deep_learning_nlp.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# First, read original data sets."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "from os.path import join as path_join\n",
    "import numpy as np\n",
    "\n",
    "# 'From Group to Individual Labels using Deep Features', Kotzias et. al,. KDD 2015\n",
    "DATA_SET_FOLDER = 'sentiment-labelled-sentences'\n",
    "\n",
    "AMAZON_CELLS_LABELLED = path_join(DATA_SET_FOLDER, 'amazon_cells_labelled.txt')\n",
    "IMDB_LABELLED = path_join(DATA_SET_FOLDER, 'imdb_labelled.txt')\n",
    "YELP_LABELLED = path_join(DATA_SET_FOLDER, 'yelp_labelled.txt')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def load(file_name):\n",
    "    \"\"\"\n",
    "    Read txt file line by line and return lines as a list and sentiments as another list.\n",
    "    \"\"\"\n",
    "    with open(file_name, 'r') as f:\n",
    "        sentences_list = []\n",
    "        reviews_list = []\n",
    "        for line in f:\n",
    "            sentence, review = line.split('\\t')[:2]\n",
    "            # Remove leading and trailing characters, space here.\n",
    "            sentences_list.append(sentence.strip())\n",
    "            reviews_list.append(int(review.strip()))\n",
    "    return sentences_list, reviews_list"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "amazon_sentences_list, amazon_reviews_list = load(AMAZON_CELLS_LABELLED)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(['So there is no way for me to plug it in here in the US unless I go by a converter.',\n",
       "  'Good case, Excellent value.',\n",
       "  'Great for the jawbone.',\n",
       "  'Tied to charger for conversations lasting more than 45 minutes.MAJOR PROBLEMS!!',\n",
       "  'The mic is great.',\n",
       "  'I have to jiggle the plug to get it to line up right to get decent volume.',\n",
       "  'If you have several dozen or several hundred contacts, then imagine the fun of sending each of them one by one.',\n",
       "  'If you are Razr owner...you must have this!',\n",
       "  'Needless to say, I wasted my money.',\n",
       "  'What a waste of money and time!.'],\n",
       " [0, 1, 1, 0, 1, 0, 0, 1, 0, 0])"
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "amazon_sentences_list[:10], amazon_reviews_list[:10]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "imdb_sentences_list, imdb_reviews_list = load(IMDB_LABELLED)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(['A very, very, very slow-moving, aimless movie about a distressed, drifting young man.',\n",
       "  'Not sure who was more lost - the flat characters or the audience, nearly half of whom walked out.',\n",
       "  'Attempting artiness with black & white and clever camera angles, the movie disappointed - became even more ridiculous - as the acting was poor and the plot and lines almost non-existent.',\n",
       "  'Very little music or anything to speak of.',\n",
       "  'The best scene in the movie was when Gerardo is trying to find a song that keeps running through his head.',\n",
       "  \"The rest of the movie lacks art, charm, meaning... If it's about emptiness, it works I guess because it's empty.\",\n",
       "  'Wasted two hours.',\n",
       "  'Saw the movie today and thought it was a good effort, good messages for kids.',\n",
       "  'A bit predictable.',\n",
       "  'Loved the casting of Jimmy Buffet as the science teacher.'],\n",
       " [0, 0, 0, 0, 1, 0, 0, 1, 0, 1])"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "imdb_sentences_list[:10], imdb_reviews_list[:10]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "yelp_sentences_list, yelp_reviews_list = load(YELP_LABELLED)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(['Wow... Loved this place.',\n",
       "  'Crust is not good.',\n",
       "  'Not tasty and the texture was just nasty.',\n",
       "  'Stopped by during the late May bank holiday off Rick Steve recommendation and loved it.',\n",
       "  'The selection on the menu was great and so were the prices.',\n",
       "  'Now I am getting angry and I want my damn pho.',\n",
       "  \"Honeslty it didn't taste THAT fresh.)\",\n",
       "  'The potatoes were like rubber and you could tell they had been made up ahead of time being kept under a warmer.',\n",
       "  'The fries were great too.',\n",
       "  'A great touch.'],\n",
       " [1, 0, 0, 1, 1, 0, 0, 0, 1, 1])"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "yelp_sentences_list[:10], yelp_reviews_list[:10]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Second, combine all three data sets and split into train and test."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "from sklearn.model_selection import train_test_split\n",
    "\n",
    "raw_X = amazon_sentences_list + imdb_sentences_list + yelp_sentences_list\n",
    "raw_y = amazon_reviews_list + imdb_reviews_list + yelp_reviews_list\n",
    "\n",
    "# Randomly split data set into train (70%) and test (30%).\n",
    "raw_X_train, raw_X_test, y_train, y_test = train_test_split(raw_X, raw_y, test_size=0.3, random_state=99)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(2100, 2100, 900, 900)"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(raw_X_train), len(y_train), len(raw_X_test), len(y_test)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(['This movie is well-balanced with comedy and drama and I thoroughly enjoyed myself.',\n",
       "  \"The instructions didn't explain that a microphone jack could be used.\",\n",
       "  \"If you haven't choked in your own vomit by the end (by all the cheap drama and worthless dialogue) you've must have bored yourself to death with this waste of time.\",\n",
       "  'MANNA FROM HEAVEN is a terrific film that is both predictable and unpredictable at the same time.',\n",
       "  'We had so much to say about the place before we walked in that he expected it to be amazing, but was quickly disappointed.',\n",
       "  'Just consider the excellent story, solid acting and look of the film as added bonuses.',\n",
       "  'Not good by any stretch of the imagination.',\n",
       "  'Service was fine and the waitress was friendly.',\n",
       "  'The scallop dish is quite appalling for value as well.',\n",
       "  \"This is hands down the best phone I've ever had.\"],\n",
       " [1, 0, 0, 1, 0, 1, 0, 1, 0, 1],\n",
       " ['Highly recommend for any one who has a blue tooth phone.',\n",
       "  'This particular model would not work with my Motorola Q smartphone.',\n",
       "  'Good show, Samsung.',\n",
       "  'The sound quality is good and functionality is awesome.',\n",
       "  \"So don't go there if you are looking for good food...\",\n",
       "  'The Buffet at Bellagio was far from what I anticipated.',\n",
       "  \"We thought you'd have to venture further away to get good sushi, but this place really hit the spot that night.\",\n",
       "  'If you want to wait for mediocre food and downright terrible service, then this is the place for you.',\n",
       "  'Everything about this film is simply incredible.',\n",
       "  'My sister has one also and she loves it.'],\n",
       " [1, 0, 1, 1, 0, 0, 1, 0, 1, 1])"
      ]
     },
     "execution_count": 11,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "raw_X_train[:10], y_train[:10], raw_X_test[:10], y_test[:10]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Remove non-letters."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# Download nltk corpus whcih includes stop words.\n",
