Ensemble.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Machine Learning Models and Ensemble Method\n",
    "---\n",
    "\n",
    "1. Split X-Features and y-labels\n",
    "2. 80-20 train-validation split\n",
    "3. fit models (on train), evaluate (on validation):\n",
    "    1. DNN\n",
    "    2. SVM\n",
    "    3. RF\n",
    "    4. XGB\n",
    "    5. LogReg\n",
    "4. __Manual__ Ensemble Method:\n",
    "    1. Evaluate Ensemble on Validation data\n",
    "5. Evaluate Ensemble on TEST data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/Users/rezanaghshineh/opt/anaconda3/lib/python3.7/site-packages/statsmodels/tools/_testing.py:19: FutureWarning: pandas.util.testing is deprecated. Use the functions in the public API at pandas.testing instead.\n",
      "  import pandas.util.testing as tm\n"
     ]
    }
   ],
   "source": [
    "import os\n",
    "os.environ[\"CUDA_VISIBLE_DEVICES\"] = \"-1\" # disable GPU\n",
    "from tqdm import tqdm # progress bar\n",
    "import pandas as pd\n",
    "import numpy as np\n",
    "import matplotlib.pyplot as plt\n",
    "import seaborn as sns\n",
    "%matplotlib inline\n",
    "\n",
    "# processing / validation\n",
    "from sklearn.model_selection import train_test_split\n",
    "from sklearn.preprocessing import MinMaxScaler\n",
    "\n",
    "# keras/tf\n",
    "import tensorflow as tf\n",
    "from tensorflow.keras.models import Sequential\n",
    "from tensorflow.keras.layers import Dense, Activation,Dropout\n",
    "\n",
    "# models\n",
    "from sklearn.svm import SVC\n",
    "from sklearn.ensemble import RandomForestRegressor\n",
    "from xgboost import XGBClassifier\n",
    "from sklearn.linear_model import LogisticRegression\n",
    "#from sklearn.model_selection import GridSearchCV # hp-tuning\n",
    "\n",
    "# metrics\n",
    "from sklearn import metrics\n",
    "from sklearn.metrics import classification_report\n",
    "from sklearn.metrics import roc_auc_score\n",
    "\n",
    "# constant seed for reproducibility\n",
    "SEED = 111 \n",
    "os.environ['PYTHONHASHSEED'] = str(SEED)\n",
    "np.random.seed(SEED)\n",
    "tf.random.set_seed(SEED)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 1 & 2 - X/y + Train/Test Split:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "df = pd.read_csv(\"data/UFC_TRAIN.csv\")\n",
    "\n",
    "# tackling imbalance issue\n",
    "#theMin = df[\"Winner\"].value_counts().min()\n",
    "#minority = df[df[\"Winner\"]==1].iloc[0:theMin]\n",
    "#undersampleMaj = df[df[\"Winner\"]==0].iloc[0:theMin]\n",
    "#df = pd.concat([minority, undersampleMaj], axis=0)\n",
    "#df[\"Winner\"].value_counts()\n",
    "\n",
    "# feature/label and train/test split\n",
    "X = df.drop([\"date\",\"Winner\",\"B_fighter\",\"R_fighter\"], axis=1).values\n",
    "y = df[\"Winner\"].values\n",
    "X_TRAIN, X_VAL, y_TRAIN, y_VAL = train_test_split(X,y, test_size=0.20, random_state=SEED)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Baseline: Always predict red (i.e: 0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.6779468544174426"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "metrics.accuracy_score(np.zeros(len(df.index)),df[\"Winner\"])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Baseline accuracy is 67.96 % in unbalanced dataset"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 3- ML Models\n",
    "### A: DNN - Using a deep neural network with early stopping functionality to prevent divergence of loss & val_loss:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "X_train_scaled shape: (3522, 42) | X_val_scaled shape: (881, 42) | y_train shape: (3522,) | y_val shape: (881,)\n"
     ]
    }
   ],
   "source": [
    "# scaling\n",
    "scaler = MinMaxScaler()\n",
    "scaler.fit(X_TRAIN)\n",
    "X_train_scaled = scaler.transform(X_TRAIN)\n",
    "X_val_scaled = scaler.transform(X_VAL)\n",
