doordash_delivery_est/doordash_est.ipynb

{
 "cells": [
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Delivery Duration Prediction\n",
    "\n",
    "When a consumer places an order on DoorDash, we show the expected time of delivery. It is very important for DoorDash to get this right, as it has a big impact on consumer experience. In this exercise, you will build a model to predict the estimated time taken for a delivery.\n",
    "\n",
    "Concretely, for a given delivery you must predict the total delivery duration seconds , i.e., the time taken from:\n",
    "\n",
    "__Start:__ the time consumer submits the order `(created_at)` to\n",
    "<br>\n",
    "__End:__ when the order will be delivered to the consumer `(actual_delivery_time)`\n",
    "\n",
    "#### My Approach\n",
    "\n",
    "The dataset comes with two other predicted values: `estimated_order_place_duration` and `estimated_store_to_consumer_driving_duration`. The missing piece is the amount of time the order spends in the store. In addition to generic feature engineering, I will be creating separate order 'prep_time' statistics tables/dataframes to simulate gold fact tables in a database or features from a feature store (i.e. Feather, Feast) that would in turn live in a data lake.\n",
    "\n",
    "This is a fairly popular exercise so it is possible to derive performance objectives. My goal is to produce a feature set and model that has a test RMSE of less than 1200 seconds and a test MAE of less than 600 seconds, with a stretch goal of 600 seconds for the test RMSE. For reference, the best models publically available produce a test RMSE around 980 - 1000 seconds with a holdout size of 0.2. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 73,
   "metadata": {},
   "outputs": [],
   "source": [
    "import warnings\n",
    "warnings.simplefilter(action='ignore',category=FutureWarning)\n",
    "\n",
    "import pandas as pd\n",
    "import numpy as np\n",
    "import matplotlib.pyplot as plt\n",
    "import seaborn as sns\n",
    "from datetime import datetime as dt\n",
    "from datetime import timedelta\n",
    "\n",
    "from statsmodels.stats.outliers_influence import variance_inflation_factor\n",
    "from statsmodels.tools import add_constant\n",
    "from sklearn.model_selection import train_test_split\n",
    "from sklearn.model_selection import KFold\n",
    "from sklearn.model_selection import cross_val_score\n",
    "from sklearn.decomposition import PCA\n",
    "from sklearn.preprocessing import StandardScaler,MinMaxScaler\n",
    "from sklearn.metrics import mean_squared_error\n",
    "from sklearn.metrics import mean_absolute_error\n",
    "from sklearn.metrics import confusion_matrix\n",
    "from sklearn.metrics import accuracy_score\n",
    "from scipy.stats import linregress\n",
    "\n",
    "from sklearn.ensemble import RandomForestRegressor\n",
    "from sklearn.ensemble import GradientBoostingRegressor\n",
    "from flaml import AutoML\n",
    "from flaml.automl.model import XGBoostEstimator"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "np.random.seed(42)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "## update types\n",
    "## create target variable\n",
    "def addFeatures_durations(historical_data:pd.DataFrame) -> pd.DataFrame:\n",
    "    try:\n",
    "        historical_data['created_at'] = pd.to_datetime(historical_data['created_at'])\n",
    "        historical_data['actual_delivery_time'] = pd.to_datetime(historical_data['actual_delivery_time'])\n",
    "        ## calculate delivery duration\n",
    "        historical_data['actual_total_delivery_duration'] = (historical_data['actual_delivery_time'] - historical_data['created_at']).dt.total_seconds()\n",
    "        ## estimated time spent outside the store/not on order preparation\n",
    "        historical_data['est_time_non-prep'] = historical_data['estimated_order_place_duration'] + historical_data['estimated_store_to_consumer_driving_duration']\n",
    "    except Exception as ex:\n",
    "        raise ex\n",
    "    return historical_data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "def createBaseline(historical_data:pd.DataFrame) -> pd.DataFrame:   #this function would be done in practice by SQL or Spark\n",
    "    historical_data = addFeatures_durations(historical_data=historical_data)\n",
    "    try:\n",
    "        baseline_data = historical_data[['store_id','actual_total_delivery_duration']]\n",
    "        baseline_data = baseline_data.dropna()\n",
    "        baseline_data = pd.merge(\n",
    "            left=baseline_data,\n",
    "            right=baseline_data.groupby('store_id').actual_total_delivery_duration.aggregate([np.mean,np.std]),\n",
    "            how='inner',\n",
    "            on='store_id'\n",
    "        )\n",
    "        baseline_data['z-score'] = (baseline_data['actual_total_delivery_duration'] - baseline_data['mean']) / baseline_data['std']\n",
    "        baseline_data = baseline_data.loc[(baseline_data['z-score'] <=3) & (baseline_data['z-score'] >= -3)]\n",
    "        baseline_avgs = baseline_data.groupby('store_id').actual_total_delivery_duration.aggregate([np.mean])\n",
    "        baseline_data = pd.merge(\n",
    "            left=historical_data[['store_id','actual_total_delivery_duration']].dropna(),\n",
    "            right=baseline_avgs,\n",
    "            on='store_id'\n",
    "        )\n",
    "        baseline_MAE = mean_absolute_error(y_true=baseline_data['actual_total_delivery_duration'],y_pred=baseline_data['mean'])\n",
    "        baseline_RMSE = mean_squared_error(y_true=baseline_data['actual_total_delivery_duration'],y_pred=baseline_data['mean'],squared=False)\n",
    "    except Exception as ex:\n",
    "        raise ex\n",
    "    return {'baseline_MAE:':baseline_MAE,'baseline_RMSE:':baseline_RMSE}"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "## TODO #4 create estimated prep time features that are anchored to stores and/or categories -- DONE\n",
    "# the est_time_prep feature depends on the actual delivery datetime, which can't be provided to a model\n",
    "# BUT a set of values derived from historical data can be given to a model on demand\n",
    "\n",
    "def addFeatures_store_prep_stats(historical_data:pd.DataFrame) -> pd.DataFrame:   #this function would be done in practice by SQL or Spark\n",
    "## returns a separate dataframe with the min, max, median and mean and standard deviation of the est_prep_time_per_item for each store_id\n",
    "    try:\n",
    "        historical_data = addFeatures_durations(historical_data=historical_data)\n",
    "        ## calculate the amount of time (in seconds) the order spends in the store, then divide by the total number of items in the order\n",