    "# Running this cell will automatically open a new window.\n",
    "# Choose 'All Packages' -> 'stopwords' and click 'Download'.\n",
    "# Finally, close the window.\n",
    "import nltk\n",
    "# TODO, uncomment fo first-time run.\n",
    "# nltk.download()\n",
    "# Or run 'python -m nltk.downloader stopwords' in shell."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "deep learning makes nlp useful true\n",
      "['deep', 'learning', 'makes', 'nlp', 'useful', 'true']\n",
      "['deep', 'learning', 'makes', 'nlp', 'more', 'useful', 'it', 'is', 'true']\n"
     ]
    }
   ],
   "source": [
    "import re\n",
    "from nltk.corpus import stopwords\n",
    "\n",
    "def clean_sentence(sentence, return_sentence=True, remove_stopwords=True):\n",
    "    \"\"\"\n",
    "    Transform a sentence to lower case and\n",
    "    remove non-letters and optionally remove stopwords.\n",
    "    If return_sentence is True, return sentence as a string.\n",
    "    Otherwise, return as a list of words.\n",
    "    \"\"\"\n",
    "    # Replace non-letters with white space.\n",
    "    sentence_no_letters = re.sub('[^a-zA-Z]', ' ', sentence)\n",
    "    # Convert to lower case.\n",
    "    sentence_lower_case = sentence_no_letters.lower()\n",
    "    # Split sentence by white space, tab, newline, return, formfeed.\n",
    "    words = sentence_lower_case.split()\n",
    "    if remove_stopwords:\n",
    "        # Remove stop words. To make things simpler and reduce noise.\n",
    "        # Try without removing stopwords. Or try stemming!\n",
    "        udf_stopwords = set(stopwords.words('english'))\n",
    "    else:\n",
    "        udf_stopwords = None\n",
    "    no_stopwords = [word for word in words if not udf_stopwords or word not in udf_stopwords]\n",
    "    if return_sentence:\n",
    "        # From list of words to single string.\n",
    "        return ' '.join(no_stopwords)\n",
    "    else:\n",
    "        # Return a list of words\n",
    "        return no_stopwords\n",
    "\n",
    "\n",
    "# A simple test.\n",
    "print(clean_sentence('Deep learning makes NLP more USEFUL! It is True 124.'))\n",
    "print(clean_sentence('Deep learning makes NLP more USEFUL! It is True 124.',\n",
    "                     return_sentence=False))\n",
    "print(clean_sentence('Deep learning makes NLP more USEFUL! It is True 124.',\n",
    "                     return_sentence=False, remove_stopwords=False))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Third, compute bag of words on train set."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(['this movie is well balanced with comedy and drama and i thoroughly enjoyed myself',\n",
       "  'the instructions didn t explain that a microphone jack could be used',\n",
       "  'if you haven t choked in your own vomit by the end by all the cheap drama and worthless dialogue you ve must have bored yourself to death with this waste of time',\n",
       "  'manna from heaven is a terrific film that is both predictable and unpredictable at the same time',\n",
       "  'we had so much to say about the place before we walked in that he expected it to be amazing but was quickly disappointed',\n",
       "  'just consider the excellent story solid acting and look of the film as added bonuses',\n",
       "  'not good by any stretch of the imagination',\n",
       "  'service was fine and the waitress was friendly',\n",
       "  'the scallop dish is quite appalling for value as well',\n",
       "  'this is hands down the best phone i ve ever had'],\n",
       " [1, 0, 0, 1, 0, 1, 0, 1, 0, 1],\n",
       " ['highly recommend for any one who has a blue tooth phone',\n",
       "  'this particular model would not work with my motorola q smartphone',\n",
       "  'good show samsung',\n",
       "  'the sound quality is good and functionality is awesome',\n",
       "  'so don t go there if you are looking for good food',\n",
       "  'the buffet at bellagio was far from what i anticipated',\n",
       "  'we thought you d have to venture further away to get good sushi but this place really hit the spot that night',\n",
       "  'if you want to wait for mediocre food and downright terrible service then this is the place for you',\n",
       "  'everything about this film is simply incredible',\n",
       "  'my sister has one also and she loves it'],\n",
       " [1, 0, 1, 1, 0, 0, 1, 0, 1, 1])"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from functools import partial\n",
    "\n",
    "# Transform to lower case and Remove non-letters and stopwords.\n",
    "# Remove words that are less meaningful \n",
    "# It will reduce vocabulary size dramastically.\n",
    "# However, for logistic regression, it will produce a worse model.\n",
    "X_train = map(partial(clean_sentence, remove_stopwords=False), raw_X_train)\n",
    "X_test = map(partial(clean_sentence, remove_stopwords=False), raw_X_test)\n",
    "\n",
    "X_train[:10], y_train[:10], X_test[:10], y_test[:10]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(<2100x4054 sparse matrix of type '<type 'numpy.int64'>'\n",
       " \twith 21866 stored elements in Compressed Sparse Row format>,\n",
       " 4054,\n",
       " [u'aailiyah', u'abhor', u'ability', u'able', u'about', u'above'])"
      ]
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from sklearn.feature_extraction.text import CountVectorizer\n",
    "\n",
    "# Simply leaving the arguments as empty is fine.\n",
    "vectorizer = CountVectorizer()\n",
    "\n",
    "# Compute vocabulary of train data and transform it to bag of words.\n",
    "X_train_bow = vectorizer.fit_transform(X_train)\n",
    "\n",
    "# Check fitted vocabulary dictionary.\n",
    "vocabulary = vectorizer.get_feature_names()\n",
    "vocabulary_size = len(vocabulary)\n",
    "X_train_bow, vocabulary_size, vocabulary[:6]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Fouth, transform test set and fit a classifier."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(<900x4054 sparse matrix of type '<type 'numpy.int64'>'\n",
       " \twith 8452 stored elements in Compressed Sparse Row format>, 8)"
      ]
     },
     "execution_count": 16,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Transform test set to bag of words.\n",
    "X_test_bow = vectorizer.transform(X_test)\n",
    "# Demonstrate element is count, not occurrence indicator.\n",
    "X_test_bow, np.max(X_test_bow)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Logistic regression."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "ground truth, predictions on default threshold, prediction probabilities\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "([1, 0, 1, 1, 0, 0, 1, 0, 1, 1, 1, 0, 1, 1, 0, 0, 1, 1, 0, 1],\n",
       " [1, 0, 1, 1, 0, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 0, 1],\n",
       " array([ 1.65828242, -2.63769067,  1.21343956,  4.61733549, -1.25440738,\n",
       "        -1.43457675,  2.19812722, -3.22527992,  2.49086451,  1.66220455,\n",
       "         1.80753189,  0.18994867,  1.06550332,  4.57109158,  0.43211265,\n",
       "        -2.81319491,  0.24646594,  2.15407803, -1.53610767,  2.86696619]))"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Fit a logistic regression model on train set and make predictions on test data.\n",
    "from sklearn.linear_model import LogisticRegression\n",
    "\n",