    "\n",
    "print(f\"X_train_scaled shape: {X_train_scaled.shape} | X_val_scaled shape: {X_val_scaled.shape} | y_train shape: {y_TRAIN.shape} | y_val shape: {y_VAL.shape}\")\n",
    "\n",
    "# model\n",
    "dnnClf = Sequential()\n",
    "\n",
    "# first hiden layer\n",
    "dnnClf.add(Dense(units=20, input_dim=42,activation='relu'))\n",
    "#dnnClf.add(Dropout(0.5)) # deactivates 50% of nodes\n",
    "\n",
    "dnnClf.add(Dense(units=10, activation='relu'))\n",
    "dnnClf.add(Dropout(0.5)) # deactivates 50% of nodes\n",
    "\n",
    "# output layer\n",
    "dnnClf.add(Dense(units=1, activation='sigmoid'))\n",
    "\n",
    "dnnClf.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n",
    "\n",
    "from tensorflow.keras.callbacks import EarlyStopping # prevent divergence of loss & val_loss\n",
    "early_stop = EarlyStopping(monitor='val_loss', mode='min', verbose=1, patience=16)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 3522 samples, validate on 881 samples\n",
      "Epoch 1/400\n",
      "3522/3522 [==============================] - 2s 619us/sample - loss: 0.6603 - accuracy: 0.6357 - val_loss: 0.6327 - val_accuracy: 0.6697\n",
      "Epoch 2/400\n",
      "3522/3522 [==============================] - 0s 126us/sample - loss: 0.6358 - accuracy: 0.6792 - val_loss: 0.6238 - val_accuracy: 0.6697\n",
      "Epoch 3/400\n",
      "3522/3522 [==============================] - 0s 126us/sample - loss: 0.6259 - accuracy: 0.6769 - val_loss: 0.6176 - val_accuracy: 0.6708\n",
      "Epoch 4/400\n",
      "3522/3522 [==============================] - 0s 130us/sample - loss: 0.6218 - accuracy: 0.6826 - val_loss: 0.6144 - val_accuracy: 0.6697\n",
      "Epoch 5/400\n",
      "3522/3522 [==============================] - 0s 127us/sample - loss: 0.6170 - accuracy: 0.6806 - val_loss: 0.6114 - val_accuracy: 0.6731\n",
      "Epoch 6/400\n",
      "3522/3522 [==============================] - 0s 137us/sample - loss: 0.6160 - accuracy: 0.6780 - val_loss: 0.6119 - val_accuracy: 0.6708\n",
      "Epoch 7/400\n",
      "3522/3522 [==============================] - 0s 126us/sample - loss: 0.6060 - accuracy: 0.6860 - val_loss: 0.6059 - val_accuracy: 0.6776\n",
      "Epoch 8/400\n",
      "3522/3522 [==============================] - 0s 123us/sample - loss: 0.6046 - accuracy: 0.6817 - val_loss: 0.6085 - val_accuracy: 0.6879\n",
      "Epoch 9/400\n",
      "3522/3522 [==============================] - 0s 124us/sample - loss: 0.6048 - accuracy: 0.6823 - val_loss: 0.6070 - val_accuracy: 0.6879\n",
      "Epoch 10/400\n",
      "3522/3522 [==============================] - 0s 123us/sample - loss: 0.6047 - accuracy: 0.6783 - val_loss: 0.6030 - val_accuracy: 0.6833\n",
      "Epoch 11/400\n",
      "3522/3522 [==============================] - 0s 125us/sample - loss: 0.6047 - accuracy: 0.6848 - val_loss: 0.6034 - val_accuracy: 0.6856\n",
      "Epoch 12/400\n",
      "3522/3522 [==============================] - 0s 124us/sample - loss: 0.6023 - accuracy: 0.6829 - val_loss: 0.6023 - val_accuracy: 0.6901\n",
      "Epoch 13/400\n",
      "3522/3522 [==============================] - 0s 124us/sample - loss: 0.6018 - accuracy: 0.6800 - val_loss: 0.6020 - val_accuracy: 0.6731\n",
      "Epoch 14/400\n",
      "3522/3522 [==============================] - 0s 124us/sample - loss: 0.6004 - accuracy: 0.6843 - val_loss: 0.6025 - val_accuracy: 0.6765\n",
      "Epoch 15/400\n",
      "3522/3522 [==============================] - 1s 201us/sample - loss: 0.5991 - accuracy: 0.6797 - val_loss: 0.6042 - val_accuracy: 0.6799\n",
      "Epoch 16/400\n",
      "3522/3522 [==============================] - 1s 148us/sample - loss: 0.5982 - accuracy: 0.6871 - val_loss: 0.6008 - val_accuracy: 0.6822\n",
      "Epoch 17/400\n",
      "3522/3522 [==============================] - 1s 165us/sample - loss: 0.5946 - accuracy: 0.6848 - val_loss: 0.6020 - val_accuracy: 0.6879\n",