    "        historical_data['est_time_prep_per_item'] = (historical_data['actual_total_delivery_duration'] - historical_data['est_time_non-prep']) / historical_data['total_items']\n",
    "        historical_data = historical_data[historical_data['est_time_prep_per_item'] > 0]\n",
    "        store_prep_stats_df = historical_data[['store_id','est_time_prep_per_item']].groupby('store_id').est_time_prep_per_item.aggregate([np.mean,np.std])\n",
    "        ## calculate z-score\n",
    "        store_prep_facts_df = pd.merge(\n",
    "            left=store_prep_stats_df.reset_index(),\n",
    "            right=historical_data[['store_id','est_time_prep_per_item']],\n",
    "            how='inner',\n",
    "            on='store_id'\n",
    "        )\n",
    "        store_prep_facts_df['z_score'] = (store_prep_facts_df['est_time_prep_per_item'] - store_prep_facts_df['mean']) / store_prep_facts_df['std']\n",
    "        ## drop outliers\n",
    "        store_prep_stats_df = store_prep_facts_df.loc[(store_prep_facts_df['z_score'] <=3) & (store_prep_facts_df['z_score'] >= -3)]\n",
    "        store_prep_stats_df = store_prep_stats_df.drop(columns=['mean','std'])\n",
    "        store_prep_stats_df = store_prep_stats_df.groupby('store_id').est_time_prep_per_item.aggregate(['min','max','median',np.mean,np.std])\n",
    "        store_prep_stats_df = store_prep_stats_df.add_prefix('store_est_time_prep_per_item_')\n",
    "    except Exception as ex:\n",
    "        raise ex\n",
    "    return store_prep_stats_df\n",
    "\n",
    "def addFeatures_category_prep_stats(historical_data:pd.DataFrame) -> pd.DataFrame:   #this function would be done in practice by SQL or Spark\n",
    "## returns a separate dataframe with the min,max, median, mean and standard deviation for each category\n",
    "    try:\n",
    "        historical_data = addFeatures_durations(historical_data=historical_data)\n",
    "        ## calculate the amount of time (in seconds) the order spends in the store, then divide by the total number of items in the order\n",
    "        historical_data['est_time_prep_per_item'] = (historical_data['actual_total_delivery_duration'] - historical_data['est_time_non-prep']) / historical_data['total_items']\n",
    "        historical_data = historical_data[historical_data['est_time_prep_per_item'] > 0]\n",
    "        category_prep_stats_df = historical_data[['clean_store_primary_category','est_time_prep_per_item']].groupby('clean_store_primary_category').est_time_prep_per_item.aggregate([np.mean,np.std])\n",
    "        ## calculate z-score\n",
    "        category_prep_facts_df = pd.merge(\n",
    "            left=category_prep_stats_df.reset_index(),\n",
    "            right=historical_data[['created_at','clean_store_primary_category','est_time_prep_per_item']],\n",
    "            how='inner',\n",
    "            on='clean_store_primary_category'\n",
    "        )\n",
    "        category_prep_facts_df['z_score'] = (category_prep_facts_df['est_time_prep_per_item'] - category_prep_facts_df['mean']) / category_prep_facts_df['std']\n",
    "        ## drop outliers\n",
    "        category_prep_stats_df = category_prep_facts_df.loc[(category_prep_facts_df['z_score'] <= 3) & (category_prep_facts_df['z_score'] >= -3)]\n",
    "        category_prep_stats_df = category_prep_stats_df.drop(columns=['mean','std','created_at'])\n",
    "        category_prep_stats_df = category_prep_stats_df.groupby('clean_store_primary_category').est_time_prep_per_item.aggregate(['min','max','median',np.mean,np.std])\n",
    "        category_prep_stats_df = category_prep_stats_df.add_prefix('category_est_time_prep_per_item_')\n",
    "    except Exception as ex:\n",
    "        raise ex\n",
    "    return category_prep_stats_df"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "## add time of day features\n",
    "\n",
    "def addFeatures_time_of_day(historical_data:pd.DataFrame) -> pd.DataFrame:\n",
    "    try:\n",
    "        historical_data['created_hour_of_day'] = historical_data['created_at'].dt.hour\n",
    "        historical_data['created_day_of_week'] = historical_data['created_at'].dt.day_of_week\n",
    "    except Exception as ex:\n",
    "        raise ex\n",
    "    return historical_data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "## add ratio features\n",
    "def addFeatures_ratios(historical_data:pd.DataFrame) -> pd.DataFrame:\n",
    "    try:\n",
    "        historical_data['available_dashers'] = historical_data['total_onshift_dashers'] - historical_data['total_busy_dashers']\n",
    "        historical_data['busy_to_onshift'] = historical_data['total_onshift_dashers'] / historical_data['total_busy_dashers']\n",
    "        historical_data['busy_to_onshift'].replace(to_replace=[np.inf,-np.inf],value=historical_data['total_onshift_dashers'][1],inplace=True)\n",
    "        historical_data['busy_to_outstanding'] = historical_data['total_busy_dashers'] / historical_data['total_outstanding_orders']\n",
    "        historical_data['busy_to_outstanding'].replace(to_replace=[np.inf,-np.inf,np.nan],value=0,inplace=True)\n",
    "        historical_data['onshift_to_outstanding'] = historical_data['total_onshift_dashers'] / historical_data['total_outstanding_orders']\n",
    "        historical_data['onshift_to_outstanding'].replace([np.inf,-np.inf],0,inplace=True)\n",
    "    except Exception as ex:\n",
    "        raise ex\n",
    "    return historical_data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "## add dummies for categories\n",
    "def addFeatures_dummies(historical_data:pd.DataFrame,dummy_column:str) -> pd.DataFrame:\n",
    "    try:\n",
    "        dumm = pd.get_dummies(historical_data[dummy_column],prefix=str(dummy_column + '_'),dtype=float)\n",
    "        # concat dummies\n",
    "        historical_data = pd.concat([historical_data,dumm],axis=1)\n",
    "        historical_data = historical_data.drop(columns=[dummy_column])\n",
    "    except Exception as ex:\n",
    "        raise ex\n",
    "    return historical_data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 64,
   "metadata": {},
   "outputs": [],
   "source": [
    "## remove confounding outliers\n",
    "def cleanFeatures_remove_outliers(historical_data:pd.DataFrame) -> pd.DataFrame:\n",
    "    try:\n",
    "        historical_data = historical_data.loc[historical_data['actual_total_delivery_duration'] > historical_data['est_time_non-prep']]\n",
    "        historical_data = historical_data[historical_data['actual_total_delivery_duration'] < 86400] # the equivalent of an entire day\n",
    "    except Exception as ex:\n",
    "        raise ex\n",