    "log_reg = LogisticRegression(solver='lbfgs')\n",
    "\n",
    "# Train.\n",
    "log_reg.fit(X_train_bow, y_train)\n",
    "pred_train = log_reg.predict(X_train_bow)\n",
    "pred_prob_train = log_reg.decision_function(X_train_bow)\n",
    "# Test.\n",
    "pred_test = log_reg.predict(X_test_bow)\n",
    "pred_prob_test = log_reg.decision_function(X_test_bow)\n",
    "print('ground truth, predictions on default threshold, prediction probabilities')\n",
    "y_test[:20], pred_test[:20].tolist(), pred_prob_test[:20]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train accuracy on default threshold: 0.977619047619\n",
      "area under curve: 0.995706118554\n",
      "Test accuracy on default threshold: 0.82\n",
      "area under curve: 0.892611815367\n"
     ]
    }
   ],
   "source": [
    "# Compute classification performance.\n",
    "from sklearn.metrics import accuracy_score, roc_auc_score, roc_curve\n",
    "\n",
    "train_accuracy = accuracy_score(y_train, pred_train)\n",
    "test_accuracy = accuracy_score(y_test, pred_test)\n",
    "train_auc = roc_auc_score(y_train, pred_prob_train)\n",
    "test_auc = roc_auc_score(y_test, pred_prob_test)\n",
    "test_fpr, test_tpr, thresholds = roc_curve(y_test, pred_prob_test)\n",
    "\n",
    "print('Train accuracy on default threshold: {}\\narea under curve: {}'.format(train_accuracy, train_auc))\n",
    "print('Test accuracy on default threshold: {}\\narea under curve: {}'.format(test_accuracy, test_auc))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYoAAAEKCAYAAAAMzhLIAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n2gAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f627a0e1410>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Plot roc.\n",
    "import matplotlib.pylab as plt\n",
    "\n",
    "def plot_auc(fpr, tpr):\n",
    "    figure, ax = plt.subplots(1)\n",
    "    ax.plot(fpr, tpr)\n",
    "    ax.plot([0.0, 1.0], [0.0, 1.0])\n",
    "    ax.set_xlabel('false positive rate')\n",
    "    ax.set_ylabel('true positive rate')\n",
    "    ax.set_xlim([0.0, 1.0])\n",
    "    ax.set_ylim([0.0, 1.0])\n",
    "    plt.show()\n",
    "\n",
    "plot_auc(test_fpr, test_tpr)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Support vector machine."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "ground truth, predictions on default threshold, prediction probabilities\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "([1, 0, 1, 1, 0, 0, 1, 0, 1, 1, 1, 0, 1, 1, 0, 0, 1, 1, 0, 1],\n",
       " [1, 0, 1, 1, 0, 0, 1, 0, 1, 1, 1, 0, 0, 1, 1, 0, 1, 1, 0, 1],\n",
       " array([ 0.32125453, -0.90261352,  0.84132678,  1.50155935, -0.3143413 ,\n",
       "        -0.95283443,  0.19351253, -0.52335077,  0.50071088,  0.94522789,\n",
       "         0.32869754, -0.19940717, -0.11342933,  0.63613904,  0.07898237,\n",
       "        -1.31081585,  0.07816099,  0.56349722, -1.10091062,  0.20735569]))"
      ]
     },
     "execution_count": 20,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Fit an svc model on train set and make predictions on test data.\n",
    "from sklearn.svm import SVC\n",
    "\n",
    "svc = SVC(C=5, gamma=0.1, probability=True, decision_function_shape='ovr')\n",
    "\n",
    "# Train.\n",
    "svc.fit(X_train_bow, y_train)\n",
    "pred_train = svc.predict(X_train_bow)\n",
    "pred_prob_train = svc.decision_function(X_train_bow)\n",
    "# Test.\n",
    "pred_test = svc.predict(X_test_bow)\n",
    "pred_prob_test = svc.decision_function(X_test_bow)\n",
    "print('ground truth, predictions on default threshold, prediction probabilities')\n",
    "y_test[:20], pred_test[:20].tolist(), pred_prob_test[:20]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train accuracy on default threshold: 0.998095238095\n",
      "area under curve: 1.0\n",
      "Test accuracy on default threshold: 0.807777777778\n",
      "area under curve: 0.890888350066\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYoAAAEKCAYAAAAMzhLIAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f627a9b44d0>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Compute classification performance.\n",
    "from sklearn.metrics import accuracy_score, roc_auc_score, roc_curve\n",
    "\n",
    "train_accuracy = accuracy_score(y_train, pred_train)\n",
    "test_accuracy = accuracy_score(y_test, pred_test)\n",
    "train_auc = roc_auc_score(y_train, pred_prob_train)\n",
    "test_auc = roc_auc_score(y_test, pred_prob_test)\n",
    "test_fpr, test_tpr, thresholds = roc_curve(y_test, pred_prob_test)\n",
    "\n",
    "print('Train accuracy on default threshold: {}\\narea under curve: {}'.format(train_accuracy, train_auc))\n",
    "print('Test accuracy on default threshold: {}\\narea under curve: {}'.format(test_accuracy, test_auc))\n",
    "\n",
    "plot_auc(test_fpr, test_tpr)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Neural network."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Minimum test loss 2.29713225365 at step 1000, and training loss 6.10517236055e-06\n",
      "Maximal test auc 0.850016543292 at step 1000, and training auc 1.0\n",
      "Maximal test accuracy 0.773333333333 at step 1000, and training accuracy 1.0\n"
     ]
    }
   ],
   "source": [
    "import tensorflow as tf\n",
    "\n",
    "# Build computation graph.\n",
    "with tf.Graph().as_default() as g:\n",
    "    X = tf.placeholder(tf.float32, shape=[None, vocabulary_size])\n",
    "    y = tf.placeholder(tf.float32, shape=None)\n",
    "    with tf.name_scope('hiddel_layer_1'):\n",
    "        weights = tf.Variable(tf.truncated_normal([vocabulary_size, 1024]))\n",
    "        weights_reg_loss = tf.reduce_mean(tf.square(weights))\n",
    "        biases = tf.Variable(tf.zeros([1024]))\n",
    "        output = tf.matmul(X, weights) + biases\n",
    "        activation = tf.nn.relu(output)\n",
    "    \n",
    "    with tf.variable_scope('softmax'):\n",
    "        weights = tf.Variable(tf.truncated_normal([1024, 1]))\n",
    "        weights_reg_loss += tf.reduce_mean(tf.square(weights))\n",
    "        biases = tf.Variable(0.0)\n",
    "        final_output = tf.squeeze(tf.matmul(activation, weights), axis=1) + biases\n",
    "        pred_prob = tf.nn.sigmoid(final_output)\n",
    "        loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels=y, logits=final_output), axis=0)\n",
    "        final_loss = loss + weights_reg_loss * 0.01\n",
    "    \n",
    "    init_learning_rate = 0.01\n",
    "    opt = tf.train.RMSPropOptimizer(init_learning_rate)\n",
    "    train_op = opt.minimize(final_loss)\n",
    "\n",
    "    init_op = tf.global_variables_initializer()\n",
    "\n",
    "# Create a session to invoke computations.\n",
    "sess = tf.Session(graph=g)\n",
    "\n",
    "sess.run(init_op)\n",
    "\n",
    "# From sparse array to dense array.\n",
    "X_train_bow_arr = X_train_bow.toarray().astype(np.float32)\n",
    "X_test_bow_arr = X_test_bow.toarray().astype(np.float32)\n",
    "y_train_arr = np.array(y_train, dtype=np.float32)\n",
    "y_test_arr = np.array(y_test, dtype=np.float32)\n",
    "\n",
    "loss_train = []\n",
    "loss_test = []\n",
    "auc_train = []\n",
    "auc_test = []\n",
    "accuracy_train = []\n",
    "accuracy_test = []\n",
    "\n",
    "for step in xrange(1001):\n",
    "    if step % 100 == 0:\n",
    "        _, train_loss_val, train_pred_prob_val = sess.run([train_op, loss, pred_prob],\n",
    "                                                          feed_dict={X: X_train_bow_arr, y: y_train})\n",
