\n",
       "      <td>13.0</td>\n",
       "      <td>0.500000</td>\n",
       "      <td>0.692308</td>\n",
       "      <td>0.500000</td>\n",
       "      <td>0.615385</td>\n",
       "      <td>0.000050</td>\n",
       "      <td>0.000048</td>\n",
       "      <td>0.000100</td>\n",
       "      <td>0.000004</td>\n",
       "      <td>0.000067</td>\n",
       "      <td>0.000073</td>\n",
       "      <td>0.000100</td>\n",
       "      <td>0.000019</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>73.0</td>\n",
       "      <td>59.0</td>\n",
       "      <td>0.109589</td>\n",
       "      <td>0.372881</td>\n",
       "      <td>0.041096</td>\n",
       "      <td>0.067797</td>\n",
       "      <td>10.0</td>\n",
       "      <td>14.0</td>\n",
       "      <td>0.600000</td>\n",
       "      <td>0.714286</td>\n",
       "      <td>0.600000</td>\n",
       "      <td>0.714286</td>\n",
       "      <td>0.000070</td>\n",
       "      <td>0.000084</td>\n",
       "      <td>0.000100</td>\n",
       "      <td>0.000019</td>\n",
       "      <td>0.000043</td>\n",
       "      <td>0.000037</td>\n",
       "      <td>0.000099</td>\n",
       "      <td>0.000003</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>50.0</td>\n",
       "      <td>65.0</td>\n",
       "      <td>0.340000</td>\n",
       "      <td>0.492308</td>\n",
       "      <td>0.180000</td>\n",
       "      <td>0.369231</td>\n",
       "      <td>11.0</td>\n",
       "      <td>9.0</td>\n",
       "      <td>1.000000</td>\n",
       "      <td>1.000000</td>\n",
       "      <td>1.000000</td>\n",
       "      <td>1.000000</td>\n",
       "      <td>0.000045</td>\n",
       "      <td>0.000044</td>\n",
       "      <td>0.000099</td>\n",
       "      <td>0.000004</td>\n",
       "      <td>0.000035</td>\n",
       "      <td>0.000028</td>\n",
       "      <td>0.000099</td>\n",
       "      <td>0.000000</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>76.0</td>\n",
       "      <td>39.0</td>\n",
       "      <td>0.039474</td>\n",
       "      <td>1.025641</td>\n",
       "      <td>0.013158</td>\n",
       "      <td>0.230769</td>\n",
       "      <td>7.0</td>\n",
       "      <td>13.0</td>\n",
       "      <td>0.714286</td>\n",
       "      <td>0.846154</td>\n",
       "      <td>0.714286</td>\n",
       "      <td>0.846154</td>\n",
       "      <td>0.000082</td>\n",
       "      <td>0.000094</td>\n",
       "      <td>0.000097</td>\n",
       "      <td>0.000030</td>\n",
       "      <td>0.000083</td>\n",
       "      <td>0.000100</td>\n",
       "      <td>0.000100</td>\n",
       "      <td>0.000007</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "   len_1  len_2  uniq_1_s  uniq_2_s  uniq_symb_1_not_intercept  \\\n",
       "0   66.0   57.0  0.000000  0.157895                   0.000000   \n",
       "1   51.0   88.0  0.039216  0.443182                   0.019608   \n",
       "2   73.0   59.0  0.109589  0.372881                   0.041096   \n",
       "3   50.0   65.0  0.340000  0.492308                   0.180000   \n",
       "4   76.0   39.0  0.039474  1.025641                   0.013158   \n",
       "\n",
       "   uniq_symb_2_not_intercept  len(words_1)  len(words_2)  uniq_1_w  uniq_2_w  \\\n",
       "0                   0.000000          12.0          14.0  0.083333  0.214286   \n",
       "1                   0.102273           8.0          13.0  0.500000  0.692308   \n",
       "2                   0.067797          10.0          14.0  0.600000  0.714286   \n",
       "3                   0.369231          11.0           9.0  1.000000  1.000000   \n",
       "4                   0.230769           7.0          13.0  0.714286  0.846154   \n",
       "\n",
       "   uniq_w_1_not_intercept  uniq_w_2_not_intercept    mean_1  median_1  \\\n",
       "0                0.083333                0.142857  0.000091  0.000091   \n",
       "1                0.500000                0.615385  0.000050  0.000048   \n",
       "2                0.600000                0.714286  0.000070  0.000084   \n",
       "3                1.000000                1.000000  0.000045  0.000044   \n",
       "4                0.714286                0.846154  0.000082  0.000094   \n",
       "\n",
       "      max_1     min_1    mean_2  median_2     max_2     min_2  \n",
       "0  0.000091  0.000091  0.000063  0.000063  0.000088  0.000038  \n",
       "1  0.000100  0.000004  0.000067  0.000073  0.000100  0.000019  \n",
       "2  0.000100  0.000019  0.000043  0.000037  0.000099  0.000003  \n",
       "3  0.000099  0.000004  0.000035  0.000028  0.000099  0.000000  \n",
       "4  0.000097  0.000030  0.000083  0.000100  0.000100  0.000007  "
      ]
     },