    "    return historical_data        "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 65,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>total_items</th>\n",
       "      <th>subtotal</th>\n",
       "      <th>num_distinct_items</th>\n",
       "      <th>min_item_price</th>\n",
       "      <th>max_item_price</th>\n",
       "      <th>total_onshift_dashers</th>\n",
       "      <th>total_busy_dashers</th>\n",
       "      <th>total_outstanding_orders</th>\n",
       "      <th>estimated_order_place_duration</th>\n",
       "      <th>estimated_store_to_consumer_driving_duration</th>\n",
       "      <th>...</th>\n",
       "      <th>created_hour_of_day__21</th>\n",
       "      <th>created_hour_of_day__22</th>\n",
       "      <th>created_hour_of_day__23</th>\n",
       "      <th>created_day_of_week__0</th>\n",
       "      <th>created_day_of_week__1</th>\n",
       "      <th>created_day_of_week__2</th>\n",
       "      <th>created_day_of_week__3</th>\n",
       "      <th>created_day_of_week__4</th>\n",
       "      <th>created_day_of_week__5</th>\n",
       "      <th>created_day_of_week__6</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>1.0</td>\n",
       "      <td>1900.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1400.0</td>\n",
       "      <td>1400.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>2.0</td>\n",
       "      <td>2.0</td>\n",
       "      <td>446.0</td>\n",
       "      <td>690.0</td>\n",
       "      <td>...</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>1.0</td>\n",
       "      <td>1900.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1900.0</td>\n",
       "      <td>1900.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>446.0</td>\n",
       "      <td>690.0</td>\n",
       "      <td>...</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>6.0</td>\n",
       "      <td>6900.0</td>\n",
       "      <td>5.0</td>\n",
       "      <td>600.0</td>\n",
       "      <td>1800.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>2.0</td>\n",
       "      <td>446.0</td>\n",
       "      <td>289.0</td>\n",
       "      <td>...</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>3.0</td>\n",
       "      <td>3900.0</td>\n",
       "      <td>3.0</td>\n",
       "      <td>1100.0</td>\n",
       "      <td>1600.0</td>\n",
       "      <td>6.0</td>\n",
       "      <td>6.0</td>\n",
       "      <td>9.0</td>\n",
       "      <td>446.0</td>\n",
       "      <td>650.0</td>\n",
       "      <td>...</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>1.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>3.0</td>\n",
       "      <td>5000.0</td>\n",
       "      <td>3.0</td>\n",
       "      <td>1500.0</td>\n",
       "      <td>1900.0</td>\n",
       "      <td>2.0</td>\n",
       "      <td>2.0</td>\n",
       "      <td>2.0</td>\n",
       "      <td>446.0</td>\n",
       "      <td>338.0</td>\n",
       "      <td>...</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>5 rows × 139 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "   total_items  subtotal  num_distinct_items  min_item_price  max_item_price  \\\n",
       "0          1.0    1900.0                 1.0          1400.0          1400.0   \n",
       "1          1.0    1900.0                 1.0          1900.0          1900.0   \n",
       "2          6.0    6900.0                 5.0           600.0          1800.0   \n",
       "3          3.0    3900.0                 3.0          1100.0          1600.0   \n",
       "4          3.0    5000.0                 3.0          1500.0          1900.0   \n",
       "\n",
       "   total_onshift_dashers  total_busy_dashers  total_outstanding_orders  \\\n",
       "0                    1.0                 2.0                       2.0   \n",
       "1                    1.0                 0.0                       0.0   \n",
       "2                    1.0                 1.0                       2.0   \n",
       "3                    6.0                 6.0                       9.0   \n",
       "4                    2.0                 2.0                       2.0   \n",
       "\n",
       "   estimated_order_place_duration  \\\n",
       "0                           446.0   \n",
       "1                           446.0   \n",
       "2                           446.0   \n",
       "3                           446.0   \n",
       "4                           446.0   \n",
       "\n",
       "   estimated_store_to_consumer_driving_duration  ...  created_hour_of_day__21  \\\n",
       "0                                         690.0  ...                      1.0   \n",
       "1                                         690.0  ...                      0.0   \n",
       "2                                         289.0  ...                      1.0   \n",
       "3                                         650.0  ...                      0.0   \n",
       "4                                         338.0  ...                      0.0   \n",
       "\n",
       "   created_hour_of_day__22  created_hour_of_day__23  created_day_of_week__0  \\\n",
       "0                      0.0                      0.0                     0.0   \n",
       "1                      0.0                      0.0                     0.0   \n",
       "2                      0.0                      0.0                     0.0   \n",
       "3                      0.0                      0.0                     0.0   \n",
       "4                      0.0                      0.0                     0.0   \n",
       "\n",
       "   created_day_of_week__1  created_day_of_week__2  created_day_of_week__3  \\\n",
       "0                     1.0                     0.0                     0.0   \n",
       "1                     0.0                     0.0                     1.0   \n",
       "2                     1.0                     0.0                     0.0   \n",
       "3                     0.0                     0.0                     0.0   \n",
       "4                     0.0                     1.0                     0.0   \n",
       "\n",
       "   created_day_of_week__4  created_day_of_week__5  created_day_of_week__6  \n",
       "0                     0.0                     0.0                     0.0  \n",
       "1                     0.0                     0.0                     0.0  \n",
       "2                     0.0                     0.0                     0.0  \n",
       "3                     0.0                     0.0                     1.0  \n",
       "4                     0.0                     0.0                     0.0  \n",
       "\n",
       "[5 rows x 139 columns]"
      ]
     },