    "        \n",
    "        loss_train.append(train_loss_val)\n",
    "        auc_train.append(roc_auc_score(y_train, train_pred_prob_val))\n",
    "        accuracy_train.append(accuracy_score(y_train, train_pred_prob_val >= 0.5))\n",
    "        # print('Step: {}, training loss: {}'.format(step, train_loss_val))\n",
    "        test_loss_val, test_pred_prob_val = sess.run([loss, pred_prob],\n",
    "                                                      feed_dict={X: X_test_bow_arr, y: y_test})\n",
    "        loss_test.append(test_loss_val)\n",
    "        auc_test.append(roc_auc_score(y_test, test_pred_prob_val))\n",
    "        accuracy_test.append(accuracy_score(y_test, test_pred_prob_val >= 0.5))\n",
    "        # print('Step: {}, test loss: {}'.format(step, test_loss_val))\n",
    "    else:\n",
    "        _ = sess.run(train_op, feed_dict={X: X_train_bow_arr, y: y_train})\n",
    "\n",
    "# Make some simple statistics.\n",
    "min_loss_test = min(loss_test)\n",
    "min_loss_test_idx = loss_test.index(min_loss_test)\n",
    "max_auc_test = max(auc_test)\n",
    "max_auc_test_idx = auc_test.index(max_auc_test)\n",
    "max_accuracy_test = max(accuracy_test)\n",
    "max_accuracy_test_idx = accuracy_test.index(max_accuracy_test)\n",
    "print('Minimum test loss {} at step {}, and training loss {}'.format(min_loss_test,\n",
    "                                                                     100 * min_loss_test_idx,\n",
    "                                                                     loss_train[min_loss_test_idx]))\n",
    "print('Maximal test auc {} at step {}, and training auc {}'.format(max_auc_test,\n",
    "                                                                  100 * max_auc_test_idx,\n",
    "                                                                  auc_train[max_auc_test_idx]))\n",
    "print('Maximal test accuracy {} at step {}, and training accuracy {}'.format(max_accuracy_test,\n",
    "                                                                             100 * max_accuracy_test_idx,\n",
    "                                                                             accuracy_train[max_accuracy_test_idx]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYwAAAEKCAYAAAAB0GKPAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f627660f850>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "def plots(loss_train, loss_test, auc_train, auc_test, accuracy_train, accuracy_test):\n",
    "    figure, (ax1, ax2, ax3) = plt.subplots(1, 3)\n",
    "    # Plot loss.\n",
    "    ax1.plot(loss_train, label='train')\n",
    "    ax1.plot(loss_test, label='test')\n",
    "    ax1.legend()\n",
    "    ax1.set_ylabel('loss')\n",
    "    ax1.set_xlabel('train steps (X100)')\n",
    "    # Plot auc.\n",
    "    ax2.plot(auc_train, label='train')\n",
    "    ax2.plot(auc_test, label='test')\n",
    "    ax2.legend()\n",
    "    ax2.set_ylabel('auc')\n",
    "    ax2.set_xlabel('train steps (X100)')\n",
    "    # Plot accuracy.\n",
    "    ax3.plot(accuracy_train, label='train')\n",
    "    ax3.plot(accuracy_test, label='test')\n",
    "    ax3.legend()\n",
    "    ax3.set_ylabel('accuracy')\n",
    "    ax3.set_xlabel('train steps (X100)')\n",
    "\n",
    "    plt.show()\n",
    "\n",
    "plots(loss_train, loss_test, auc_train, auc_test, accuracy_train, accuracy_test)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Fifth, make use of Word2Vec"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Countinuous bag of words or predictive word embedding."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[['movie', 'well', 'balanced', 'comedy', 'drama', 'thoroughly', 'enjoyed'],\n",
      " ['instructions', 'explain', 'microphone', 'jack', 'could', 'used']]\n",
      "[['highly', 'recommend', 'one', 'blue', 'tooth', 'phone'],\n",
      " ['particular', 'model', 'would', 'work', 'motorola', 'q', 'smartphone']]\n"
     ]
    }
   ],
   "source": [
    "from functools import partial\n",
    "import pprint\n",
    "\n",
    "\n",
    "# Map sentences to words.\n",
    "X_train_words = map(partial(clean_sentence, return_sentence=False, remove_stopwords=True), raw_X_train)\n",
    "X_test_words = map(partial(clean_sentence, return_sentence=False, remove_stopwords=True), raw_X_test)\n",
    "\n",
    "pp = pprint.PrettyPrinter(indent=1, width=120)\n",
    "pp.pprint(X_train_words[:2])\n",
    "pp.pprint(X_test_words[:2])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "2017-06-13 21:32:15,890 : INFO : collecting all words and their counts\n",
      "2017-06-13 21:32:15,891 : INFO : PROGRESS: at sentence #0, processed 0 words, keeping 0 word types\n",
      "2017-06-13 21:32:15,898 : INFO : collected 3939 word types from a corpus of 12474 raw words and 2100 sentences\n",
      "2017-06-13 21:32:15,899 : INFO : Loading a fresh vocabulary\n",
      "2017-06-13 21:32:15,913 : INFO : min_count=1 retains 3939 unique words (100% of original 3939, drops 0)\n",
      "2017-06-13 21:32:15,914 : INFO : min_count=1 leaves 12474 word corpus (100% of original 12474, drops 0)\n",
      "2017-06-13 21:32:15,933 : INFO : deleting the raw counts dictionary of 3939 items\n",
      "2017-06-13 21:32:15,935 : INFO : sample=0.001 downsamples 38 most-common words\n",
      "2017-06-13 21:32:15,936 : INFO : downsampling leaves estimated 11579 word corpus (92.8% of prior 12474)\n",
      "2017-06-13 21:32:15,937 : INFO : estimated required memory for 3939 words and 300 dimensions: 11423100 bytes\n",
      "2017-06-13 21:32:15,947 : INFO : resetting layer weights\n",
      "2017-06-13 21:32:16,015 : INFO : training model with 4 workers on 3939 vocabulary and 300 features, using sg=1 hs=0 sample=0.001 negative=20 window=5\n",
      "2017-06-13 21:32:17,032 : INFO : PROGRESS: at 50.71% examples, 115864 words/s, in_qsize 7, out_qsize 0\n",
      "2017-06-13 21:32:17,990 : INFO : worker thread finished; awaiting finish of 3 more threads\n",
      "2017-06-13 21:32:18,005 : INFO : worker thread finished; awaiting finish of 2 more threads\n",
      "2017-06-13 21:32:18,006 : INFO : worker thread finished; awaiting finish of 1 more threads\n",
      "2017-06-13 21:32:18,013 : INFO : worker thread finished; awaiting finish of 0 more threads\n",
      "2017-06-13 21:32:18,014 : INFO : training on 249480 raw words (231599 effective words) took 2.0s, 116048 effective words/s\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "(3939, 300)"
      ]
     },
     "execution_count": 25,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from gensim.models.word2vec import Word2Vec\n",
    "import logging\n",
    "\n",
    "logging.basicConfig(format='%(asctime)s : %(levelname)s : %(message)s', level=logging.INFO)\n",
    "\n",
    "# sg, 0 for countinuous bag of words, 1 for skip-gram which works better for logistic regression here.\n",
    "model = Word2Vec(sentences=X_train_words, size=300, window=5, min_count=1,\n",
    "                 max_vocab_size=None, sample=0.001, workers=4,\n",
    "                 min_alpha=0.0001, sg=1, hs=0, negative=20, cbow_mean=1,\n",
    "                 iter=20, null_word=0, trim_rule=None, sorted_vocab=1, batch_words=1000)\n",
    "\n",
    "# syn0 is a numpy array that stores a feature vector for each word.\n",
    "model.wv.syn0.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((300,),\n",
       " array([ 0.12181101, -0.04280083, -0.21245392, -0.11450065,  0.05007749,\n",