      "Epoch 18/400\n",
      "3522/3522 [==============================] - 0s 125us/sample - loss: 0.5955 - accuracy: 0.6877 - val_loss: 0.6017 - val_accuracy: 0.6754\n",
      "Epoch 19/400\n",
      "3522/3522 [==============================] - 0s 130us/sample - loss: 0.5946 - accuracy: 0.6871 - val_loss: 0.6012 - val_accuracy: 0.6799\n",
      "Epoch 20/400\n",
      "3522/3522 [==============================] - 0s 124us/sample - loss: 0.5952 - accuracy: 0.6885 - val_loss: 0.6037 - val_accuracy: 0.6799\n",
      "Epoch 21/400\n",
      "3522/3522 [==============================] - 0s 124us/sample - loss: 0.5911 - accuracy: 0.6885 - val_loss: 0.6016 - val_accuracy: 0.6754\n",
      "Epoch 22/400\n",
      "3522/3522 [==============================] - 1s 163us/sample - loss: 0.5965 - accuracy: 0.6848 - val_loss: 0.6014 - val_accuracy: 0.6856\n",
      "Epoch 23/400\n",
      "3522/3522 [==============================] - 0s 140us/sample - loss: 0.5923 - accuracy: 0.6894 - val_loss: 0.6003 - val_accuracy: 0.6822\n",
      "Epoch 24/400\n",
      "3522/3522 [==============================] - 0s 132us/sample - loss: 0.5930 - accuracy: 0.6860 - val_loss: 0.5998 - val_accuracy: 0.6822\n",
      "Epoch 25/400\n",
      "3522/3522 [==============================] - 0s 127us/sample - loss: 0.5894 - accuracy: 0.6880 - val_loss: 0.6005 - val_accuracy: 0.6776\n",
      "Epoch 26/400\n",
      "3522/3522 [==============================] - 0s 127us/sample - loss: 0.5914 - accuracy: 0.6917 - val_loss: 0.6003 - val_accuracy: 0.6867\n",
      "Epoch 27/400\n",
      "3522/3522 [==============================] - 0s 131us/sample - loss: 0.5879 - accuracy: 0.6854 - val_loss: 0.5993 - val_accuracy: 0.6879\n",
      "Epoch 28/400\n",
      "3522/3522 [==============================] - 0s 123us/sample - loss: 0.5833 - accuracy: 0.6888 - val_loss: 0.5995 - val_accuracy: 0.6822\n",
      "Epoch 29/400\n",
      "3522/3522 [==============================] - 0s 124us/sample - loss: 0.5878 - accuracy: 0.6848 - val_loss: 0.5988 - val_accuracy: 0.6867\n",
      "Epoch 30/400\n",
      "3522/3522 [==============================] - 0s 133us/sample - loss: 0.5862 - accuracy: 0.6914 - val_loss: 0.5983 - val_accuracy: 0.6901\n",
      "Epoch 31/400\n",
      "3522/3522 [==============================] - 0s 140us/sample - loss: 0.5880 - accuracy: 0.6882 - val_loss: 0.5980 - val_accuracy: 0.6822\n",
      "Epoch 32/400\n",
      "3522/3522 [==============================] - 1s 188us/sample - loss: 0.5875 - accuracy: 0.6877 - val_loss: 0.5980 - val_accuracy: 0.6822\n",
      "Epoch 33/400\n",
      "3522/3522 [==============================] - 1s 164us/sample - loss: 0.5851 - accuracy: 0.6934 - val_loss: 0.5995 - val_accuracy: 0.6856\n",
      "Epoch 34/400\n",
      "3522/3522 [==============================] - 1s 362us/sample - loss: 0.5857 - accuracy: 0.6914 - val_loss: 0.5985 - val_accuracy: 0.6856\n",
      "Epoch 35/400\n",
      "3522/3522 [==============================] - 1s 327us/sample - loss: 0.5815 - accuracy: 0.6894 - val_loss: 0.5989 - val_accuracy: 0.6844\n",
      "Epoch 36/400\n",
      "3522/3522 [==============================] - 1s 163us/sample - loss: 0.5813 - accuracy: 0.6931 - val_loss: 0.5983 - val_accuracy: 0.6810\n",
      "Epoch 37/400\n",
      "3522/3522 [==============================] - 0s 138us/sample - loss: 0.5825 - accuracy: 0.6931 - val_loss: 0.5996 - val_accuracy: 0.6879\n",
      "Epoch 38/400\n",
      "3522/3522 [==============================] - 0s 135us/sample - loss: 0.5861 - accuracy: 0.6826 - val_loss: 0.6004 - val_accuracy: 0.6799\n",
      "Epoch 39/400\n",
      "3522/3522 [==============================] - 0s 138us/sample - loss: 0.5806 - accuracy: 0.6905 - val_loss: 0.5979 - val_accuracy: 0.6879\n",
      "Epoch 40/400\n",