     "execution_count": 20,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "X.head()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": true
   },
   "source": [
    "Добавим дополнительные hacks из kaggle kernels, у которых хорошие отзывы"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "from nltk.corpus import stopwords\n",
    "\n",
    "stops = set(stopwords.words(\"english\"))\n",
    "\n",
    "def word_match_share(row):\n",
    "    q1words = {}\n",
    "    q2words = {}\n",
    "    for word in str(row['question1']).lower().split():\n",
    "        if word not in stops:\n",
    "            q1words[word] = 1\n",
    "    for word in str(row['question2']).lower().split():\n",
    "        if word not in stops:\n",
    "            q2words[word] = 1\n",
    "    if len(q1words) == 0 or len(q2words) == 0:\n",
    "        # The computer-generated chaff includes a few questions that are nothing but stopwords\n",
    "        return 0\n",
    "    shared_words_in_q1 = [w for w in q1words.keys() if w in q2words]\n",
    "    shared_words_in_q2 = [w for w in q2words.keys() if w in q1words]\n",
    "    R = (len(shared_words_in_q1) + len(shared_words_in_q2))/float(len(q1words) + len(q2words))\n",
    "    return R\n",
    "\n",
    "\n",
    "def tfidf_word_match_share(row):\n",
    "    q1words = {}\n",
    "    q2words = {}\n",
    "    for word in str(row['question1']).lower().split():\n",
    "        if word not in stops:\n",
    "            q1words[word] = 1\n",
    "    for word in str(row['question2']).lower().split():\n",
    "        if word not in stops:\n",
    "            q2words[word] = 1\n",
    "    if len(q1words) == 0 or len(q2words) == 0:\n",
    "        # The computer-generated chaff includes a few questions that are nothing but stopwords\n",
    "        return 0\n",
    "    \n",
    "    shared_weights = [weights.get(w, 0) for w in q1words.keys() if w in q2words] + [weights.get(w, 0) for w in q2words.keys() if w in q1words]\n",
    "    total_weights = [weights.get(w, 0) for w in q1words] + [weights.get(w, 0) for w in q2words]\n",
    "    \n",
    "    R = np.sum(shared_weights) / float(np.sum(total_weights))\n",
    "    return R"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "C:\\Users\\agoncharov\\Anaconda2\\lib\\site-packages\\ipykernel\\__main__.py:39: RuntimeWarning: invalid value encountered in double_scalars\n"
     ]
    }
   ],
   "source": [
    "train_word_match = data.apply(word_match_share, axis=1, raw=True)\n",
    "tfidf_train_word_match = data.apply(tfidf_word_match_share, axis=1, raw=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "X_additional = pd.concat((\n",
    "                    X,\n",
    "                    train_word_match,\n",
    "                    tfidf_train_word_match\n",
    "                        ),axis=1)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "y = data['is_duplicate'].values"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "X_additional = X_additional.fillna(0)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Классификация по полученным признакам"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Для начала нормируем все значения на 0 - 1"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "from sklearn.preprocessing import MinMaxScaler\n",
    "\n",
    "X_norm = MinMaxScaler().fit_transform(X_additional)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Разделим всю выборку на обучение и контроль. В соотношении 3:7"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "from sklearn.cross_validation import train_test_split\n",
    "\n",
    "x_train, x_valid, y_train, y_valid = train_test_split(X_norm, y, test_size=0.3, random_state=4242)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": false
   },
   "source": [
    "Произведем классификацию при помощи различных классификаторов:"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### XGBoost\n",
    "Вычисления производились на рабочей машине, где стоит windows, так что данный ноутбук без данного раздела"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "#import xgboost as xgb\n",
    "# # Set our parameters for xgboost\n",
    "#params = {}\n",
    "#params['objective'] = 'binary:logistic'\n",
    "#params['eval_metric'] = 'logloss'\n",
    "#params['eta'] = 0.02\n",
    "#params['max_depth'] = 4\n",
    "\n",
    "#d_train = xgb.DMatrix(x_train, label=y_train)\n",
    "#d_valid = xgb.DMatrix(x_valid, label=y_valid)\n",
    "\n",
    "#watchlist = [(d_train, 'train'), (d_valid, 'valid')]\n",
    "\n",
    "#bst = xgb.train(params, d_train, 400, watchlist, early_stopping_rounds=50, verbose_eval=10)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### RandomForest"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 56,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "from sklearn.ensemble import RandomForestClassifier\n",