       "        -0.12880118, -0.06752747, -0.19197069,  0.08410849, -0.01657801], dtype=float32))"
      ]
     },
     "execution_count": 26,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model['phone'].shape, model['phone'][:10]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.99292146561030181"
      ]
     },
     "execution_count": 27,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# This doesn't make much sense.\n",
    "model.similarity('phone', 'good')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Transform sentence words to dense vector."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Represent a sentence as mean of the vectors of its words."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(300,)"
      ]
     },
     "execution_count": 28,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# First method - averaging words vectors.\n",
    "def ws2meanvec(words, embedding_model):\n",
    "    \"\"\"\n",
    "    Compute mean vector of a list of words.\n",
    "    \"\"\"\n",
    "    vectors = [embedding_model[w] for w in words if w in embedding_model]\n",
    "    if len(vectors) > 0:\n",
    "        return np.mean(vectors, axis=0)\n",
    "    else:\n",
    "        # To avoid empty words causing error. For example, raw sentence is 10/10 at imdb.\n",
    "        return np.zeros(embedding_model.vector_size, dtype=np.float32)\n",
    "\n",
    "ws2meanvec(['phone', 'good', 'work'], model).shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(array([ 0.15610416, -0.05606126, -0.2159145 , -0.12296281,  0.05785468,\n",
       "        -0.10106085, -0.06802104, -0.18052582,  0.08882451, -0.00321173], dtype=float32),\n",
       " array([ 0.161189  , -0.04263185, -0.22173673, -0.12244137,  0.05912674,\n",
       "        -0.11804243, -0.07963934, -0.18222256,  0.09119293,  0.01295353], dtype=float32))"
      ]
     },
     "execution_count": 29,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "X_train_cbow = map(partial(ws2meanvec, embedding_model=model), X_train_words)\n",
    "X_test_cbow = map(partial(ws2meanvec, embedding_model=model), X_test_words)\n",
    "X_train_cbow[0][:10], X_test_cbow[0][:10]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# Some test sentences don't have a word in the vocabulary.\n",
    "ensure_no_error = [len(e) for e in X_train_cbow]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "ground truth, predictions on default threshold, prediction probabilities\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "([1, 0, 1, 1, 0, 0, 1, 0, 1, 1, 1, 0, 1, 1, 0, 0, 1, 1, 0, 1],\n",
       " [1, 1, 1, 1, 1, 1, 1, 0, 1, 0, 1, 0, 0, 0, 0, 0, 1, 1, 0, 1],\n",
       " array([ 0.26391771,  0.28913893,  0.36542769,  1.85804354,  0.43771857,\n",
       "         0.59934859,  0.0592134 , -0.11441425,  0.06177486, -0.03211928,\n",
       "         0.13187559, -0.15628349, -0.25878999, -0.03729056, -0.37299931,\n",
       "        -0.93954224,  0.45756042,  0.07359714, -0.29783902,  0.07049611]))"
      ]
     },
     "execution_count": 31,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Fit a logistic regression model on train set and make predictions on test data.\n",
    "from sklearn.linear_model import LogisticRegression\n",
    "\n",
    "log_reg = LogisticRegression(C=10)\n",
    "\n",
    "# Train.\n",
    "log_reg.fit(X_train_cbow, y_train)\n",
    "# Test.\n",
    "pred_test = log_reg.predict(X_test_cbow)\n",
    "pred_prob_test = log_reg.decision_function(X_test_cbow)\n",
    "print('ground truth, predictions on default threshold, prediction probabilities')\n",
    "y_test[:20], pred_test[:20].tolist(), pred_prob_test[:20]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "accuracy on default threshold: 0.621111111111\n",
      "area under curve: 0.687065121309\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYoAAAEKCAYAAAAMzhLIAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f62766e4410>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Compute classification performance.\n",
    "from sklearn.metrics import accuracy_score, roc_auc_score, roc_curve\n",
    "\n",
    "accuracy = accuracy_score(y_test, pred_test)\n",
    "\n",
    "auc = roc_auc_score(y_test, pred_prob_test)\n",
    "\n",
    "fpr, tpr, thresholds = roc_curve(y_test, pred_prob_test)\n",
    "\n",
    "print('accuracy on default threshold: {}\\narea under curve: {}'.format(accuracy, auc))\n",
    "\n",
    "plot_auc(fpr, tpr)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Let's try transfer learning"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "2017-06-13 21:32:18,503 : INFO : loading projection weights from ./GoogleNews-vectors-negative300.bin.gz\n",
      "2017-06-13 21:35:31,265 : INFO : loaded (3000000, 300) matrix from ./GoogleNews-vectors-negative300.bin.gz\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "(3000000, 300)"
      ]
     },
     "execution_count": 33,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from gensim.models import KeyedVectors\n",
    "\n",
    "# Load google pre-trained model.\n",
    "# From http://mccormickml.com/2016/04/12/googles-pretrained-word2vec-model-in-python/\n",
    "google_model = KeyedVectors.load_word2vec_format('./GoogleNews-vectors-negative300.bin.gz', binary=True)\n",
    "google_model.syn0.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.074366691017249367"
      ]
     },
     "execution_count": 34,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "google_model.similarity('phone', 'good')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "2017-06-13 21:35:31,280 : INFO : precomputing L2-norms of word weight vectors\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "[(u'kings', 0.7138045430183411),\n",
       " (u'queen', 0.6510956287384033),\n",
       " (u'monarch', 0.6413193941116333),\n",
       " (u'crown_prince', 0.6204220056533813),\n",
       " (u'prince', 0.6159994006156921),\n",
       " (u'sultan', 0.5864824056625366),\n",
       " (u'ruler', 0.5797566771507263),\n",
       " (u'princes', 0.5646552443504333),\n",
       " (u'Prince_Paras', 0.543294370174408),\n",
       " (u'throne', 0.5422105193138123)]"
      ]
     },
     "execution_count": 35,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "google_model.most_similar('king')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[(u'universities', 0.7003918886184692),\n",
       " (u'faculty', 0.6780906915664673),\n",
       " (u'unversity', 0.6758289933204651),\n",
       " (u'undergraduate', 0.6587094068527222),\n",
       " (u'univeristy', 0.6585439443588257),\n",
       " (u'campus', 0.6434986591339111),\n",
       " (u'college', 0.638526976108551),\n",
       " (u'academic', 0.6317198276519775),\n",
       " (u'professors', 0.6298649311065674),\n",
       " (u'undergraduates', 0.6149813532829285)]"
      ]
     },
     "execution_count": 36,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "google_model.most_similar('university')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Represent a sentence as mean of the vectors of its words."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(array([-0.03456334,  0.0370745 , -0.11201695,  0.16207449, -0.00889369,\n",
       "        -0.02577427,  0.01097979,  0.00630406,  0.07080078,  0.00959996], dtype=float32),\n",
       " array([-0.01264445, -0.08089193,  0.04174805,  0.01407878, -0.16902669,\n",
       "         0.03644816,  0.06310526, -0.10062663,  0.11686198,  0.01391602], dtype=float32))"
      ]
     },