      "3522/3522 [==============================] - 0s 137us/sample - loss: 0.5807 - accuracy: 0.6988 - val_loss: 0.5985 - val_accuracy: 0.6867\n",
      "Epoch 41/400\n",
      "3522/3522 [==============================] - 0s 137us/sample - loss: 0.5812 - accuracy: 0.6951 - val_loss: 0.6056 - val_accuracy: 0.6810\n",
      "Epoch 42/400\n",
      "3522/3522 [==============================] - 0s 139us/sample - loss: 0.5814 - accuracy: 0.6914 - val_loss: 0.5982 - val_accuracy: 0.6844\n",
      "Epoch 43/400\n",
      "3522/3522 [==============================] - 0s 137us/sample - loss: 0.5801 - accuracy: 0.6902 - val_loss: 0.5984 - val_accuracy: 0.6833\n",
      "Epoch 44/400\n",
      "3522/3522 [==============================] - 1s 143us/sample - loss: 0.5798 - accuracy: 0.6922 - val_loss: 0.5973 - val_accuracy: 0.6890\n",
      "Epoch 45/400\n",
      "3522/3522 [==============================] - 0s 138us/sample - loss: 0.5806 - accuracy: 0.6880 - val_loss: 0.5972 - val_accuracy: 0.6890\n",
      "Epoch 46/400\n",
      "3522/3522 [==============================] - 0s 137us/sample - loss: 0.5789 - accuracy: 0.6917 - val_loss: 0.5975 - val_accuracy: 0.6890\n",
      "Epoch 47/400\n",
      "3522/3522 [==============================] - 0s 136us/sample - loss: 0.5766 - accuracy: 0.6919 - val_loss: 0.5973 - val_accuracy: 0.6890\n",
      "Epoch 48/400\n",
      "3522/3522 [==============================] - 0s 138us/sample - loss: 0.5780 - accuracy: 0.6988 - val_loss: 0.5997 - val_accuracy: 0.6901\n",
      "Epoch 49/400\n",
      "3522/3522 [==============================] - 0s 139us/sample - loss: 0.5759 - accuracy: 0.6979 - val_loss: 0.5988 - val_accuracy: 0.6924\n",
      "Epoch 50/400\n",
      "3522/3522 [==============================] - 0s 137us/sample - loss: 0.5744 - accuracy: 0.6948 - val_loss: 0.5995 - val_accuracy: 0.6856\n",
      "Epoch 51/400\n",
      "3522/3522 [==============================] - 0s 136us/sample - loss: 0.5696 - accuracy: 0.6985 - val_loss: 0.5980 - val_accuracy: 0.6879\n",
      "Epoch 52/400\n",
      "3522/3522 [==============================] - 0s 137us/sample - loss: 0.5724 - accuracy: 0.6985 - val_loss: 0.5988 - val_accuracy: 0.6867\n",
      "Epoch 53/400\n",
      "3522/3522 [==============================] - 0s 139us/sample - loss: 0.5783 - accuracy: 0.6905 - val_loss: 0.5970 - val_accuracy: 0.6844\n",
      "Epoch 54/400\n",
      "3522/3522 [==============================] - 0s 138us/sample - loss: 0.5751 - accuracy: 0.6959 - val_loss: 0.5973 - val_accuracy: 0.6924\n",
      "Epoch 55/400\n",
      "3522/3522 [==============================] - 0s 136us/sample - loss: 0.5734 - accuracy: 0.6914 - val_loss: 0.5972 - val_accuracy: 0.6924\n",
      "Epoch 56/400\n",
      "3522/3522 [==============================] - 0s 138us/sample - loss: 0.5699 - accuracy: 0.6965 - val_loss: 0.5991 - val_accuracy: 0.6890\n",
      "Epoch 57/400\n",
      "3522/3522 [==============================] - 0s 136us/sample - loss: 0.5714 - accuracy: 0.6979 - val_loss: 0.5971 - val_accuracy: 0.6833\n",
      "Epoch 58/400\n",
      "3522/3522 [==============================] - 0s 138us/sample - loss: 0.5701 - accuracy: 0.6990 - val_loss: 0.5964 - val_accuracy: 0.6890\n",
      "Epoch 59/400\n",
      "3522/3522 [==============================] - 0s 137us/sample - loss: 0.5708 - accuracy: 0.6990 - val_loss: 0.5976 - val_accuracy: 0.6856\n",
      "Epoch 60/400\n",
      "3522/3522 [==============================] - 0s 138us/sample - loss: 0.5707 - accuracy: 0.6979 - val_loss: 0.5978 - val_accuracy: 0.6856\n",
      "Epoch 61/400\n",
      "3522/3522 [==============================] - 0s 140us/sample - loss: 0.5731 - accuracy: 0.6931 - val_loss: 0.5970 - val_accuracy: 0.6890\n",
      "Epoch 62/400\n",
      "3522/3522 [==============================] - 0s 138us/sample - loss: 0.5668 - accuracy: 0.6939 - val_loss: 0.5964 - val_accuracy: 0.6844\n",