    "from sklearn.metrics import accuracy_score\n",
    "from sklearn.metrics import log_loss\n",
    "\n",
    "clf = RandomForestClassifier()\n",
    "clf.fit(x_train, y_train)\n",
    "y_pred_proba = clf.predict_proba(x_valid)\n",
    "y_pred = clf.predict(x_valid)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 73,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "y_pred_2 = y_pred.copy()+0.\n",
    "alpha = 0.2\n",
    "y_pred_2[y_pred_2 < alpha] = alpha\n",
    "y_pred_2[y_pred_2 > 1-alpha] = 1-alpha"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 75,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.769010693644\n",
      "0.625735594681\n",
      "0.543362724195\n"
     ]
    }
   ],
   "source": [
    "print accuracy_score(y_valid,y_pred)\n",
    "print log_loss(y_valid,y_pred_proba)\n",
    "print log_loss(y_valid,y_pred_2)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Logistic Regression"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 77,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Fitting 3 folds for each of 6 candidates, totalling 18 fits\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "[Parallel(n_jobs=-1)]: Done  18 out of  18 | elapsed:  3.3min finished\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "GridSearchCV(cv=None, error_score='raise',\n",
       "       estimator=LogisticRegression(C=1.0, class_weight=None, dual=False, fit_intercept=True,\n",
       "          intercept_scaling=1, max_iter=100, multi_class='ovr', n_jobs=1,\n",
       "          penalty='l2', random_state=None, solver='liblinear', tol=0.0001,\n",
       "          verbose=0, warm_start=False),\n",
       "       fit_params={}, iid=True, n_jobs=-1,\n",
       "       param_grid={'penalty': ['l1', 'l2'], 'C': [1e-06, 0.001, 1.0]},\n",
       "       pre_dispatch='2*n_jobs', refit=True, return_train_score=True,\n",
       "       scoring='neg_log_loss', verbose=1)"
      ]
     },
     "execution_count": 77,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from sklearn.linear_model import LogisticRegression\n",
    "from sklearn.metrics import precision_recall_curve, auc, roc_curve\n",
    "from sklearn.model_selection import GridSearchCV, train_test_split\n",
    "\n",
    "clf = LogisticRegression()\n",
    "grid = {\n",
    "    'C': [1e-6, 1e-3, 1e0],\n",
    "    'penalty': ['l1', 'l2']\n",
    "}\n",
    "cv = GridSearchCV(clf, grid, scoring='neg_log_loss', n_jobs=-1, verbose=1)\n",
    "cv.fit(X_norm, y)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 78,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "6. Mean validation neg log loss: -0.693 (std: 0.000) - {'penalty': 'l1', 'C': 1e-06}\n",
      "5. Mean validation neg log loss: -0.666 (std: 0.000) - {'penalty': 'l2', 'C': 1e-06}\n",
      "4. Mean validation neg log loss: -0.533 (std: 0.001) - {'penalty': 'l1', 'C': 0.001}\n",
      "3. Mean validation neg log loss: -0.532 (std: 0.001) - {'penalty': 'l2', 'C': 0.001}\n",
      "1. Mean validation neg log loss: -0.513 (std: 0.001) - {'penalty': 'l1', 'C': 1.0}\n",
      "2. Mean validation neg log loss: -0.515 (std: 0.001) - {'penalty': 'l2', 'C': 1.0}\n"
     ]
    }
   ],
   "source": [
    "for i in range(1, len(cv.cv_results_['params'])+1):\n",
    "    rank = cv.cv_results_['rank_test_score'][i-1]\n",
    "    s = cv.cv_results_['mean_test_score'][i-1]\n",
    "    sd = cv.cv_results_['std_test_score'][i-1]\n",
    "    params = cv.cv_results_['params'][i-1]\n",
    "    print(\"{0}. Mean validation neg log loss: {1:.3f} (std: {2:.3f}) - {3}\".format(\n",
    "        rank,\n",
    "        s,\n",
    "        sd,\n",
    "        params\n",
    "    ))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": true
   },
   "source": [
    "### Выводы"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Таким образом, достигнут плохой результат качества, судя по Leader Board. Применение XGBoost Позволит увеличить это качество на 0.1 - 0.2, что выведет результат в середину списка.\n",
    "\n",
    "Были попробованы различные методы выбора признаков.\n",
    "На мой взгляд, их качество достаточно хорошее, но необходимо улучшить предобработку вопросов. Сейчас вся предобработка заключается в приведении к нижнему регистру. Нужно сделать лемматизацию или стемминг, хорошее удаление стоп-слов, можно использовать словари синонимов, что является лучшим решением для данной задачи."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 2",
   "language": "python",
   "name": "python2"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 2
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython2",
   "version": "2.7.13"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}