     "execution_count": 65,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "historical_data = pd.read_csv('historical_data.csv')\n",
    "\n",
    "\n",
    "store_id_unique = historical_data[\"store_id\"].unique().tolist()\n",
    "store_id_and_category = {store_id: historical_data[historical_data.store_id == store_id].store_primary_category.mode() \n",
    "                         for store_id in store_id_unique}\n",
    "def fill(store_id):\n",
    "    try:\n",
    "        return store_id_and_category[store_id].values[0]\n",
    "    except:\n",
    "        return np.nan\n",
    "# fill null values\n",
    "historical_data[\"clean_store_primary_category\"] = historical_data.store_id.apply(fill)\n",
    "\n",
    "## build store, category prep time statistics DataFrames\n",
    "store_prep_stats_df = addFeatures_store_prep_stats(historical_data=historical_data)\n",
    "category_prep_stats_df = addFeatures_category_prep_stats(historical_data=historical_data)\n",
    "## build the training dataset using the functions above\n",
    "train_df = addFeatures_durations(historical_data=historical_data)\n",
    "train_df = addFeatures_ratios(historical_data=train_df)\n",
    "train_df = addFeatures_time_of_day(historical_data=train_df)\n",
    "## merge store, category prep stats DataFrames with train_df\n",
    "train_df = pd.merge(\n",
    "    left=train_df,\n",
    "    right=store_prep_stats_df.reset_index(),\n",
    "    how='inner',\n",
    "    on='store_id'\n",
    ")\n",
    "train_df = pd.merge(\n",
    "    left=train_df,\n",
    "    right=category_prep_stats_df.reset_index(),\n",
    "    how='inner',\n",
    "    on='clean_store_primary_category'\n",
    ")\n",
    "## remove confounding outliers\n",
    "train_df = cleanFeatures_remove_outliers(historical_data=train_df)\n",
    "## add total prep time estimations\n",
    "train_df['store_est_median_total_prep_time'] = train_df['total_items'] * train_df['store_est_time_prep_per_item_median']\n",
    "train_df['store_est_mean_total_prep_time'] = train_df['total_items'] * train_df['store_est_time_prep_per_item_mean']\n",
    "## add dummy columns for clean_store_primary_category, market_id, order_protocol\n",
    "train_df = addFeatures_dummies(historical_data=train_df,dummy_column='clean_store_primary_category')\n",
    "train_df = addFeatures_dummies(historical_data=train_df,dummy_column='market_id')\n",
    "train_df = addFeatures_dummies(historical_data=train_df,dummy_column='order_protocol')\n",
    "train_df = addFeatures_dummies(historical_data=train_df,dummy_column='created_hour_of_day')\n",
    "train_df = addFeatures_dummies(historical_data=train_df,dummy_column='created_day_of_week')\n",
    "## prep time training dataframe\n",
    "prep_time_df = train_df.copy(deep=True)\n",
    "prep_time_df['est_order_total_prep_time'] = prep_time_df['actual_total_delivery_duration'] - prep_time_df['est_time_non-prep']\n",
    "#prep_time_df['est_order_time_prep_per_item'] = prep_time_df['est_order_total_prep_time'] / prep_time_df['total_items']\n",
    "## drop the native 'store_primary_category' column and all remaining non-feature columns, then do a headcheck\n",
    "train_df.drop(columns=['store_primary_category','created_at','actual_delivery_time','store_id','est_time_prep_per_item'],inplace=True)\n",
    "prep_time_df.drop(columns=['store_primary_category','created_at','actual_delivery_time','store_id','est_time_prep_per_item','actual_total_delivery_duration'],inplace=True)\n",
    "## drop NaN's\n",
    "train_df.dropna(inplace=True)\n",
    "train_df = train_df.astype(\"float32\")\n",
    "prep_time_df.dropna(inplace=True)\n",
    "prep_time_df = prep_time_df.astype('float32')\n",
    "\n",
    "train_df.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>total_items</th>\n",
       "      <th>subtotal</th>\n",
       "      <th>num_distinct_items</th>\n",
       "      <th>min_item_price</th>\n",
       "      <th>max_item_price</th>\n",
       "      <th>total_onshift_dashers</th>\n",
       "      <th>total_busy_dashers</th>\n",
       "      <th>total_outstanding_orders</th>\n",
       "      <th>estimated_order_place_duration</th>\n",
       "      <th>estimated_store_to_consumer_driving_duration</th>\n",
       "      <th>...</th>\n",
       "      <th>created_hour_of_day__22</th>\n",
       "      <th>created_hour_of_day__23</th>\n",
       "      <th>created_day_of_week__0</th>\n",
       "      <th>created_day_of_week__1</th>\n",
       "      <th>created_day_of_week__2</th>\n",
       "      <th>created_day_of_week__3</th>\n",
       "      <th>created_day_of_week__4</th>\n",
       "      <th>created_day_of_week__5</th>\n",
       "      <th>created_day_of_week__6</th>\n",
       "      <th>est_order_total_prep_time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>1.0</td>\n",
       "      <td>1900.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1400.0</td>\n",
       "      <td>1400.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>2.0</td>\n",
       "      <td>2.0</td>\n",
       "      <td>446.0</td>\n",
       "      <td>690.0</td>\n",
       "      <td>...</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>2888.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>1.0</td>\n",
       "      <td>1900.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1900.0</td>\n",
       "      <td>1900.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>446.0</td>\n",
       "      <td>690.0</td>\n",
       "      <td>...</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>645.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>6.0</td>\n",
       "      <td>6900.0</td>\n",
       "      <td>5.0</td>\n",
       "      <td>600.0</td>\n",
       "      <td>1800.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>2.0</td>\n",
       "      <td>446.0</td>\n",
       "      <td>289.0</td>\n",
       "      <td>...</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>2340.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>3.0</td>\n",
       "      <td>3900.0</td>\n",
       "      <td>3.0</td>\n",
       "      <td>1100.0</td>\n",
       "      <td>1600.0</td>\n",
       "      <td>6.0</td>\n",
       "      <td>6.0</td>\n",
       "      <td>9.0</td>\n",
       "      <td>446.0</td>\n",
       "      <td>650.0</td>\n",
       "      <td>...</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>1294.0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>3.0</td>\n",
       "      <td>5000.0</td>\n",
       "      <td>3.0</td>\n",
       "      <td>1500.0</td>\n",
       "      <td>1900.0</td>\n",
       "      <td>2.0</td>\n",
       "      <td>2.0</td>\n",
       "      <td>2.0</td>\n",
       "      <td>446.0</td>\n",
       "      <td>338.0</td>\n",