      "Epoch 63/400\n",
      "3522/3522 [==============================] - 0s 139us/sample - loss: 0.5675 - accuracy: 0.7005 - val_loss: 0.5972 - val_accuracy: 0.6822\n",
      "Epoch 64/400\n",
      "3522/3522 [==============================] - 0s 138us/sample - loss: 0.5720 - accuracy: 0.6979 - val_loss: 0.5962 - val_accuracy: 0.6856\n",
      "Epoch 65/400\n",
      "3522/3522 [==============================] - 0s 137us/sample - loss: 0.5658 - accuracy: 0.6993 - val_loss: 0.5972 - val_accuracy: 0.6799\n",
      "Epoch 66/400\n",
      "3522/3522 [==============================] - 0s 138us/sample - loss: 0.5659 - accuracy: 0.7039 - val_loss: 0.5984 - val_accuracy: 0.6810\n",
      "Epoch 67/400\n",
      "3522/3522 [==============================] - 0s 138us/sample - loss: 0.5641 - accuracy: 0.7013 - val_loss: 0.5978 - val_accuracy: 0.6833\n",
      "Epoch 68/400\n",
      "3522/3522 [==============================] - 0s 134us/sample - loss: 0.5654 - accuracy: 0.6999 - val_loss: 0.5981 - val_accuracy: 0.6822\n",
      "Epoch 69/400\n",
      "3522/3522 [==============================] - 0s 136us/sample - loss: 0.5699 - accuracy: 0.6979 - val_loss: 0.5994 - val_accuracy: 0.6867\n",
      "Epoch 70/400\n",
      "3522/3522 [==============================] - 0s 139us/sample - loss: 0.5666 - accuracy: 0.6942 - val_loss: 0.5973 - val_accuracy: 0.6867\n",
      "Epoch 71/400\n",
      "3522/3522 [==============================] - 0s 139us/sample - loss: 0.5655 - accuracy: 0.6993 - val_loss: 0.5990 - val_accuracy: 0.6822\n",
      "Epoch 72/400\n",
      "3522/3522 [==============================] - 0s 137us/sample - loss: 0.5669 - accuracy: 0.6985 - val_loss: 0.5999 - val_accuracy: 0.6822\n",
      "Epoch 73/400\n",
      "3522/3522 [==============================] - 0s 139us/sample - loss: 0.5620 - accuracy: 0.6939 - val_loss: 0.6007 - val_accuracy: 0.6867\n",
      "Epoch 74/400\n",
      "3522/3522 [==============================] - 0s 137us/sample - loss: 0.5591 - accuracy: 0.6999 - val_loss: 0.6025 - val_accuracy: 0.6822\n",
      "Epoch 75/400\n",
      "3522/3522 [==============================] - 0s 138us/sample - loss: 0.5557 - accuracy: 0.7118 - val_loss: 0.6017 - val_accuracy: 0.6890\n",
      "Epoch 76/400\n",
      "3522/3522 [==============================] - 0s 133us/sample - loss: 0.5593 - accuracy: 0.7016 - val_loss: 0.6025 - val_accuracy: 0.6901\n",
      "Epoch 77/400\n",
      "3522/3522 [==============================] - 0s 138us/sample - loss: 0.5601 - accuracy: 0.6976 - val_loss: 0.6045 - val_accuracy: 0.6822\n",
      "Epoch 78/400\n",
      "3522/3522 [==============================] - 0s 137us/sample - loss: 0.5610 - accuracy: 0.7002 - val_loss: 0.6048 - val_accuracy: 0.6765\n",
      "Epoch 79/400\n",
      "3522/3522 [==============================] - 0s 137us/sample - loss: 0.5581 - accuracy: 0.7007 - val_loss: 0.6023 - val_accuracy: 0.6856\n",
      "Epoch 80/400\n",
      "3522/3522 [==============================] - 0s 137us/sample - loss: 0.5593 - accuracy: 0.6965 - val_loss: 0.6037 - val_accuracy: 0.6947\n",
      "Epoch 00080: early stopping\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<matplotlib.axes._subplots.AxesSubplot at 0x1a458729d0>"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "dnnClf.fit(x=X_train_scaled, \n",
    "          y=y_TRAIN, \n",
    "          epochs=400,\n",
    "          validation_data=(X_val_scaled, y_VAL), verbose=1,\n",
    "          callbacks=[early_stop]\n",
    "          )\n",
    "model_loss = pd.DataFrame(dnnClf.history.history)\n",
    "model_loss.plot()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "DNN Validation Performance on UNBALANCED(!):\n",
      "------------------\n",
      "               precision    recall  f1-score   support\n",
      "\n",