     "execution_count": 37,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Remove stopwords or not, please check previous cells.\n",
    "X_train_tl = np.array(map(partial(ws2meanvec, embedding_model=google_model), X_train_words))\n",
    "X_test_tl = np.array(map(partial(ws2meanvec, embedding_model=google_model), X_test_words))\n",
    "X_train_tl[0][:10], X_test_tl[0][:10]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "ground truth, predictions on default threshold, prediction probabilities\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "([1, 0, 1, 1, 0, 0, 1, 0, 1, 1, 1, 0, 1, 1, 0, 0, 1, 1, 0, 1],\n",
       " [1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 0, 0, 1, 0, 0, 1, 1, 0, 0],\n",
       " array([ 1.10976065,  0.57030594,  2.54076074,  3.00765064, -0.3820346 ,\n",
       "        -0.26290903, -0.25651127, -3.05186962,  1.99199124,  2.74571328,\n",
       "         0.39430714, -0.89937996, -0.55053778,  2.9119668 , -2.02392607,\n",
       "        -2.53171542,  1.31356882,  2.44599221, -3.08721885, -0.87593437]))"
      ]
     },
     "execution_count": 38,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Fit a logistic regression model on train set and make predictions on test data.\n",
    "from sklearn.linear_model import LogisticRegression\n",
    "\n",
    "log_reg = LogisticRegression(solver='lbfgs')\n",
    "\n",
    "# Train.\n",
    "log_reg.fit(X_train_tl, y_train)\n",
    "# Test.\n",
    "pred_test = log_reg.predict(X_test_tl)\n",
    "pred_prob_test = log_reg.decision_function(X_test_tl)\n",
    "print('ground truth, predictions on default threshold, prediction probabilities')\n",
    "y_test[:20], pred_test[:20].tolist(), pred_prob_test[:20]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "accuracy on default threshold: 0.82\n",
      "area under curve: 0.903888414264\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYoAAAEKCAYAAAAMzhLIAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f627a9b4890>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Compute classification performance.\n",
    "from sklearn.metrics import accuracy_score, roc_auc_score, roc_curve\n",
    "\n",
    "accuracy = accuracy_score(y_test, pred_test)\n",
    "\n",
    "auc = roc_auc_score(y_test, pred_prob_test)\n",
    "\n",
    "fpr, tpr, thresholds = roc_curve(y_test, pred_prob_test)\n",
    "\n",
    "print('accuracy on default threshold: {}\\narea under curve: {}'.format(accuracy, auc))\n",
    "\n",
    "plot_auc(fpr, tpr)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Support vector machine."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "ground truth, predictions on default threshold, prediction probabilities\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "([1, 0, 1, 1, 0, 0, 1, 0, 1, 1, 1, 0, 1, 1, 0, 0, 1, 1, 0, 1],\n",
       " [1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 0, 0, 1, 0, 0, 1, 1, 0, 0],\n",
       " array([ 0.788459  ,  0.26117121,  1.69120277,  2.17865987, -0.05390308,\n",
       "        -0.25132504, -0.11454365, -1.73419407,  1.08987599,  1.55242931,\n",
       "         0.30301341, -0.56123168, -0.41550107,  1.59683295, -1.46200424,\n",
       "        -1.35474506,  0.58426006,  1.0485971 , -1.88305036, -0.64584004]))"
      ]
     },
     "execution_count": 40,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Fit an svc model on train set and make predictions on test data.\n",
    "from sklearn.svm import SVC\n",
    "\n",
    "svc = SVC(C=1.0, gamma=0.1, probability=True, decision_function_shape='ovr')\n",
    "\n",
    "# Train.\n",
    "svc.fit(X_train_tl, y_train)\n",
    "pred_train = svc.predict(X_train_tl)\n",
    "pred_prob_train = svc.decision_function(X_train_tl)\n",
    "# Test.\n",
    "pred_test = svc.predict(X_test_tl)\n",
    "pred_prob_test = svc.decision_function(X_test_tl)\n",
    "print('ground truth, predictions on default threshold, prediction probabilities')\n",
    "y_test[:20], pred_test[:20].tolist(), pred_prob_test[:20]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train accuracy on default threshold: 0.857619047619\n",
      "area under curve: 0.928112406451\n",
      "Test accuracy on default threshold: 0.827777777778\n",
      "area under curve: 0.906490896251\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYoAAAEKCAYAAAAMzhLIAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz\nAAALEgAACxIB0t1+/AAAIABJREFUeJzt3Xd8FHX+x/HXJwk9EJDegkgVBGlSRYqggBTbYff09DwL\n9oKevf2wH2JHz/M4C3rqSZGOFGmCSK8GRDpIDyEJSfbz+2NWWClhUmZnd/N5Ph55sDs7u/vOkOwn\nM9+Zz1dUFWOMMeZk4vwOYIwxJrJZoTDGGJMrKxTGGGNyZYXCGGNMrqxQGGOMyZUVCmOMMbnyrFCI\nyIcislNElp/kcRGRYSKSIiJLRaSVV1mMMcbkn5d7FB8BvXJ5vDfQIPh1C/COh1mMMcbkk2eFQlVn\nAntyWWUAMEId84DyIlLdqzzGGGPyJ8HH964JbAq5vzm4bNuxK4rILTh7HZQpU6Z148aNwxLQGGPy\n6tDhbNIPBwrltfanZ5GRlYNI3p6XHXA6bsQRoCp7qSQHWLgtZ5eqVs5PDj8LhWuqOhwYDtCmTRv9\n8ccffU5kjIlFgYDyxY+bOJCR9YflG/cc4uN5GylVLJ6E+Nw/tQ9lZBdannLBr6vbJef5uc0zf+Ly\nrS+TcGAXaWcPIvHS13/Nbw4/C8UWoHbI/VrBZcaYIkpVGblgE7tSMwv9tV+f+jPZAaV4wsmPuB/O\nzn1PoEO9itSpWDrXdVSha6PKNK9VPl85j5VYIiHXzMdJ3wsTH4PVH0PFBnDjBMrU6QC8nu8MfhaK\n0cAgERkJtAP2q+pxh52MMf7IzM7hVD1DJ63cwdrtqa5fc+2OVCat3FHAZPlXtmQC17Srk+s6xeKF\n6zrUoUzxP348JsQLJRLivYxXcKvGwLf3Q9ouOPc+6DIYipUs8Mt6VihE5DOgK1BJRDYDTwLFAFT1\nXWAc0AdIAQ4BN3qVxRjj3uHsAO/NWMerk9e6fk58nLuD6DnBY+dXtKlNtaQTf4AVixeuaptMUqli\nrt/frYT4GL10LHUHjH8QVo6Cas3g6i+gRotCe3nPCoWqXnWKxxW4w6v3NyYaHMzM5tDhwjum7dac\nlN0s/HXvcYOkWTnKZ/M3Hrn/UK9Gub6OIPRpVo06Fct4EdOciios+QwmPAJZ6XD+E9DxLogv3CIb\nFYPZxsSiTXsO0fmlab5mqFD6jx8oAYVSxeLpUK8iN3Y6nc4N8nWSjAmHfRthzD2wbirUbgf934TK\nDT15KysUxngg/XAOW/alAzB68RY27jl03DrfLN4KwHkNK3NBk6phzQfQ5vQKNK5WLuzvawooEIAF\nH8CUp5z7vV+Gc26GOO8Oq1mhMMal3Qcz2X4g44SPLd+ynymrdpIQPFY/fvn249Y59myZmuVL0bxW\nEkOvbBH5g6QmMuz6GUYNgk3zoN750G8olM/7qbN5ZYXCmKBfd6exO+3wH5aNmLOBA8Hz4r9bvfOU\nr9GwaiIA9askUqtCKS5tVQuA9nVPo0q5gp99YoqonCyYMwymvwjFSsHF78DZV5HnK/HyyQqFKdIC\nAeWN71KYtHI7K7YeOOl6zWomcWb1crRMLk+Xhic+bp98WmnOrG6Hckwh27bE2YvYvhSaDHAONZUN\n76FKKxSmSNh5IIP7/7uEQ4dzCP0bbOHGvUeuFTirZjkublGT+lUSjzwuIrSuU4HEEvarYsIsKwNm\nvAizX4fSFWHgf6BJf1+i2E+/iSmqytx1u/lw9ga2H0hHgmVh2Zb9gHPBVfNaSUfW71ivIocO5/Da\nwBbUrWSneJoI8etcGD0IdqdAi2vhwuegVAXf4lihMFHr+59/Y88xYwr3fbHkyEVdAN0bVznyb9mS\nCTx/STPbOzCRKzMVpjwNC953Bqmv/Rrqn+93KisUJrpMW72TD2at58cNe8nMpS/Ppze34+za5Slj\nRcFEi5QpznUR+zdDu1uh++NQIvHUzwsD+y0yYREIKGOWbj1yBtH3a39j/oY9FI+Pc33ixt5DWX9o\n2ta3eXWubV+HymVLHFkWJ0Kd00oT57KlhDG+O7QHJv7ducK6UkP4y0RIbud3qj+wQmE8MydlFyu3\nHeCz+RtZ91vaCdcZ2KYWcXk4xS8noFzVLpkm1ctRsphde2Ci3IpvYNwDTsfXzg/AeQ8WShO/wmaF\nwuRJ+uEcPpu/kfSsnOMe+3LhZjbuOUTxYOO1Y9e5oePpXN+hDmVLOm0jypZMsA97UzSlbne6vK4e\nC9XPdsYiqjf3O9VJWaEwrs1bv5sh41axZPP+XNe77lynjbOqcmHTajSsVpYyxRNcdxg1JmapwuJP\nnENNWRnQ4ynocCfER/ZHcWSnM747dDibtMwcRszdwBvfpQBQpng8E+45jyrlShy3frG4OBsfMOZE\n9m6AMXfD+umQ3BH6vwGV6vudyhUrFAaA/YeyyA4E2LQ3nW8WORMNTli+/bjeRrd1rcfgXjZnuTGu\nBXJg/vsw9WmQOLjoVWj9F0+b+BU2KxRFXFZOgH5vzGL1MbOUlS5+dOzg0T5nklgygT+1rhW7E78Y\n44Xf1jjtNzbPh/o9oe8/oHztUz8vwlihKKJUlS370lmzPfVIkXiibxMS4oVaFUrRvXH4214bEzNy\nsmD2UJjxEhQvA5cMh+YDw9bEr7BZoSiixi/fzu2f/HTk/r9uOIduwauYjTEFsHWRsxexYzk0vcRp\n4pcY3RNAWaEogrJyAtz3xWIAnhnQlCplS9CpfiWfUxkT5bLSYfoQmPMGlKkCV3wCZ/b1O1WhsEJR\nxKRlZnPl8HlkZAWOTGJfzMYdjCmYDbNh9J2wZx20uh56PgulyvudqtBYoSgCDmZms2DDHl6btPZI\nF1WASfd2sSJhTEFkHHCmJP3xn1C+Dlw/Cs7o6nOowmeFIsYEAsqdIxexec+hIwNnSzbt+8M6vZpW\n45WBZ1sXVWMKYu0kGHsvHNgC7e+A7o86A9cxyD4pYswLE1bz7dJtAEdmYuvSsDJJpYpxc+e6nFUj\nyS6IM6Yg0nbDxEdg6edQuTHcNBlqn+N3Kk9ZoYhi2/dncPOIBRzKPNpTaf0up/nelPu6/GGmNmNM\nAanCiq9h3EOQsQ+6DIbO90PC8R0KYo0Viii2ftdBlm85QMd6FamY6PywNq2ZxCUta1iRMKYwHdgG\n394Ha8ZBjZbQfxRUO8vvVGFjhSIG3HV+A9qfUdHvGMbEHlX4aQRMehxyMuGC56DdbRHfxK+wFa3v\nNoYs37KfMUu2+h3DmNi15xcYcxf8MhPqnAv9h0HFen6n8oUViii040AGfd+YBUCJhDiqlou8iU6M\niVqBHPjhXZj6LMQlQN+h0OrPUdXEr7BZoYhCny/YBDhTgb42sAXFE4ruD7AxhWrnKqf9xpYfocGF\nThO/pJp+p/KdFYoo89rktQyb+jMAQ69oYd1cjSkM2Ydh1j9g5stQshxc9k8467KobeJX2KxQRJE5\nKbuOFImHejWyImFMYdiy0NmL2LkSmv0Jer0AZaz3WSgrFBEuOyfA0Ck/s+NABv9duBmAN69uSd/m\nNXxOZkyUO3wIpj0P896GxGpw1Uho1NvvVBHJCkUEysoJsCF44dyj3yxn/i97AKhargQXt6hpRcKY\ngvrle6eJ395foPWN0PNpKJnkd6qIZYUiQmRm5/Dc2FXsTM1g4oodxz2+8LEeRy6qM8bkU8Z+mPwE\nLPwIKtSFP4+Buuf5nSriWaHwUUZWDiu2HgCU575dxaKNTvO+RlXLUqp4PDd3rgtAk+rlrEgYU1Br\nxsPY++Dgduh4J3T9OxQv7XeqqOBpoRCRXsDrQDzwgaq+cMzjScDHQHIwyyuq+i8vM0WKn3ek8sCX\nS4/r7Lr4iZ6UL13cp1TGxKC0XTB+MCz/Eqo0gSs+hlqt/U4VVTwrFCISD7wF9AQ2AwtEZLSqrgxZ\n7Q5gpar2E5HKwBoR+URVD3uVKxKEnuIKMOIvbQGoW6mMFQljCosqLPsSxj8EmanOHsS590KC/Y7l\nlZd7FG2BFFVdDyAiI4EBQGihUKCsiAiQCOwBsj3M5Jt3pq9jzJKtiBA83ATPDmjK+WdWpUb5Uj6n\nMybG7N/iNPFbOwFqtoEBb0KVM/1OFbW8LBQ1gU0h9zcD7Y5Z501gNLAVKAtcoaqBY19IRG4BbgFI\nTk72JKwXNu4+xLjl2/hw1i/sTM0EoMeZVaieVJLLW9em11nVfE5oTIwJBOCnj2DSExDIhgv/D9rd\nCnHxfieLan4PZl8ILAa6A/WAySLyvaoeCF1JVYcDwwHatGmjYU+ZR98u3cbDXy8lNePozlHjamW5\np0dDKw7GeGX3OhhzN2z43jmTqd8wOK2u36ligpeFYgtQO+R+reCyUDcCL6iqAiki8gvQGJjvYS5P\nDZv6M69NXgtA0xrluLhFTfq3qGGN+4zxSk62c9HctOchvrhTIFpdb+03CpGXhWIB0EBE6uIUiCuB\nq49ZZyNwPvC9iFQFGgHrPczkqWlrdh4pEp/f0p52NkeEMd7avhxGD4Kti6BRH7joVShnF6QWNs8K\nhapmi8ggYCLO6bEfquoKEbk1+Pi7wLPARyKyDBBgsKru8iqT11ZvSwXg3WtbWZEwxkvZmfD9q85X\nyfJw+b+g6SW2F+ERT8coVHUcMO6YZe+G3N4KXOBlBq/9lppJ79dnsuvg0TN6uzaq4mMiY2LcpgXO\nXsRvq6H5FU4Tv9Kn+Z0qpvk9mB31xi3bdqRI3NW9PrVPK03JYnaGhTGF7nAafBds4leuBlz9X2gY\n1X9nRg0rFAV0ID0LgCVPXEBS6WI+pzEmRq2fDqPvgn2/QpuboMdTzrwRJiysUBTAr7vTeDU4eB0f\nb8dGjSl06ftg8uPw0wg4rR7cMA5O7+R3qiLHCkU+zU7ZxTUf/ADA1e2SSSxhm9KYQrX6W6eJX9pO\n6HQ3dH0EilkXAz/Yp1s+zFu/+0iRuKhZdZ7s18TnRMbEkIM7nf5MK/4HVc+Cqz6Dmq38TlWkWaFw\nKRBQVm0/wNZ9Gfx1xI8AXNMumecvaeZzMmNihCos/QImDHYGrrs/Bp3ugXgb+/ObFQqX3pmxjpcn\nrjly/29dzuCR3tZkzJhCsW8TjL0XUiZDrbZOE7/KjfxOZYKsULiwN+3wkSLxyp/Opkb5krQ93c7b\nNqbAAgFY+CFMfhI0AL1ehLZ/tSZ+EcYKhQvpWTkA3NjpdC5vXcvnNMbEiF0pzrzVG+fAGV2h3+tQ\n4XSfQ5kTsUKRB42rlfU7gjHRLycb5r4B04ZAsZIw4C1ocY2134hgVihy8a/Zv/DWtHWUSIjzO4ox\nsWH7Mhh1B2xbAo37Ok38ylrr/UhnheIk5qzbxdNjnMn4OtarSMd6FTm3QWWfUxkTpbIyYObLMHso\nlDoNBo6AJgP8TmVcskJxjIkrtrNlbzrPjHWKxAuXNuPKttEzq54xEWfjD04Tv11r4eyr4cLnrYlf\nlLFCESL9cA5/+8/CI/cvbVnTioQx+ZV5EL57Fn54D5JqwbVfQf0efqcy+WCFIsRDXy0F4MELG3Ft\nuzqUK2Wbx5h8SZkKY+6B/Ruh7S1w/hNQwk4GiVb2SRhiw640AP7UupZ1gjUmP9L3wsRHYfEnULEB\n3DgB6nTwO5UpICsUQSk7D7Jsy366NapMFZvf2pi8Wzkaxj0Aabvg3Pugy2Dn9FcT9axQBP134SYA\nmtcq73MSY6JM6g6nQKwaDdWawTX/hepn+53KFKJTFgoREeAa4AxVfUZEkoFqqjrf83Rhsj89i/dm\nrCc+Tri3Z0O/4xgTHVRhyWcw4RHISnfGITreZU38YpCbPYq3gQDQHXgGSAW+As7xMFdYLdq4F4A6\np5X2OYkxUWLfRmewet1UqN0e+r8Ble2PrFjlplC0U9VWIrIIQFX3ikhxj3P54pWBtrtsTK4CAVjw\nAUx5yrnf+2U452aIs+4FscxNocgSkXhAAUSkMs4ehjGmKPltrdPEb9M8qHc+9BsK5e06o6LATaEY\nBvwPqCIizwOXA497msoYEzlysmD26zDjRShWGi5+F86+0pr4FSGnLBSq+omILATOBwS4WFVXeZ4s\nTFSVx0ct9zuGMZFp2xKnid/2ZU5vpj6vQGIVv1OZMHNz1tN/VPU6YPUJlkW9rBxl0550ABpWtStH\njQGcJn4zXoDZw6BMJRj4H2jS3+9UxiduDj01Db0THK9o7U0c/zx4YSMSS9hlJcbw61ynid/uFGhx\nLVz4HJSq4Hcq46OTfjKKyCPA34FSInIA57ATwGFgeBiyGWPCKTMVpjwNC953Bqmv+x/U6+53KhMB\nTlooVHUIMEREhqjqI2HMZIwJt5+nwNh7YP9maHcbdH8MSiT6ncpECDeD2Y+ISAWgAVAyZPlML4MZ\nY8Lg0B6Y+HfnCutKjeCmSVC7rd+pTIRxM5h9M3A3UAtYDLQH5uJcqW2MiUaqsHKU06MpfS+c96Dz\nlVDC72QmArkZvb0bp13HPFXtJiKNgf/zNlb4qHMdoTFFR+p2+PZ+WD0WqrdwxiKqNfM7lYlgbgpF\nhqpmiAgiUkJVV4tII8+ThcGa7alcONQ5ghYfZxcPmRinCos+hkmPQnYm9HgaOgyCeDvbz+TOzU/I\nZhEpD3wDTBaRvcCv3sby3oqt+7lo2CwAqieV5NJWNX1OZIyH9m6AMXfD+umQ3NFp4lepvt+pTJRw\nM5h9SfDmUyIyDUgCJniaykOqysgFm3jk62UADOpWnwcujIkdJGOOF8iB+cNh6jMg8XDRq9D6L9bE\nz+RJroUieHHdClVtDKCqM/Ly4iLSC3gdiAc+UNUXTrBOV2AoUAzYpapd8vIeefHejHX8Y8paMrKc\nnoYDWtTgnh4NvHo7Y/y1c7XTxG/zfKjf02nil1TL71QmCuVaKFQ1R0TWiEiyqm7MywsHi8xbQE9g\nM7BAREar6sqQdcrjzHfRS1U3iohnTWQmrtjOkPFOF5KKZYrz+d86UL+KnSduYlBOFswaCjNfguKJ\ncOn70OxP1sTP5JubMYoKwAoRmQ+k/b5QVU/V+KUtkKKq6wFEZCQwAFgZss7VwNe/FyFV3ZmH7K6t\n2Z7K3/6zEIAPb2hD98ZVvXgbY/y3dRGMGgQ7lkPTS6H3S5BY2e9UJsq5KRT5bSleE9gUcn8z0O6Y\ndRoCxURkOlAWeF1VRxz7QiJyC3ALQHJy3vvff//zbwDc3rWeFQkTm7LSYfoQmPMGlKkCV34KjS/y\nO5WJEW4Gs/M0LpGP92+N08K8FDBXROap6tpjMgwn2F+qTZs2+b7w4bau9QoQ1ZgItWGWMxaxZz20\nuh56PgulyvudysQQL0+g3gLUDrlfK7gs1GZgt6qmAWkiMhM4G1iLMSZ3GQdgypPw44dQ4XS4fhSc\n0dXnUCYWeXmO3AKggYjUDc6xfSUw+ph1RgHnikiCiJTGOTQVM5MiGeOZtZPg7faw8CPnornb5liR\nMJ5xtUchIqWAZFVd4/aFVTVbRAYBE3FOj/1QVVeIyK3Bx99V1VUiMgFYijMP9weqWqjTze0/lMVz\n31rtMTEibTdMeBiWfQGVG8PAEVCrjd+pTIxz0xSwH/AKUByoKyItgGdcnPWEqo4Dxh2z7N1j7r8M\nvJyX0G6kZmQxbtk2ft19CICmNcrZxEQmeqnCiq9h3EOQsQ+6PAyd77MmfiYs3HxyPoVzqut0AFVd\nLCJ1PcxUYAcysmj+1KQj90Xg2YvPQuw8chONDmx1mvitGQc1WsKA0VC16amfZ0whcVMoslR1/zEf\nshHdcnXngUwA6lYqw+d/a0+JhHiSShXzOZUxeaQKP/0bJj0OOYfhguecSYWsiZ8JMzc/cStE5Gog\nXkQaAHcBc7yNVTBTVu0A4N6eDalStuQp1jYmAu1ZD6Pvgg3fw+mdod/rUNFO7zb+cHPW051AUyAT\n+BTYD9zjZaiCenfGOgBqVyjlcxJj8iiQA3PehLc7wrYl0HcoXD/aioTxlZs9isaq+ijwqNdhCkvx\n+DiualublskV/I5ijHs7VsLoQbBlITTsBRe9BknW/t74z02heFVEqgFfAp8X9umrhS0QUHYdzPQ7\nhjHuZR+GWa/BzFegZDm47J9w1mXWxM9EDDctPLoFC8VA4D0RKYdTMJ7zPF0+3PnZIgLq7FUYE/E2\nL3T2InaudDq89noBylTyO5Uxf+Dq01RVt6vqMOBWYDHwhKep8mnyyh18u2wbALd3s9m7TAQ7fAgm\nPgr/7AHp++Cqz+GyD6xImIjk5oK7M4ErgMuA3cDnwP0e58qXvWmHAaeVeNVydraTiVC/zHSa+O3d\nAK1vhJ5PQ8kkv1MZc1Juxig+xCkOF6rqVo/zFIpG1cr5HcGY42Xsh8lPOP2ZKtSFP4+Fup39TmXM\nKbkZo+gQjiCFYeW2A35HMObE1oyHsffCwR3Q8U7o+ncoXtrvVMa4ctJCISJfqOpAEVnGH6/EFkBV\ntbnn6fJg18FMPpqzAYDSxeL9DWPM79J2wfjBsPxLqNIUrvwEarb2O5UxeZLbHsXdwX/7hiNIQaUf\nzgGcWewqlCnucxpT5KnCsi9h/EOQmersQZx7LyTYz6aJPictFKq6LXjzdlUdHPqYiLwIDD7+Wf5Z\nFTzsVLdSGZ+TmCJv/2YYex/8PBFqtoEBb0KVM/1OZUy+uTk9tucJlvUu7CAF9fSYlQBUSrS2y8Yn\ngYAz29xb7Z0eTRcOgZsmWZEwUS+3MYrbgNuBM0RkachDZYHZXgfLjw5nVKRb4yp+xzBF0e51ThO/\nX2dB3S5OE7/TIrobvzGu5TZG8SkwHhgCPByyPFVV93iaKh/i4qB6kl07YcIsJxvmvQ3Tnof4EtD/\nDWh5nbXfMDElt0KhqrpBRO449gEROS0Si4UxYbV9udN+Y+siaHQRXPQqlKvudypjCt2p9ij6Agtx\nTo8N/RNJgTM8zGVM5MrOdBr4zXoNSpaHy/8FTS+xvQgTs3I766lv8F870GrM7zYtcPYiflsNza+E\nXkOg9Gl+pzLGU256PXUCFqtqmohcC7QChqrqRs/TGRMpDqfBd8/BvHegXE245ktocKITAo2JPW5O\nj30HOCQiZ+M0A1wH/MfTVMZEkvXT4e0OzqD1OTfB7XOtSJgixU1TwGxVVREZALypqv8UkZu8DmaM\n79L3waTHYNF/4LR6cMM4OL2T36mMCTs3hSJVRB4BrgM6i0gcUMzbWMb4bNVY+PZ+SPsNOt0DXR+G\nYjYHuyma3BSKK4Crgb+o6nYRSQZe9jaWMT45uBPGPQgrv4GqzeDqkVCjpd+pjPHVKccoVHU78AmQ\nJCJ9gQxVHeF5sjyYumoHm/ak+x3DRDNVWDIS3moLa8ZB98fglmlWJIzB3VlPA3H2IKbjXEvxhog8\nqKpfepzNtUkrdgBwQdOqPicxUWnfJmeuiJTJUKut08SvciO/UxkTMdwcenoUOEdVdwKISGVgChAx\nhQKgWrmS9DrLroo1eRAIwI//hClPOXsUvV+Cc26GOJvPxJhQbgpF3O9FImg37k6rDYvM7Bymrt6B\n2FWxJi92pTjzVm+cA2d0c5r4VajjdypjIpKbQjFBRCYCnwXvXwGM8y5S3ny3aie7Dh62hoDGnZxs\nmPsGTBsCxUrCgLehxdXWfsOYXLiZM/tBEbkUODe4aLiq/s/bWO5lZgcAeP/6Nj4nMRFv21Kn/ca2\nJdC4r9PEr2w1v1MZE/Hc7FEAzAFygACwwLs4+VemhNtvxRQ5WRkw8yWYNRRKV4SBI6DJAL9TGRM1\nTjnWICI3A/OBS4DLgXki8hevgxlTKDb+AO91hu9fheZXwB0/WJEwJo/c/Bn+INBSVXcDiEhFnD2M\nD70MZkyBZB6Eqc/A/OGQVAuu/Qrq9/A7lTFRyU2h2A2khtxPDS4zJjKlTIUx98D+TdD2r3D+E1Ci\nrN+pjIlabk5zTQF+EJGnRORJYB6wVkTuE5H7cnuiiPQSkTUikiIiD+ey3jkiki0il+ctvjEhDu2B\nb26Hjy+FhBJw43jo87IVCWMKyM0exbrg1+9GBf/N9bdPROKBt4CewGZggYiMVtWVJ1jvRWCS29DG\nHGflKPj2ATi0G869D7oMdk5/NcYUmJvTY5/O52u3BVJUdT2AiIwEBgArj1nvTuAr4Jz8vMnutMP5\njGdiQuoOGPcArBoN1ZrDtV9C9bP9TmVMTPHynNKawKaQ+5uBdqEriEhNnLOpupFLoRCRW4BbAJKT\nk48sP5iZzbNjnbpTIiFiLhY34aAKiz+FiX+HrHQ4/0noeCfEWwd8Ywqb3xcfDAUGq2ogtxYcqjoc\nGA7Qpk0b/X35ocxsAHqfVY0a5W2ugCJj768w9h5Y9x0kd4D+b0ClBn6nMiZmeVkotgC1Q+7XCi4L\n1QYYGSwSlYA+IpKtqt/k5Y3ObVCpIDlNtAgEYMH7MOVpp+VGn1egzU0QZ3uTxnjJTZvxhjjzZldV\n1bNEpDnQX1WfO8VTFwANRKQuToG4EmcCpCNUtW7I+3wEjM1Lkdiyz+agKDJ+W+s08ds0D+qdD/2G\nQvnkUz/PGFNgbv4Uex94BMgCUNWlOB/6uVLVbGAQMBFYBXyhqitE