       "      <td>...</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>1.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>1516.0</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>5 rows × 139 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "   total_items  subtotal  num_distinct_items  min_item_price  max_item_price  \\\n",
       "0          1.0    1900.0                 1.0          1400.0          1400.0   \n",
       "1          1.0    1900.0                 1.0          1900.0          1900.0   \n",
       "2          6.0    6900.0                 5.0           600.0          1800.0   \n",
       "3          3.0    3900.0                 3.0          1100.0          1600.0   \n",
       "4          3.0    5000.0                 3.0          1500.0          1900.0   \n",
       "\n",
       "   total_onshift_dashers  total_busy_dashers  total_outstanding_orders  \\\n",
       "0                    1.0                 2.0                       2.0   \n",
       "1                    1.0                 0.0                       0.0   \n",
       "2                    1.0                 1.0                       2.0   \n",
       "3                    6.0                 6.0                       9.0   \n",
       "4                    2.0                 2.0                       2.0   \n",
       "\n",
       "   estimated_order_place_duration  \\\n",
       "0                           446.0   \n",
       "1                           446.0   \n",
       "2                           446.0   \n",
       "3                           446.0   \n",
       "4                           446.0   \n",
       "\n",
       "   estimated_store_to_consumer_driving_duration  ...  created_hour_of_day__22  \\\n",
       "0                                         690.0  ...                      0.0   \n",
       "1                                         690.0  ...                      0.0   \n",
       "2                                         289.0  ...                      0.0   \n",
       "3                                         650.0  ...                      0.0   \n",
       "4                                         338.0  ...                      0.0   \n",
       "\n",
       "   created_hour_of_day__23  created_day_of_week__0  created_day_of_week__1  \\\n",
       "0                      0.0                     0.0                     1.0   \n",
       "1                      0.0                     0.0                     0.0   \n",
       "2                      0.0                     0.0                     1.0   \n",
       "3                      0.0                     0.0                     0.0   \n",
       "4                      0.0                     0.0                     0.0   \n",
       "\n",
       "   created_day_of_week__2  created_day_of_week__3  created_day_of_week__4  \\\n",
       "0                     0.0                     0.0                     0.0   \n",
       "1                     0.0                     1.0                     0.0   \n",
       "2                     0.0                     0.0                     0.0   \n",
       "3                     0.0                     0.0                     0.0   \n",
       "4                     1.0                     0.0                     0.0   \n",
       "\n",
       "   created_day_of_week__5  created_day_of_week__6  est_order_total_prep_time  \n",
       "0                     0.0                     0.0                     2888.0  \n",
       "1                     0.0                     0.0                      645.0  \n",
       "2                     0.0                     0.0                     2340.0  \n",
       "3                     0.0                     1.0                     1294.0  \n",
       "4                     0.0                     0.0                     1516.0  \n",
       "\n",
       "[5 rows x 139 columns]"
      ]
     },
     "execution_count": 11,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "prep_time_df.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "## dimensionality reduction functions with variance inflaction\n",
    "\n",
    "def calcVIF(feature_frame:pd.DataFrame,target_feature:str):\n",
    "    try:\n",
    "        feature_frame = add_constant(feature_frame)\n",
    "        feature_list = feature_frame.drop(columns=[target_feature]).columns.to_list()\n",
    "        vif_frame = pd.DataFrame()\n",
    "        vif_frame['feature'] = feature_list\n",
    "        vif_frame['VIF'] = [variance_inflation_factor(feature_frame[feature_list].values,i) for i in range(len(feature_list))]\n",
    "    except Exception as ex:\n",
    "        raise ex\n",
    "    return vif_frame.sort_values(by=['VIF'],ascending=False).reset_index(drop=True)\n",
    "\n",
    "def VIF_dim_reduction(feature_frame:pd.DataFrame,target_feature:str,vif_limit:int):\n",
    "    vif_table = calcVIF(feature_frame=feature_frame,target_feature=target_feature)\n",
    "    try:\n",
    "        ## drop nan columns from feature set, nan VIF values from base VIF table\n",
    "        vif_table['VIF'].replace(to_replace=[np.inf,-np.inf],value=np.nan,inplace=True)\n",
    "        nan_list = vif_table[vif_table['VIF'].isnull()]['feature'].to_list()\n",
    "        feature_frame = feature_frame.drop(columns=nan_list)\n",
    "        vif_table = vif_table.set_index(vif_table.feature)\n",
    "        vif_table = vif_table.drop(labels=nan_list).reset_index(drop=True)\n",
    "        ## start\n",
    "        multicollinearity = True if vif_table['VIF'].values.tolist()[1] > vif_limit else False\n",
    "        while multicollinearity:\n",
    "            lead_vif_feature = vif_table['feature'].values.tolist()[1]\n",
    "            feature_frame = feature_frame.drop(columns=lead_vif_feature)\n",
    "            vif_table = calcVIF(feature_frame=feature_frame,target_feature=target_feature)\n",
    "            multicollinearity = True if vif_table['VIF'].values.tolist()[1] > vif_limit else False\n",
    "    except Exception as ex:\n",
    "        raise ex\n",
    "    return feature_frame,vif_table"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "## feature selection/dimensionality reduction with PCA\n",
    "##TODO #2 make a class for PCA\n",
    "##TODO #3 retain row numbers in scaler --DONE\n",
    "\n",
    "def build_train_test(dataset:pd.DataFrame,target_feature:str,test_size:int):\n",
    "    try:\n",
    "        X = dataset.drop(columns=target_feature)\n",
    "        y = dataset[target_feature]\n",
    "        x_train,x_test,y_train,y_test = train_test_split(X,y,test_size=test_size,random_state=42)\n",
    "        x_train,x_test = pd.DataFrame(x_train),pd.DataFrame(x_test)\n",