      "     class 0       0.71      0.91      0.80       590\n",
      "     class 1       0.58      0.25      0.35       291\n",
      "\n",
      "    accuracy                           0.69       881\n",
      "   macro avg       0.64      0.58      0.57       881\n",
      "weighted avg       0.67      0.69      0.65       881\n",
      "\n",
      "AUC:  0.5787960859688974\n"
     ]
    }
   ],
   "source": [
    "dnnPreds = dnnClf.predict(scaler.transform(X_VAL))\n",
    "dnnPreds = [round(i[0]) for i in dnnPreds]\n",
    "target_names = ['class 0', 'class 1']\n",
    "print(\"DNN Validation Performance on UNBALANCED(!):\\n------------------\\n\",classification_report(y_VAL, dnnPreds , target_names=target_names))\n",
    "print(\"AUC: \",roc_auc_score(y_VAL, dnnPreds))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### B: SVM - Support Vector Machine:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "SVM Validation Performance on UNBALANCED(!):\n",
      "------------------\n",
      "               precision    recall  f1-score   support\n",
      "\n",
      "     class 0       0.68      0.99      0.81       590\n",
      "     class 1       0.71      0.05      0.10       291\n",
      "\n",
      "    accuracy                           0.68       881\n",
      "   macro avg       0.70      0.52      0.45       881\n",
      "weighted avg       0.69      0.68      0.57       881\n",
      "\n",
      "AUC:  0.5206884501135768\n"
     ]
    }
   ],
   "source": [
    "svmClf = SVC(kernel=\"linear\")\n",
    "svmClf.fit(X_TRAIN,y_TRAIN)\n",
    "svmPreds = svmClf.predict(X_VAL)\n",
    "print(\"SVM Validation Performance on UNBALANCED(!):\\n------------------\\n\",classification_report(y_VAL, svmPreds , target_names=target_names))\n",
    "print(\"AUC: \",roc_auc_score(y_VAL, svmPreds))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### C: RF - RandomForest:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "RF Validation Performance on UNBALANCED(!):\n",
      "------------------\n",
      "               precision    recall  f1-score   support\n",
      "\n",
      "     class 0       0.71      0.89      0.79       590\n",
      "     class 1       0.55      0.26      0.36       291\n",
      "\n",
      "    accuracy                           0.68       881\n",
      "   macro avg       0.63      0.58      0.57       881\n",
      "weighted avg       0.66      0.68      0.65       881\n",
      "\n",
      "AUC:  0.5780651173626886\n"
     ]
    }
   ],
   "source": [
    "rfClf = RandomForestRegressor(n_estimators = 1000)\n",
    "rfClf.fit(X_TRAIN, y_TRAIN)\n",
    "rfPreds = rfClf.predict(X_VAL)\n",
    "rfPreds = [round(i) for i in rfPreds]\n",
    "print(\"RF Validation Performance on UNBALANCED(!):\\n------------------\\n\",classification_report(y_VAL, rfPreds , target_names=target_names))\n",
    "print(\"AUC: \",roc_auc_score(y_VAL, rfPreds))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### D: XGB - Gradient Boost:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "XGB Validation Performance on UNBALANCED(!):\n",
      "------------------\n",
      "               precision    recall  f1-score   support\n",
      "\n",
      "     class 0       0.71      0.91      0.79       590\n",
      "     class 1       0.56      0.23      0.33       291\n",
      "\n",
      "    accuracy                           0.69       881\n",
      "   macro avg       0.63      0.57      0.56       881\n",
      "weighted avg       0.66      0.69      0.64       881\n",
      "\n",
      "AUC:  0.5702050206768011\n"
     ]
    }
   ],
   "source": [
    "xgbClf = XGBClassifier(n_estimators=200)\n",
    "xgbClf.fit(X_TRAIN, y_TRAIN)\n",
    "xgbPreds = xgbClf.predict(X_VAL)\n",
    "print(\"XGB Validation Performance on UNBALANCED(!):\\n------------------\\n\",classification_report(y_VAL, xgbPreds , target_names=target_names))\n",