5FYRuTX/kY968MulACRa+47Y\nlZMFM1+BdzvBb6vh4nedi+esSBgTNm4+YUur6vxjxhCy3by4qo7jmE6zqvruSda9wc1rhgoElLqV\nytCveY28PtVEg62LnSZ+25dBk4udayISq/idypgix02h2CUi9QAFCF4Ut83TVC6JwJnVyxEXZy2i\nY0pWOsx4EWYPgzKV4IqP4cx+fqcypshyUyjuwDnjqLGIbAF+Aa71NJUpun6d6+xF7E6BltfCBc9B\nqQp+pzKmSHNzwd16oIeIlMGZ7S71VM8Jhy370ln3WxqNq5fzO4opDJmpztlMC953xh+u+wbqdfM7\nlTEGd2c9PXHMfQBU9RmPMrnywffrAahp109Ev58nO038DmyBdrdB98egRKLfqYwxQW4OPaWF3C4J\n9MU5i8k3aZnZ/HvOBuLjhEd6N/YziimIQ3tgwiOwdCRUagQ3TYLabf1OZYw5hptDT6+G3heRV3BO\nefXNrJRdBBRqJpUktyu6TYRShZXfwLgHIX0vnPcQnPeA0/HVGBNx8nMBQmmcq6x9Ewg4XTw++LPN\nkx11Urc9r0nrAAARIElEQVTDt/fD6rFQvQVc9z+o1szvVMaYXLgZo1hG8NRYIB6oDPg6PvE725mI\nIqqw6GOY+CjkZELPZ6D9HRBvF0saE+nc/Jb2DbmdDewIXnXtm/3pWX6+vcmrvRtgzN2wfjrU6QT9\nhkGl+n6nMsa4lGuhCE4qNFFVI2bEWBUe/noZACUS4n1OY3IVyHHmrJ76DEg8XPQatL7RmvgZE2Vy\nLRSqmhOcyjRZVTeGK1RuAuocBevcoBKnVyztcxpzUjtXOxfObV4ADS6Avv+AJF+Htowx+eTm0FMF\nYIWIzCfkVFlV7e9ZKhe6NapiZzxFouzDMHsozHwZiifCpe9Dsz/ZgJIxUcxNoXjc8xQmNmz5yWkF\nvmM5nHUZ9HoREiv7ncoYU0BuCkUfVR0cukBEXgRmeBPJRJ2sdJj2fzD3TUisCld+Bo37+J3KGFNI\n3Iwq9jzBst6FHcREqQ2z4J2OMGcYtLwObp9nRcKYGHPSPQoRuQ24HThDRJaGPFQWmO11MBPhMg7A\nlCfhxw+hwulw/Wg4o4vfqYwxHsjt0NOnwHhgCPBwyPJUVd3jaSoT2dZOhLH3Quo26DAIuv0dipfx\nO5UxxiMnLRSquh/YD1wVvjgmoqXthgkPw7IvoPKZMHAE1LI2KsbEOuufYE5NFZZ/BeMfcg45dXkY\nOt8PCcX9TmaMCQMrFCZ3B7Y6TfzWjIMarWDAm1C1qd+pjDFhZIXCnJgq/PRvmPQ45GQ5U5K2vx3i\nrG2KMUWNFQpzvD3rYfRdsOF7OL0z9HsdKtbzO5UxxidWKMxRgRyY9w589xzEF3MKRMvrrYmfMUWc\nFQrj2LHSaeK3ZSE07A19X4NyNfxOZYyJAFYoirrswzDrNZj5CpQsB5f90+nTZE38jDFBViiKss0L\nnb2InSudDq+9XoQyFf1OZYyJMFYoiqLDh2Da8zDvbUisBld9Do16+Z3KGBOhrFAUNb/MdFqB793g\nzDbX82komeR3KmNMBLNCUVRk7Heuifjp33DaGfDnsVC3s9+pjDFRwApFUbBmvNPE7+AO6HgXdH0E\nits0ssYYd6xQxLK0XU5/puVfQZWmcOWnULOV36mMMVHGCkUsUoVlXzpFIjMVuj0Kne6xJn7GmHyx\nQhFr9m+GsffBzxOhZhuniV+VM/1OZYyJYlYoYkUgAAv/BZOfBM2BC4dAu79ZEz9jTIF52sRHRHqJ\nyBoRSRGRh0/w+DUislRElonIHBE528s8MWv3Ovh3P/j2PmcM4rY50ME6vRpjCodnexQiEg+8BfQE\nNgMLRGS0qq4MWe0XoIuq7hWR3sBwoJ1XmWJOTjbMewum/R/El4D+b0LLa639hjGmUHl56KktkKKq\n6wFEZCQwADhSKFR1Tsj684BaHuaJLduXO+03ti6CRhfBRa9Cuep+pzLGxCAvC0VNYFPI/c3kvrdw\nEzD+RA+IyC3ALQC1k+t4e7ws0mVnOg38Zr0GpSrAnz6CJhfbXoQxxjMRMZgtIt1wCsW5J3pcVYfj\nHJaiZavWujeM2SLKpvkwahDsWgPNr4ReQ6D0aX6nMsbEOC8LxRagdsj9WsFlfyAizYEPgN6quvtU\nL7p5bzplgJLFitBA7eE0mPos/PAulKsJ13wJDXr6ncoYU0R4WSgWAA1EpC5OgbgSuDp0BRFJBr4G\nrlPVtW5e9EBGFte1rMmAFkVkUp1102DMXbBvI5xzM5z/pDNvhDHGhIlnhUJVs0VkEDARiAc+VNUV\nInJr8PF3gSeAisDb4hxjz1bVNrm9bpwIr13RwqvYkSN9H0x6FBZ9DKfVgxvHQ52OfqcyxhRBoqp+\nZ8iTUjUaavpWVzsf0WvVWPj2fkj7DTrdBV0GQ7FSfqcyxkQxEVl4qj/ETyYiBrNN0MGdMO5BWPkN\nVG0GV4+EGi39TmWMKeKsUEQCVVj6OUx42Bm47v44dLob4ov5ncwYY6xQ+G7fJhh7D6RMgdrtoP8b\nULmR36mMMeYIKxR+CQTgx3/ClKecPYreL8E5f4W4In05oTEmAlmh8MOun515qzfOhTO6Qb/XoUId\nv1MZY8wJWaEIp5xsmDMMpr8AxUrCgLehxdXWfsMYE9GsUITLtqVOE79tS+DMftDnVShb1e9Uxhhz\nSlYovJaVATNfgllDoXRFGDgCmgzwO5UxxrhmhcJLG+c5Tfx2/wwtroELnrMmfsaYqGOFwguZB2Hq\nMzB/OCTVhmu/hvrn+53KGGPyxQpFYUuZCmPugf2boO0tcP4TUCLR71TGGJNvVigKy6E9MOkxWPwJ\nVGwAf5kAye39TmWMMQVmhaIwrBwF3z4Ah3ZD5/vhvIec01+NMSYGWKEoiNTtMO4BWDUGqjWHa7+C\n6s39TmWMMYXKCkV+qMLiT2HiI87prz2egg6DrImfMSYmWaHIq72/wpi7Yf00SO7gNPGr1MDvVMYY\n4xkrFG4FArDgfZjytNNyo88r0OYma+JnjIl5Vijc+G2N08Rv0w9Qvwf0/QeUT/Y7lTHGhIUVitzk\nZMHs12HGi1C8DFzyHjS/wpr4GWOKFCsUJ7N1sdN+Y8cyaHIx9HkZEqv4ncoYY8LOCsWxstKdNuBz\n3oAyleCKj51ur8YYU0RZoQj16xxnLGJ3CrS8Di54FkpV8DuVMcb4ygoFQGaqMyXpgg+cQerrvoF6\n3fxOZYwxEcEKxc+TnSZ+B7ZA+9uh+2POwLUxxhigKBeKQ3tgwiOwdCRUagQ3TYLabf1OZYwxEafo\nFQpVWPE/GPcgZOxzGvid9wAklPA7mTHGRKSiVSgObHOa+K0eC9VbwPWjoNpZfqcyxpiIVjQKhSos\n+g9MfAxyMqHnM9D+DogvGt++McYUROx/Uu75xWni98sMqNPJaeJXsZ7fqYwxJmrEbqEI5MAP78F3\nz4LEw0WvQesbrYmfMcbkUWwWip2rYfQg2LwAGlzgNPFLquV3KmOMiUqxVSiyD8PsoTDjJShRFi59\nH5r9yZr4GWNMAcROodiyEEbdCTtXwFmXQa8XIbGy36mMMSbqRX+hOHwIpg+BuW9CYlW48jNo3Mfv\nVMYYEzOiu1BsmOU08duzHlr92WniVzLJ71TGGBNTPD0FSER6icgaEUkRkYdP8LiIyLDg40tFpJWr\nF844AGPvhY8uAg3A9aOh/zArEsYY4wHP9ihEJB54C+gJbAYWiMhoVV0ZslpvoEHwqx3wTvDfkypL\nGrzdHlK3QYdB0O1RKF7am2/CGGOMp4ee2gIpqroeQERGAgOA0EIxABihqgrME5HyIlJdVbed7EWT\n2Q4lasPAEVCrjYfxjTHGgLeFoiawKeT+Zo7fWzjROjWBPxQKEbkFuCV4N1MG/bCcQecUbtroVAnY\n5XeICGHb4ijbFkfZtjiqUX6fGBWD2ao6HBgOICI/qqrtSmDbIpRti6NsWxxl2+IoEfkxv8/1cjB7\nC1A75H6t4LK8rmOMMcZHXhaKBUADEakrIsWBK4HRx6wzGrg+ePZTe2B/buMTxhhjws+zQ0+qmi0i\ng4CJQDzwoaquEJFbg4+/C4wD+gApwCHgRhcvPdyjyNHItsVRti2Osm1xlG2Lo/K9LcQ54cgYY4w5\nMeu5bYwxJldWKIwxxuQqYguFZ+0/opCLbXFNcBssE5E5InK2HznD4VTbImS9c0QkW0QuD2e+cHKz\nLUSkq4gsFpEVIjIj3BnDxcXvSJKIjBGRJcFt4WY8NOqIyIcislNElp/k8fx9bqpqxH3hDH6vA84A\nigNLgCbHrNMHGA8I0B74we/cPm6LjkCF4O3eRXlbhKz3Hc7JEpf7ndvHn4vyOJ0QkoP3q/id28dt\n8XfgxeDtysAeoLjf2T3YFucBrYDlJ3k8X5+bkbpHcaT9h6oeBn5v/xHqSPsPVZ0HlBeR6uEOGgan\n3BaqOkdV9wbvzsO5HiUWufm5ALgT+ArYGc5wYeZmW1wNfK2qGwFUNVa3h5ttoUBZEREgEadQZIc3\npvdUdSbO93Yy+frcjNRCcbLWHnldJxbk9fu8Cecvhlh0ym0hIjWBS3AaTMYyNz8XDYEKIjJdRBaK\nyPVhSxdebrbFm8CZwFZgGXC3qgbCEy+i5OtzMypaeBh3RKQbTqE41+8sPhoKDFbVgNgUuAlAa+B8\noBQwV0Tmqepaf2P54kJgMdAdqAdMFpHvVfWAv7GiQ6QWCmv/cZSr71NEmgMfAL1VdXeYsoWbm23R\nBhgZLBKVgD4ikq2q34QnYti42Rabgd2qmgakichM4Gwg1gqFm21xI/CCOgfqU0TkF6AxMD88ESNG\nvj43I/XQk7X/OOqU20JEkoGvgeti/K/FU24LVa2rqqer6unAl8DtMVgkwN3vyCjgXBFJEJHSON2b\nV4U5Zzi42RYbcfasEJGqOJ1U14c1ZWTI1+dmRO5RqHftP6KOy23xBFAReDv4l3S2xmDHTJfbokhw\nsy1UdZWITACWAgHgA1U94WmT0czlz8WzwEcisgznjJ/Bqhpz7cdF5DOgK1BJRDYDTwLFoGCfm9bC\nwxhjTK4i9dCTMcaYCGGFwhhjTK6sUBhjjMmVFQpjjDG5skJhjDEmV1YoTEQTkbtEZJWIfJLLOl1F\nZGw4c52MiPT/vXupiFwsIk1CHntGRHqEMUtXEekYrvczsSsir6MwJsTtQA9V3ex3EDdUdTRHL/a6\nGBiL08EVVX2isN9PRBJU9WTN7boCB4E5hf2+pmixPQoTsUTkXZzW0eNF5F4RaSsic0VkUXDejUYn\neE6X4PwLi4PrlQ0uf1BEFgR78D99kvc7KCL/CM5XMFVEKgeXtxCRecHn/k9EKgSX3yUiK4PLRwaX\n3SAibwb/ku8PvBzMUk9EPhKRy8WZO+G/Ie97ZI9IRC4Ifo8/ich/RSTxBDmni8hQEfkRuFtE+onI\nD8Hvd4qIVBWR04FbgXuD799ZRCqLyFfB7bBARDoV4L/HFCV+90+3L/vK7QvYAFQK3i4HJARv9wC+\nCt7uCowN3h4DdAreTsTZa74AZ2J5wfnjaCxw3gneS4FrgrefAN4M3l4KdAnefgYYGry9FSgRvF0+\n+O8NIc/7iJD5MH6/H8y0ESgTXP4OcC1Ob6qZIcsHA0+cIOd04O2Q+xU4evHszcCrwdtPAQ+ErPcp\ncG7wdjKwyu//X/uKji879GSiSRLwbxFpgPOhXuwE68wGXguOaXytqptF5AKcYrEouE4i0ADnQzlU\nAPg8ePtj4GsRScIpAr/PDvdv4Pe9gaXAJyLyDeC6n5Q6LScmAP1E5EvgIuAhoAvQBJgdbMVSHJh7\nkpf5POR2LeBzceYVKA78cpLn9ACayNGuuuVEJFFVD7rNboomKxQmmjwLTFPVS4KHVqYfu4KqviAi\n3+L0s5ktIhfi7EkMUdX38vh+p+pvcxHOjGL9gEdFpFkeXnskMAhnkpkfVTVVnE/wyap6lYvnp4Xc\nfgN4TVVHi0hXnD2JE4kD2qtqRh5yGmNjFCaqJHG0JfINJ1pBROqp6jJVfRGnq2hjnGZxf/n9eL+I\n1BSRKid4ehzOoSFwZoebpar7gb0i0jm4/DpghojEAbVVdRrOIaIknD2VUKlA2ZN8LzNwpqz8K07R\nAGd2wk4iUj+Ys4yINDzJ80OFbpc/5/L+k3Bm/yP4+i1cvLYxVihMVHkJGCIiizj53vA9IrJcRJYC\nWcB4VZ2Ec3x+brB76Jec+AM8DWgrzsT03XHGI8D58H05+JotgsvjgY+Dr7cIGKaq+455vZHAg8FB\n5nqhD6hqDs5YSe/gv6jqbzgF8LPge83FKXSn8hTwXxFZCIR2RB0DXPL7YDZwF9AmOPi+Emew25hT\nsu6xxgSJyEFVPe4sI2OKOtujMMYYkyvbozDGGJMr26MwxhiTKysUxhhjcmWFwhhjTK6sUBhjjMmV\nFQpjjDG5+n9WlPXjkL+VvQAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f627a988550>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Compute classification performance.\n",
    "from sklearn.metrics import accuracy_score, roc_auc_score, roc_curve\n",
    "\n",
    "train_accuracy = accuracy_score(y_train, pred_train)\n",
    "test_accuracy = accuracy_score(y_test, pred_test)\n",
    "train_auc = roc_auc_score(y_train, pred_prob_train)\n",
    "test_auc = roc_auc_score(y_test, pred_prob_test)\n",
    "test_fpr, test_tpr, thresholds = roc_curve(y_test, pred_prob_test)\n",
    "\n",
    "print('Train accuracy on default threshold: {}\\narea under curve: {}'.format(train_accuracy, train_auc))\n",
    "print('Test accuracy on default threshold: {}\\narea under curve: {}'.format(test_accuracy, test_auc))\n",
    "\n",
    "plot_auc(test_fpr, test_tpr)"
   ]
  },
  {
   "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
}