    "print(\"AUC: \",roc_auc_score(y_VAL, xgbPreds))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### E: LR - Logistic Regression:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "LogReg Validation Performance on UNBALANCED(!):\n",
      "------------------\n",
      "               precision    recall  f1-score   support\n",
      "\n",
      "     class 0       0.70      0.92      0.80       590\n",
      "     class 1       0.56      0.20      0.29       291\n",
      "\n",
      "    accuracy                           0.68       881\n",
      "   macro avg       0.63      0.56      0.54       881\n",
      "weighted avg       0.65      0.68      0.63       881\n",
      "\n",
      "AUC:  0.5606733065408586\n"
     ]
    }
   ],
   "source": [
    "lrClf = LogisticRegression(solver=\"newton-cg\")\n",
    "lrClf.fit(X_TRAIN, y_TRAIN)\n",
    "lrPreds = lrClf.predict(X_VAL)\n",
    "print(\"LogReg Validation Performance on UNBALANCED(!):\\n------------------\\n\",classification_report(y_VAL, lrPreds , target_names=target_names))\n",
    "print(\"AUC: \",roc_auc_score(y_VAL, lrPreds))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 4- Ensemble Method with Validation Performance:\n",
    "Ensemble method aggregates the votes of each model and gives the most frequent vote as output"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 881/881 [02:23<00:00,  6.13it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Ensemble Validation Performance on UNBALANCED(!):\n",
      "------------------\n",
      "               precision    recall  f1-score   support\n",
      "\n",
      "     class 0       0.70      0.92      0.80       590\n",
      "     class 1       0.57      0.21      0.31       291\n",
      "\n",
      "    accuracy                           0.69       881\n",
      "   macro avg       0.64      0.57      0.55       881\n",
      "weighted avg       0.66      0.69      0.64       881\n",
      "\n",
      "AUC:  0.5658279457161163\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "def predictEnsemble(sample, models=0):\n",
    "    \"\"\"predicts the label of a given sample by aggregating votes of number of models.\n",
    "    by default, models = 0, takes into account all models. Otherwise, for a given list of codes,\n",
    "    it involves the corresponsing model. codes:\n",
    "    1: dnn | 2: svm | 3: rf | 4: xgb | 5: lr\n",
    "    \"\"\"\n",
    "    modelsDict = {\n",
    "        # models predictions dictionary\n",
    "        1:dnnClf.predict(scaler.transform(sample.reshape(1,-1))).tolist()[0][0],\n",
    "        2:svmClf.predict(sample.reshape(1,-1)).tolist()[0],\n",
    "        3:rfClf.predict(sample.reshape(1,-1)).tolist()[0],\n",
    "        4:xgbClf.predict(sample.reshape(1,-1)).tolist()[0],\n",
    "        5:lrClf.predict(sample.reshape(1,-1)).tolist()[0]\n",
    "    }\n",
    "    preds = []\n",
    "    if models == 0: # use all models\n",
    "        [preds.append(model) for model in modelsDict.values()]\n",
    "    else: # use only specified models\n",
    "        for model_code in models:\n",
    "            preds.append(modelsDict[model_code])\n",
    "\n",
    "    #print(preds)\n",
    "    preds = [round(i) for i in preds] # transform probability to label (threshold 0.5)\n",
    "    #print(preds)\n",
    "    #print(max(set(preds), key=preds.count))\n",
    "    return(max(set(preds), key=preds.count))\n",
    "\n",
    "\n",
    "ensPreds = []\n",
    "[ensPreds.append(predictEnsemble(sample, models=[3,4,5])) for sample in tqdm(X_VAL)]\n",
    "print(\"Ensemble Validation Performance on UNBALANCED(!):\\n------------------\\n\",classification_report(y_VAL, ensPreds , target_names=target_names))\n",
    "print(\"AUC: \",roc_auc_score(y_VAL, ensPreds))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 5- Performance Evaluation on TEST (unseen data)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 232/232 [00:36<00:00,  6.43it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Ensemble TEST Performance on UNBALANCED(!):\n",
      "------------------\n",
      "               precision    recall  f1-score   support\n",
      "\n",
      "     class 0       0.72      0.93      0.81       157\n",
      "     class 1       0.62      0.24      0.35        75\n",
      "\n",
      "    accuracy                           0.71       232\n",
      "   macro avg       0.67      0.58      0.58       232\n",
      "weighted avg       0.69      0.71      0.66       232\n",
      "\n",
      "AUC:  0.584968152866242\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "TEST = pd.read_csv(\"data/UFC_TEST.csv\")\n",
    "X_TEST = TEST.drop([\"date\",\"B_fighter\",\"R_fighter\",\"Winner\"],axis=1).values\n",
    "y_TEST = TEST[\"Winner\"].values\n",
    "\n",
    "ensPreds_TEST = []\n",
    "[ensPreds_TEST.append(predictEnsemble(test_sample, models=[3,4,5])) for test_sample in tqdm(X_TEST)]\n",
    "print(\"Ensemble TEST Performance on UNBALANCED(!):\\n------------------\\n\",classification_report(y_TEST, ensPreds_TEST , target_names=target_names))\n",
    "print(\"AUC: \",roc_auc_score(y_TEST, ensPreds_TEST))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# save models to disk\n",
    "#import pickle\n",
    "#dnnClf.save('resources/dnn_model.h5')\n",
    "#pickle.dump(svmClf, open('resources/svm_model.sav', 'wb'))\n",
    "#pickle.dump(rfClf, open('resources/rf_model.sav', 'wb'))\n",
    "#pickle.dump(xgbClf, open('resources/xgb_model.sav', 'wb'))\n",
    "#pickle.dump(lrClf, open('resources/lr_model.sav', 'wb'))\n",
    "#pickle.dump(scaler, open('resources/scaler.pkl', 'wb'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# notes:\n",
    "'''\n",
    "dnnPreds2 = dnnClf.predict(scaler.transform(X_TEST))\n",
    "dnnPreds2 = [round(i[0]) for i in dnnPreds2]\n",
    "dnnAcc2 = metrics.accuracy_score(dnnPreds2, y_TEST)\n",
    "print(\"DNN Accuracy:\",round(dnnAcc2,3))\n",
    "\n",
    "svmPreds2 = svmClf.predict(X_TEST)\n",
    "svmAcc2 = metrics.accuracy_score(svmPreds2, y_TEST)\n",
    "print(\"SVM Accuracy:\",round(svmAcc2,3))\n",
    "\n",
    "rfPreds2 = rfClf.predict(X_TEST)\n",
    "rfPreds2 = [round(i) for i in rfPreds2]\n",
    "rfAcc2 = metrics.accuracy_score(rfPreds2, y_TEST)\n",
    "print(\"RF Accuracy:\",round(rfAcc2,3))\n",
    "\n",
    "xgbPreds2 = xgbClf.predict(X_TEST)\n",
    "xgbAcc2 = metrics.accuracy_score(xgbPreds2, y_TEST)\n",
    "print(\"XGB Accuracy:\",round(xgbAcc2,3))\n",
    "\n",
    "lrPreds2 = lrClf.predict(X_TEST)\n",
    "lrAcc2 = metrics.accuracy_score(lrPreds2, y_TEST)\n",
    "print(\"LogReg Accuracy:\",round(lrAcc2,3))\n",
    "\n",
    "accTable = pd.DataFrame({\"Model\":[\"DNN\", \"SVM\", \"RF\", \"XGB\", \"LogReg\", \"Ensemble\"], \n",
    "                         \"Val_Accuracy\":[dnnAcc, svmAcc, rfAcc, xgbAcc, lrAcc, ensAcc],\n",
    "                         \"Test_Accuracy\":[dnnAcc2, svmAcc2, rfAcc2, xgbAcc2, lrAcc2, ensAcc2]})\n",
    "accTable.plot(kind=\"bar\",ylim=(0.5,0.8),x=\"Model\",title=\"Models Performance on Validation and Test Data\")\n",
    "\n",
    "# grid-search hyper-parameter tuning\n",
    "# svm hp-tuning with gridSearch\n",
    "#svm_param = {\"kernel\":(\"linear\",\"poly\",\"rbf\", \"sigmoid\"),\n",
    "#            \"C\":[1,52,10],\n",
    "#            \"degree\":[3,8],\n",
    "#            \"gamma\":(\"auto\",\"scale\"),\n",
    "#            \"coef0\":[0.001,10,0.5]}\n",
    "#svmClf = SVC()\n",
    "#svmGrid = GridSearchCV(svmClf, svm_param,cv=2)\n",
    "#svmGrid.fit(X_TRAIN, y_TRAIN)\n",
    "\n",
    "'''\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.4"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}