    "    except Exception as ex:\n",
    "        raise ex\n",
    "    return x_train,x_test,y_train,y_test\n",
    "\n",
    "def apply_standard_scaler(dataset):\n",
    "#applies the standard scaler to a single dataset\n",
    "    scaler = StandardScaler()\n",
    "    try:\n",
    "        if type(dataset) == pd.DataFrame:\n",
    "            scaler = scaler.fit(dataset.values)\n",
    "            dataset_scaled = scaler.transform(dataset.values)\n",
    "            dataset_scaled = pd.DataFrame(dataset_scaled,index=dataset.index,columns=dataset.columns)\n",
    "        else:\n",
    "            scaler = scaler.fit(dataset)\n",
    "            dataset_scaled = scaler.transform(dataset)\n",
    "            test_stdScale = scaler.transform(dataset)\n",
    "    except Exception as ex:\n",
    "        raise ex\n",
    "    return dataset_scaled,scaler\n",
    "\n",
    "def transform_rmse_to_seconds(scaler,y_test,y_hat_scaled):\n",
    "    try:\n",
    "        y_hat_descaled = scaler.inverse_transform(y_hat_scaled.reshape(-1,1))\n",
    "        rmse_descaled = mean_squared_error(y_pred=y_test,y_true=y_hat_descaled[:,0],squared=False)\n",
    "    except Exception as ex:\n",
    "        raise ex\n",
    "    return y_hat_descaled,rmse_descaled"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "def PCAcalc_expl_var(x_train,target_expl_variance:int):\n",
    "## returns the number of components required to explain the target_expl_variance\n",
    "# and a feature to component ratio to support automation of further dimensionality reduction\n",
    "    pca = PCA()\n",
    "    x_train_scaled = apply_standard_scaler(x_train)[0]\n",
    "    try:\n",
    "        pca = pca.fit(x_train_scaled)\n",
    "        components = np.cumsum(pca.explained_variance_ratio_)\n",
    "        components = [i for i in components if i <= target_expl_variance]\n",
    "        feature_to_component_ratio = len(components)/x_train.shape[1]\n",
    "    except Exception as ex:\n",
    "        raise ex\n",
    "    return components,feature_to_component_ratio\n",
    "\n",
    "def PCAcalc_dim_reduction(x_train,x_test,desired_components:int):\n",
    "## you know, dimensionality reduction\n",
    "    pca = PCA(n_components=desired_components)\n",
    "    x_train_scaled,x_test_scaled = apply_standard_scaler(x_train)[0],apply_standard_scaler(x_test)[0]\n",
    "    try:\n",
    "        pca = pca.fit(x_train_scaled)\n",
    "        x_train_pca,x_test_pca = pca.transform(x_train_scaled),pca.transform(x_test_scaled)\n",
    "    except Exception as ex:\n",
    "        raise ex\n",
    "    return x_train_pca,x_test_pca"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Dimensionality reduction with VIF"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "VIF_features = VIF_dim_reduction(feature_frame=train_df,target_feature='actual_total_delivery_duration',vif_limit=20)[1]\n",
    "VIF_featureset = VIF_features.drop(VIF_features[VIF_features.feature == 'const'].index)\n",
    "VIF_featureset = VIF_featureset['feature'].to_list()\n",
    "VIF_featureset = VIF_featureset + ['actual_total_delivery_duration']\n",
    "VIF_train_df = train_df[VIF_featureset]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [],
   "source": [
    "vif_x_train,vif_x_test,vif_y_train,vif_y_test = build_train_test(dataset=VIF_train_df,target_feature='actual_total_delivery_duration',test_size=0.2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "We need 15 features (out of 25 available) to explain 90 percent of the variance after VIF dimensionality reduction\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAkAAAAGwCAYAAABB4NqyAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjUuMywgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy/NK7nSAAAACXBIWXMAAA9hAAAPYQGoP6dpAABT1UlEQVR4nO3deVhUZQMF8DPDvu87gijuKAi44L4buVupWa5lWZmZpGm5Z6ktll8uaS5lmZmlZkoq5S5qiiioiIogiyACwgCyztzvD3SSQJ2BGS/DnN/z8OTcmblzvF2d413eVyIIggAiIiIiPSIVOwARERHR08YCRERERHqHBYiIiIj0DgsQERER6R0WICIiItI7LEBERESkd1iAiIiISO8Yih3gaVMoFLh16xasrKwgkUjEjkNEREQqEAQB+fn5cHd3h1Ra++M3eleAbt26hQYNGogdg4iIiGogJSUFnp6etV6P3hUgKysrABUb0NraWuQ0REREpAqZTIYGDRoov8drS+8K0IPTXtbW1ixAREREOkZTl6/wImgiIiLSOyxAREREpHdYgIiIiEjvsAARERGR3mEBIiIiIr3DAkRERER6hwWIiIiI9A4LEBEREekdFiAiIiLSOyxAREREpHdELUBHjx7FoEGD4O7uDolEgl27dj3xPUeOHEFQUBBMTU3RqFEjfPPNN9oPSkRERPWKqAWosLAQ/v7+WLlypUqvT0xMxLPPPouuXbsiOjoaH3zwAaZOnYrffvtNy0mJiIioPhF1MtTQ0FCEhoaq/PpvvvkGXl5e+OqrrwAALVq0wNmzZ/H555/jueee01JKIiLSRwqFgPzicmQXlsDd1gymRgZiRyIN0qnZ4E+ePIl+/fpVWta/f39s2LABZWVlMDIyqvKekpISlJSUKB/LZDKt5yQiorqnXK5Azr1S3C0sQ3ZhCXIKS5FTWIrsgor/5twrRc79X2cXluLuvVLIFQIAYNdbnRHQwFbc3wBplE4VoIyMDLi4uFRa5uLigvLycmRlZcHNza3Ke5YsWYKFCxc+rYhERPQUCYKA3HtlyMwvwW1ZMTLzS5CZX4xMWcV/7+SXILugotDkFZXV6DOsTAxxr6Rcw8lJbDpVgABAIpFUeiwIQrXLH5g9ezamT5+ufCyTydCgQQPtBSQioloTBAF375UhPa8ImfkluCP7T8HJL0GmrAR38ktQKleovF6JBLAzN4a9xf0fc2PYWxrDweLfZQ4WJrCzMFL+18SQp77qI50qQK6ursjIyKi0LDMzE4aGhnBwcKj2PSYmJjAxMXka8YiISEUKhYDM/BKk5d5D6t0ipOUWIe3+f1PvFuFWbhHulcpVXp+duRGcrUzhbG3y0H8rfu3wUMGxNTeGgbT6fzCTftGpAhQSEoI//vij0rIDBw4gODi42ut/iIhIHHKFgFu5RUi5e09ZbNLuFinLTnpeEcrkwhPX42BhDGdr0/tlxgQu1g+Vm/vLnaxMeJSG1CZqASooKMD169eVjxMTE3H+/HnY29vDy8sLs2fPRlpaGjZv3gwAmDx5MlauXInp06dj0qRJOHnyJDZs2ICtW7eK9VsgItJbgiDgTn4JbmQVIvH+z407hUjKLkRy9r0nnpoykErgam0KD1szeNiZwcPWDJ52//6ad16RNolagM6ePYuePXsqHz+4VmfcuHH47rvvkJ6ejuTkZOXzPj4+CA8Px7vvvotVq1bB3d0d//vf/3gLPBGRFuXdK0NidiESswqQeKdQWXiSsgpR+JjTVMaGUng+VG4eLjoedmZwtTaFoQEnJCBxSIQHVxHrCZlMBhsbG+Tl5cHa2lrsOEREdYZCIeBGViGik+8iOiUX8Rn5SMwqRE5h6SPfI5UADezN0dDBAj6OFmjkVPFfH0cLuNmY8Xob0hhNf3/r1DVARESkObn3ShGdkovo5FxEJ9/F+ZRc5BdXf7u3i7XJ/WJjCR9H8/v/tYCXvTmMDXkUh3QPCxARkR4okysQn5GvPLpzPjkXN7IKq7zO1EiKNh62aOtli1YeNmh0/2iOhQm/Lqh+4R5NRFQP3ZYVV5Sd5IojPDFpuSguq3pRciNHCwR42aKtlx3aNrBFM1crGPG6HNIDLEBERDpOoRBwNTMfZ5LuIiopB2eS7iItt6jK66xNDRFwv+i09bJFQANb2Jobi5CYSHwsQEREOqaoVI4Lqbk4m5SDszfvIurm3SrX7kglQDNXa7T1sr1feOzQyNECUl6UTASABYiIqM67k1+CqJs5OJt0F2du3sWltDyUKyrfwGthbIC2XnYIbmiHYG97BHjZwpLX7RA9Ev90EBHVIYIgICn7Hk7fyMbZm3dxNikHSdn3qrzOxdoEwQ3t0c7bDsEN7dHc1Ypj6hCpgQWIiEhkeUVlOJmQhSNXs3Ds2h2k3q18/Y5EAjRzsUKQtx3aNbRHkLcdPO3MHjkJNBE9GQsQEdFTVi5X4EJqHo5du4OjV+/gfEouHj6jZWQgQVsvO7RrWHF0J9DLDjZmnO+QSJNYgIiInoLUu/dw7FoWjl69gxPXsyD7z0XLjZ0s0LWJE7o3dUKHRvYwN+Zfz0TaxD9hRERaUFhSjlM3spWl57+DDlqbGqJLE0d0a+KELk0c4WlnLlJSIv3EAkREpCEpOfew/1IG/oq7jaibd1Em//e8loFUgoAGtujWxAldmzrC39OW82QRiYgFiIiohgRBwNXbBdh3MQP7L2Xgcrqs0vOedmbo1tQJ3Zo4IaSxA6/jIapDWICIiNSgUAiITsnFgUsVpefhW9SlEqBdQ3v0b+WKns2d0dDBnHdqEdVRLEBERE9QJlfg1I1s7L+UgQOXbiMzv0T5nLGhFF19HdG/lSt6t3CGg6WJiEmJSFUsQERE1SgqlePI1Ts4cP+anofv2rI0MUTP5s7o38oFPZo5c8RlIh3EP7VERPcVlcpx4HIG9sak4+i1O5VmT3ewMEa/Vi7o18oVnRo7wMTQQMSkRFRbLEBEpNcEQUDUzbv4NSoVe2LSUVDy75EeTzsz9G/liv6tXBHkbce7tojqERYgItJLt3KLsDM6Db9GpSLxoTF6GtibYWiAB57xc0VLN2texExUT7EAEZHeKC6TY/+lDPwalYrj17Mg3B+mx9zYAM+2dsPzQZ5o39AeUh7pIar3WICIqF4ThIrb1n+NSsUfF24h/6GLmTv42OP5IE8829oNFryQmUiv8E88EdVLt2XF2HEuDb9GpSDhzr+nuDxszfBckCeeC/SAt4OFiAmJSEwsQERUb5TJFcpTXEev3lHOsG5qJMWzfhWnuDo2cuApLiJiASIi3ScIAiIu38aSP69UuqC5XUM75SkuK1NOQ0FE/2IBIiKddjEtDx/tuYzTiTkAAEdLY4xq54Xngjzh48hTXERUPRYgItJJGXnF+Gx/PHZEp0IQABNDKV7t6oM3evhyZGYieiL+LUFEOqWwpBxrj97AuqMJypGahwa4Y8YzzeFhayZyOiLSFSxARKQT5AoBv51Lxef745WTkQZ722HOwJYIaGArbjgi0jksQERU50Vez8LivXG4nC4DAHjZm2N2aHM84+fKkZqJqEZYgIiozkq4U4Al4XH4Ky4TAGBlaoipvZpgbCdvTkZKRLXCAkREdU5OYSlW/HUVW04no1whwEAqwZiO3pjauwnsLYzFjkdE9QALEBHVGSXlcmyOvIn/HbymnLKiTwtnzAptAV9nS5HTEVF9wgJERHVCbGoepv4crRzIsIWbNeYMaIHOvo4iJyOi+ogFiIhEJQgCfjh1E4v3xKFUroCzlQne698MzwV6woBTVhCRlrAAEZFoZMVlmP1bLPbGpgMA+rV0wWfP+8PGnNNWEJF2sQARkSgupuXhrZ/O4Wb2PRhKJZj9bAtM7NyQt7UT0VPBAkRET5UgCNhyOhmL9lxGabkCHrZmWDm6Ldp62YkdjYj0CAsQET01BSXl+GBHLHZfuAWg4g6vz1/wh605b20noqeLBYiInoq4dBne2nION7IKYSCV4P1nmmFS10Y85UVEomABIiKtEgQB286kYP7uSygpV8DNxhQrR7dFkLe92NGISI+xABGR1hSWlGPOrovYGZ0GAOjRzAnLRwRwNGciEh0LEBFpRXxGPt7cEoWEOxWnvML6NcXkbo0h5dg+RFQHsAARkcZtP5uCub9fRHGZAi7WJvj6xUC09+EpLyKqO1iAiEhjikrlmPv7RfwalQoA6NrEEV+ODICjpYnIyYiIKmMBIiKNuJ5ZgDe3ROHq7QJIJcC7fZrirZ6+POVFRHUSCxAR1VpkQhYm/xAFWXE5nKxMsGJUADo15iSmRFR3sQARUa38FpWKWTtiUCYXEORth29eDoKTFU95EVHdxgJERDUiCAL+9/d1fPnXVQDAgNZu+GKEP0yNDERORkT0ZCxARKS20nIFZu+IxW/nKi52nty9MWb2b8brfYhIZ7AAEZFa8orKMPmHKJy8kQ0DqQSLhrTCSx28xY5FRKQWFiAiUllKzj1M/O4MrmUWwMLYACtfCkTPZs5ixyIiUhsLEBGpJCY1FxO/O4usghK4Wpti4/h2aOluLXYsIqIaYQEioieKuHwbU7dGo6hMjuauVtg0oR3cbMzEjkVEVGMsQET0WN+dSMTCPZchCEC3pk5YNbotrEyNxI5FRFQrLEBEVC25QsDHe+Ow8UQiAODF9g2waIgfjAykIicjIqq9Gheg69evIyEhAd26dYOZmRkEQYBEwltgieqDolI53vk5Ggcu3wYAvP9Mc0zu3oh/xomo3lD7n3LZ2dno06cPmjZtimeffRbp6ekAgFdffRVhYWEaD0hET9ed/BKMWncSBy7fhrGhFF+/2BZv9GjM8kNE9YraBejdd9+FoaEhkpOTYW5urlw+cuRI7Nu3T6PhiOjpup6Zj2GrT+BCah5szY2w5dUOGOTvLnYsIiKNU/sU2IEDB7B//354enpWWt6kSRPcvHlTY8GI6Ol6eELThg7m2DShPXwcLcSORUSkFWoXoMLCwkpHfh7IysqCiQknQCTSRXtibuHdbeeVE5p+OzYY9hbGYsciItIatU+BdevWDZs3b1Y+lkgkUCgU+Oyzz9CzZ0+NhiMi7dty+ibe3hqNMrmAZ1u7YsurHVh+iKjeU/sI0GeffYYePXrg7NmzKC0txcyZM3Hp0iXk5OTgxIkT2shIRFogCAJWH07AZ/vjAQCjO3jhoyF+MOCEpkSkB9Q+AtSyZUvExMSgffv26Nu3LwoLCzF8+HBER0ejcePG2shIRBomCAI+CY9Tlp8pPX3x8VCWHyLSHxJBEASxQzxNMpkMNjY2yMvLg7U15zEi/VMuV2DWjlj8GpUKAJgzoAVe7dpI5FRERI+n6e9vtY8Abdq0Cdu3b6+yfPv27fj+++9rHYiItKe4TI43tpzDr1GpMJBK8PkL/iw/RKSX1C5AS5cuhaOjY5Xlzs7O+OSTTzQSiog0L7+4DBM2nUHE/QEO17wUiOeDPJ/8RiKiekjti6Bv3rwJHx+fKsu9vb2RnJyskVBEpFnZBSUYv+kMYtPyYGliiG/HBiOksYPYsYiIRKP2ESBnZ2fExMRUWX7hwgU4OKj/F+rq1avh4+MDU1NTBAUF4dixY499/ZYtW+Dv7w9zc3O4ublhwoQJyM7OVvtzifRFWm4RXlh7ErFpebC3MMbWSR1ZfohI76ldgEaNGoWpU6fi0KFDkMvlkMvlOHjwIN555x2MGjVKrXVt27YN06ZNw4cffojo6Gh07doVoaGhjzySdPz4cYwdOxavvPIKLl26hO3bt+PMmTN49dVX1f1tEOmF65kFeH5NJG7cKYS7jSm2Tw5Ba08bsWMREYlO7bvASktLMWbMGGzfvh2GhhVn0BQKBcaOHYtvvvkGxsaqD6DWoUMHBAYGYs2aNcplLVq0wNChQ7FkyZIqr//888+xZs0aJCQkKJd9/fXX+PTTT5GSkqLSZ/IuMNIXMam5GLfxH9y9V4bGThb44ZUOcLc1EzsWEVGNiH4XmLGxMbZt24YrV65gy5Yt2LFjBxISErBx40a1yk9paSmioqLQr1+/Ssv79euHyMjIat/TqVMnpKamIjw8HIIg4Pbt2/j1118xYMCAR35OSUkJZDJZpR+i+i4yIQsvrjuFu/fK0MbTBtsnd2L5ISJ6iNoXQT/QtGlTNG3atMYfnJWVBblcDhcXl0rLXVxckJGRUe17OnXqhC1btmDkyJEoLi5GeXk5Bg8ejK+//vqRn7NkyRIsXLiwxjmJdM3+Sxl4+6dolMoV6NTYAevGBsPSpMZ/1ImI6iW1/1aUy+X47rvv8PfffyMzMxMKhaLS8wcPHlRrfRJJ5ZFnBUGosuyBy5cvY+rUqZg3bx769++P9PR0zJgxA5MnT8aGDRuqfc/s2bMxffp05WOZTIYGDRqolZFIV/xyNgWzfouBQgD6t3LBilFtYWpkIHYsIqI6R+0C9M477+C7777DgAED4Ofn98iy8iSOjo4wMDCocrQnMzOzylGhB5YsWYLOnTtjxowZAIA2bdrAwsICXbt2xeLFi+Hm5lblPSYmJpylnvTCt0dv4OPwOADAiGBPfDKsNQwN1D7LTUSkF9QuQD///DN++eUXPPvss7X6YGNjYwQFBSEiIgLDhg1TLo+IiMCQIUOqfc+9e/eUF14/YGBQ8a9bPZvRg0hJEAR8fiAeqw5V3BzwWrdGmB3avMb/OCEi0gdqFyBjY2P4+vpq5MOnT5+OMWPGIDg4GCEhIVi3bh2Sk5MxefJkABWnr9LS0rB582YAwKBBgzBp0iSsWbNGeQps2rRpaN++Pdzd3TWSiUiXKBQCFvxxCZtP3gQAvP9Mc7zRg5MSExE9idoFKCwsDCtWrMDKlStr/S/MkSNHIjs7G4sWLUJ6ejr8/PwQHh4Ob29vAEB6enqlMYHGjx+P/Px8rFy5EmFhYbC1tUWvXr2wbNmyWuUg0kXlcgVm/hqDHdFpkEiAj4b44eWO3mLHIiLSCWqPAzRs2DAcOnQI9vb2aNWqFYyMjCo9v2PHDo0G1DSOA0T1QUm5HFO3RmP/pdswkEqwfIQ/hgR4iB2LiEhrNP39rfYRIFtb20rX7BDR03WvtByv/xCFY9eyYGwoxarRgejbsvobB4iIqHpqF6BNmzZpIwcRqSCvqAyvfHcGZ2/ehbmxAb4dG4zOvo5ixyIi0jkcHY1IR2QVlGDshn9wOV0Ga1NDbJrQHkHedmLHIiLSSTUqQL/++it++eUXJCcno7S0tNJz586d00gwIvpXel4RXlp/GjfuFMLR0hibJ3ZAS3dew0ZEVFNqj5L2v//9DxMmTICzszOio6PRvn17ODg44MaNGwgNDdVGRiK9lpRViOfXnFTO6P7L6yEsP0REtaR2AVq9ejXWrVuHlStXwtjYGDNnzkRERASmTp2KvLw8bWQk0lvxGfl4Ye1JpOUWwcfRAtvf6IRGTpZixyIi0nlqF6Dk5GR06tQJAGBmZob8/HwAwJgxY7B161bNpiPSYxdScjFy3UncyS9Bc1cr/PJ6CDw4ozsRkUaoXYBcXV2RnZ0NAPD29sapU6cAAImJiZyOgkhDTt3IxuhvTyH3XhkCGtji59c6wsmKc9oREWmK2gWoV69e+OOPPwAAr7zyCt5991307dsXI0eO5PhARBpw6Eomxm38B4WlcoQ0csCPr3aArbmx2LGIiOoVtUeCVigUUCgUyklJf/nlFxw/fhy+vr6YPHkyjI3r9l/UHAma6rI9Mbcw7efzKFcI6NPCGStHB8LUyEDsWEREotP097faBUjXsQBRXbXtTDJm74iFQgAG+7vjixH+MDJQ+yAtEVG9JMpUGDExMfDz84NUKkVMTMxjX9umTZtahyLSNxuOJ+KjPZcBAC+298LioX4wkNZusmEiIno0lQpQQEAAMjIy4OzsjICAAEgkkmoveJZIJJDL5RoPSVRfCYKA//19HV/+dRUA8Fq3Rpgd2hwSCcsPEZE2qVSAEhMT4eTkpPw1EdWeIAhY8ucVrDt6AwAwvW9TvN3Ll+WHiOgpUKkAeXt7AwDKysqwYMECzJ07F40aNdJqMKL6TKEQMPf3i9hyOhkAMHdgS7zSxUfkVERE+kOtKyyNjIywc+dObWUh0gvlcgXe234BW04nQyIBlg5vzfJDRPSUqX2LybBhw7Br1y4tRCGq/0rK5ZjyUzR2RKfBQCrBVyMDMKq9l9ixiIj0jtqzwfv6+uKjjz5CZGQkgoKCYGFhUen5qVOnaiwcUX1SVCrH6z9G4ejVOzA2kGLVS4Ho29JF7FhERHpJ7XGAfHwefaheIpHgxo0btQ6lTRwHiMSQX1yGV74/i38Sc2BmZIBvxwajSxNHsWMREekMUcYBehjvAiNST+69Uozb+A8upObBysQQmya0Q3BDe7FjERHpNbULEBGpLjO/GGPW/4P42/mwMzfCD690gJ+HjdixiIj0Xo0KUGpqKnbv3o3k5GSUlpZWem758uUaCUak69Jyi/Dy+tNIzCqEs5UJfny1A5q6WIkdi4iIUIMC9Pfff2Pw4MHw8fFBfHw8/Pz8kJSUBEEQEBgYqI2MRDonKasQL60/jbTcInjYmuGnSR3g7WDx5DcSEdFTofZt8LNnz0ZYWBguXrwIU1NT/Pbbb0hJSUH37t3xwgsvaCMjkU6Jz8jHC2tPIi23CI0cLbB9cgjLDxFRHaN2AYqLi8O4ceMAAIaGhigqKoKlpSUWLVqEZcuWaTwgkS6JSc3FyHUncSe/BM1drbDt9RC425qJHYuIiP5D7QJkYWGBkpISAIC7uzsSEhKUz2VlZWkuGZGOOZOUg9HfnkbuvTL4N7DFz691hJOVidixiIioGmpfA9SxY0ecOHECLVu2xIABAxAWFobY2Fjs2LEDHTt21EZGojrv2LU7mLT5LIrLFOjgY48N49vB0oQ3WRIR1VVq/w29fPlyFBQUAAAWLFiAgoICbNu2Db6+vvjyyy81HpCorjtwKQNTfopGqVyBHs2csOalIJgZG4gdi4iIHkPtkaB1HUeCJk0Kj03H1K3RKFcICPVzxYpRbWFsqPaZZSIiegJNf3+r/Tf1hAkT8Pfff0PPehNRFbsv3MLb98vPsLYe+PpFlh8iIl2h9t/W2dnZGDBgADw9PREWFobz589rIRZR3bbjXCqm/RwNuULA80Ge+PwFfxgasPwQEekKtf/G3r17NzIyMjB//nxERUUhKCgILVu2xCeffIKkpCQtRCSqW345k4Kw7RegEIAX2zfAp8+1gYFUInYsIiJSQ62vAUpNTcXWrVuxceNGXLt2DeXl5ZrKphW8Bohq46fTyfhgZywA4OWOXlg02A9Slh8iIq0TfTb4h5WVleHs2bM4ffo0kpKS4OLiUutARHXV5pNJmPf7JQDAhM4NMW9gS0gkLD9ERLqoRhctHDp0CJMmTYKLiwvGjRsHKysr/PHHH0hJSdF0PqI6YcPxRGX5ea1bI5YfIiIdp/YRIE9PT2RnZ6N///5Yu3YtBg0aBFNTU21kI6oT1h5JwJI/rwAA3uzRGDP6N2P5ISLScWoXoHnz5uGFF16AnZ2dNvIQ1SmrDl3HZ/vjAQBTezfBu32asPwQEdUDaheg1157TRs5iOqcFX9dw5d/XQUATO/bFFN7NxE5ERERaQonKyL6D0EQsDziKr4+eB0A8P4zzfFGj8YipyIiIk1iASJ6iCAIWLrvCtYeuQEAmDOgBV7t2kjkVEREpGksQET3CYKAxXvjsOF4IgBg/qCWmNDZR+RURESkDSxARKgoPwv/uIzvIpMAAB8N9cOYjt7ihiIiIq1RqQDt3r1b5RUOHjy4xmGIxKBQCJj7+0VsOZ0MiQRYMqw1RrX3EjsWERFpkUoFaOjQoZUeSySSSrPBP3xbsFwu10wyoqdAoRAwe0cstp1NgUQCfPa8P54P8hQ7FhERaZlKI0ErFArlz4EDBxAQEIA///wTubm5yMvLQ3h4OAIDA7Fv3z5t5yXSGEEQMOf3i9h2NgVSCfDliACWHyIiPaH2NUDTpk3DN998gy5duiiX9e/fH+bm5njttdcQFxen0YBE2iAIAj4Jj8NP9097fTkyAEMCPMSORURET4nac4ElJCTAxsamynIbGxskJSVpIhOR1q34+xq+PVZxt9ey4W1YfoiI9IzaBahdu3aYNm0a0tPTlcsyMjIQFhaG9u3bazQckTasP3YDX/11DUDFre4j2jUQORERET1tahegjRs3IjMzE97e3vD19YWvry+8vLyQnp6ODRs2aCMjkcb8dDoZi/dWnKZ9r19TjvNDRKSn1L4GyNfXFzExMYiIiMCVK1cgCAJatmyJPn36cJJIqtN+P5+GD3fFAgAmd2+Mt3r6ipyIiIjEIhEevp9dTcXFxTAxMdGp4iOTyWBjY4O8vDxYW1uLHYeekgOXMvDGlnOQKwSM6eiNRUNa6dR+S0Sk7zT9/a32KTCFQoGPPvoIHh4esLS0RGJixYWkc+fO5SkwqpOOXbuDKT9FQ64QMDzQAwsHs/wQEek7tQvQ4sWL8d133+HTTz+FsbGxcnnr1q2xfv16jYYjqq0zSTl4bXMUSuUKhPq54tPn2kAqZfkhItJ3ahegzZs3Y926dXjppZdgYGCgXN6mTRtcuXJFo+GIaiM2NQ8TN51BUZkc3Zs6YcWotjA0UHuXJyKiekjtb4O0tDT4+la9eFShUKCsrEwjoYhq6+rtfIzdeBr5JeVo72OPb14OgrEhyw8REVVQ+xuhVatWOHbsWJXl27dvR9u2bTUSiqg2bmYX4uX1p3H3Xhn8PW2wYVwwzIwNnvxGIiLSG2rfBj9//nyMGTMGaWlpUCgU2LFjB+Lj47F582bs2bNHGxmJVHYrtwijvz2NzPwSNHe1wvcT28PK1EjsWEREVMeofQRo0KBB2LZtG8LDwyGRSDBv3jzExcXhjz/+QN++fbWRkUgld/JL8PL600jLLYKPowV+eKUDbM2Nn/xGIiLSO7UaB0gXcRyg+in3XilGrTuFKxn58LA1wy+TQ+BhayZ2LCIi0hBNf3+rfQrsgdLSUmRmZkKhUFRa7uXlVetQROooKCnHuE1ncCUjH05WJtjyageWHyIieiy1C9C1a9cwceJEREZGVlouCAIkEgnkcrnGwhE9SXGZHK9+fwYXUnJha26EH1/pgIaOFmLHIiKiOk7tAjR+/HgYGhpiz549cHNz44i6JJoyuQJvbjmHUzdyYGliiM0T26OZq5XYsYiISAeoXYDOnz+PqKgoNG/eXBt5iFSiUAh4/9cYHLySCVMjKTaOb4c2nrZixyIiIh2h9l1gLVu2RFZWljayEKlsyZ9x2BGdBgOpBKtfCkR7H3uxIxERkQ5RuwAtW7YMM2fOxOHDh5GdnQ2ZTFbph0jb1h5JwLfHKibhXfZcG/Rq7iJyIiIi0jVqnwLr06cPAKB3796VlvMiaHoafo1KxZI/K+acmx3aHM8HeYqciIiIdJHaBejQoUPayEH0RAev3Mb7v8UAACZ19cHr3RuLnIiIiHSV2gWoe/fuGg2wevVqfPbZZ0hPT0erVq3w1VdfoWvXro98fUlJCRYtWoQff/wRGRkZ8PT0xIcffoiJEydqNBfVLVE3c/DmlnOQKwQMb+uB2aEtxI5EREQ6TKUCFBMTAz8/P0ilUsTExDz2tW3atFH5w7dt24Zp06Zh9erV6Ny5M9auXYvQ0FBcvnz5kQMqjhgxArdv38aGDRvg6+uLzMxMlJeXq/yZpHuu3s7HxO/OorhMgZ7NnLDs+TaQSjn8AhER1ZxKU2FIpVJkZGTA2dkZUqkUEokE1b1N3WuAOnTogMDAQKxZs0a5rEWLFhg6dCiWLFlS5fX79u3DqFGjcOPGDdjb1+yuH06FoVvScovw3OpIZMiK0dbLFlte7QBz4xoPYE5ERDpKlKkwEhMT4eTkpPy1JpSWliIqKgqzZs2qtLxfv35VRpl+YPfu3QgODsann36KH374ARYWFhg8eDA++ugjmJlVP/VBSUkJSkpKlI95p5ruyCksxdgNp5EhK4avsyU2jmvH8kNERBqh0reJt7d3tb+ujaysLMjlcri4VL6F2cXFBRkZGdW+58aNGzh+/DhMTU2xc+dOZGVl4c0330ROTg42btxY7XuWLFmChQsXaiQzPT33Sssx8bszSLhTCDcbU2ye2B52FpzZnYiINKPG/5y+fPkykpOTUVpaWmn54MGD1VrPf6fSeHA7fXUUCgUkEgm2bNkCGxsbAMDy5cvx/PPPY9WqVdUeBZo9ezamT5+ufCyTydCgQQO1MtLTVSZX4I0fz+H8/fm9fnilPdw5uSkREWmQ2gXoxo0bGDZsGGJjYytdC/SgtKh6DZCjoyMMDAyqHO3JzMysclToATc3N3h4eCjLD1BxzZAgCEhNTUWTJk2qvMfExAQmJiYqZSLxKRQCZmy/gCNX78DMyAAbx7eDrzPn9yIiIs1SeyTod955Bz4+Prh9+zbMzc1x6dIlHD16FMHBwTh8+LDK6zE2NkZQUBAiIiIqLY+IiECnTp2qfU/nzp1x69YtFBQUKJddvXoVUqkUnp4cEE/XCYKAj8PjsOv8LRhKJVj9ciACvezEjkVERPWQ2gXo5MmTWLRoEZycnCCVSiGVStGlSxcsWbIEU6dOVWtd06dPx/r167Fx40bExcXh3XffRXJyMiZPngyg4vTV2LFjla8fPXo0HBwcMGHCBFy+fBlHjx7FjBkzMHHixEdeBE2645sjN7DheMVF9p+90AY9mzmLnIiIiOortU+ByeVyWFpaAqg4jXXr1i00a9YM3t7eiI+PV2tdI0eORHZ2NhYtWoT09HT4+fkhPDxceaF1eno6kpOTla+3tLREREQE3n77bQQHB8PBwQEjRozA4sWL1f1tUB3zy9kULNtXMcXFnAEtMKwtj+gREZH2qDQO0MO6du2KsLAwDB06FKNHj8bdu3cxZ84crFu3DlFRUbh48aK2smoExwGqe/66fBuv/xgFuULA690bcZRnIiKqQpRxgB42Z84cFBYWAgAWL16MgQMHomvXrnBwcMC2bdtqHYj0y5mkHLz1U8UUF88HeWLWM83FjkRERHpA7SNA1cnJyYGdnd0jb1+vS3gEqO64ejsfz6+JhKy4HL2bO2PtmCAYGqh9WRoREekB0Y8AVaem01KQ/rqTX4IJm85AVlyOIG87rBwdyPJDRERPjUoFaPjw4SqvcMeOHTUOQ/qhuEyO1344i7TcIjR0MMf6scEwMzYQOxYREekRlQrQwwMPEtWGQiHgve0XEJ2cCxszI2wc345TXBAR0VOnUgHatGmTtnOQnvjqr6vYE5MOQ6kE37wchEZOlmJHIiIiPVTja4AyMzMRHx8PiUSCpk2bwtmZg9bR4+04l4r/HbwOAPhkeGuENHYQOREREekrta86lclkGDNmDDw8PNC9e3d069YNHh4eePnll5GXl6eNjFQP/JOYg1m/xQIA3ujRGCOCOSEtERGJR+0C9Oqrr+L06dPYs2cPcnNzkZeXhz179uDs2bOYNGmSNjKSjkvKKsTrP5xFqVyBUD9XzOjXTOxIRESk59QeB8jCwgL79+9Hly5dKi0/duwYnnnmGeUgiXUVxwF6uvLulWHYmhO4cacQbTxtsO21EN7xRUREatP097faR4AcHByqvSvMxsYGdnacuZv+VVquwOQfo3DjTiHcbUx5uzsREdUZahegOXPmYPr06UhPT1cuy8jIwIwZMzB37lyNhiPdJQgC5u66iJM3smFhbID149rB2dpU7FhEREQAanAKrG3btrh+/TpKSkrg5eUFAEhOToaJiQmaNGlS6bXnzp3TXFIN4Smwp2PtkQQs+fMKpBJg/bhg9GruInYkIiLSYaJPhTF06NBafyjVb/suZmDpvisAgLkDW7L8EBFRnaN2AZo/f742clA9EZuah2nboiEIwNgQb4zv1FDsSERERFWofQ3QX3/99cjn1q5dW6swpNvS84rwyvdnUFymQPemTpg3sCUkEonYsYiIiKpQuwANGDAAYWFhKC0tVS67c+cOBg0ahNmzZ2s0HOmOwpJyTPzuLDLzS9DMxQorR7fl7O5ERFRnqf0NdfToUfzxxx9o164dLl26hL1798LPzw8FBQW4cOGCNjJSHSdXCJi6NRpx6TI4Whpjw/hgWJkaiR2LiIjokdQuQB06dEB0dDTatGmDoKAgDBs2DGFhYTh48CAaNOD0Bvro471x+PtKJkwMpVg3NhieduZiRyIiInqsGp2jiI+Px5kzZ+Dp6QlDQ0NcuXIF9+7d03Q20gE/nLqJjScSAQBfjPBHoBcHwyQiorpP7QK0dOlShISEoG/fvrh48SLOnDmjPCJ08uRJbWSkOurI1TtYsPsSAOC9fk0xsI27yImIiIhUo3YBWrFiBXbt2oWvv/4apqamaNWqFf755x8MHz4cPXr00EJEqosSswox5adzkCsEDA/0wFs9fcWOREREpDK1xwGKjY2Fo6NjpWVGRkb47LPPMHDgQI0Fo7rrXmk5Jv8QhfzicgR522HJ8Na83Z2IiHSK2keAHB0dkZubi/Xr12P27NnIyckBUDHtha8vjwLUd4Ig4P3fYhF/Ox9OViZY/VIgTAw5wSkREekWtY8AxcTEoE+fPrCxsUFSUhImTZoEe3t77Ny5Ezdv3sTmzZu1kZPqiI0nkvDHhVswlEqw+qVAuHCCUyIi0kFqHwGaPn06xo8fj2vXrsHU9N8vv9DQUBw9elSj4ahuOXUjG5+ExwEAPhzQAu0a2ouciIiIqGbULkBnzpzB66+/XmW5h4cHMjIyNBKK6p6MvGLlRc9DAtw5xxcREek0tQuQqakpZDJZleXx8fFwcnLSSCiqW0rK5XhjSxSyCkrR3NWKFz0TEZHOU7sADRkyBIsWLUJZWRkAQCKRIDk5GbNmzcJzzz2n8YAkvo/2XEZ0ci6sTQ2xdkwQzI3VvnSMiIioTlG7AH3++ee4c+cOnJ2dUVRUhO7du8PX1xdWVlb4+OOPtZGRRPRrVCp+PJUMAPhqVAC8HSxETkRERFR7av9T3traGsePH8fBgwdx7tw5KBQKBAYGok+fPtrIRyK6mJaHD3fGAgCm9WmCXs1dRE5ERESkGTU+l9GrVy/06tVLk1moDrlbWIrXf4hCSbkCvZo7Y2qvJmJHIiIi0pgaTYZK9ZtcIWDqz9FIyy2Cl705vhwRAKmUFz0TEVH9wQJEVXwZcRXHrmXB1EiKtWOCYGNuJHYkIiIijWIBokoOXMrAykPXAQDLnmuDFm7WIiciIiLSPBYgUrpxpwBhv1wAAIzv1BBDAjxETkRERKQdNSpACQkJmDNnDl588UVkZmYCAPbt24dLly5pNBw9PYUl5Xj9hyjkl5SjfUN7fDighdiRiIiItEbtAnTkyBG0bt0ap0+fxo4dO1BQUACgYpLU+fPnazwgaZ8gCJj5awyuZRbA2coEK19qCyMDHhwkIqL6S+1vuVmzZmHx4sWIiIiAsbGxcnnPnj1x8uRJjYajp2P9sUTsjU2HoVSCNS8HwtmKM7wTEVH9pnYBio2NxbBhw6osd3JyQnZ2tkZC0dNzMiEbS/ddAQDMG9QSQd6c4Z2IiOo/tQuQra0t0tPTqyyPjo6GhwcvmtUl6XlFyhneh7f1wJiO3mJHIiIieirULkCjR4/G+++/j4yMDEgkEigUCpw4cQLvvfcexo4dq42MpAUl5XJM/vEcsgtL0cLNGh8P4wzvRESkP9QuQB9//DG8vLzg4eGBgoICtGzZEt26dUOnTp0wZ84cbWQkLfh4bxwupOTCxswIa18OgpmxgdiRiIiInhqJIAhCTd6YkJCA6OhoKBQKtG3bFk2a6MZcUTKZDDY2NsjLy4O1tX4O8vdnbDre2HIOALBpfDv0bO4sciIiIqLH0/T3t9qToR45cgTdu3dH48aN0bhx41oHoKcrOfseZv4WAwB4vXsjlh8iItJLap8C69u3L7y8vDBr1ixcvHhRG5lIS0rLFXh76znkF5cj0MsW7/VrJnYkIiIiUahdgG7duoWZM2fi2LFjaNOmDdq0aYNPP/0Uqamp2shHGvTpviu4kJoHGzMjfD06kIMdEhGR3lL7G9DR0RFTpkzBiRMnkJCQgJEjR2Lz5s1o2LAhevXqpY2MpAF/Xb6N9ccTAQCfv+APD1szkRMRERGJp1aHAHx8fDBr1iwsXboUrVu3xpEjRzSVizQoLbcIYdsrJjmd2NkHfVu6iJyIiIhIXDUuQCdOnMCbb74JNzc3jB49Gq1atcKePXs0mY00oEyuwNSt0cgrKkMbTxvMCm0udiQiIiLRqX0X2AcffICtW7fi1q1b6NOnD7766isMHToU5ubm2shHtbQ84iqibt6FlYkhVr4YCGNDXvdDRESkdgE6fPgw3nvvPYwcORKOjo7ayEQacjg+E2sOJwAAlj3fBl4OLKlERERADQpQZGSkNnKQht2WFWP6LxXX/Yzp6I1nW7uJnIiIiKjuUKkA7d69G6GhoTAyMsLu3bsf+9rBgwdrJBjVnFwhYOrWaOTcn+frwwEtxI5ERERUp6hUgIYOHYqMjAw4Oztj6NChj3ydRCKBXC7XVDaqoRV/X8PpxBxYGBtg1ei2MDXiPF9EREQPU6kAKRSKan9NdU/k9Sx8ffAaAOCT4a3RyMlS5ERERER1j9q3BG3evBklJSVVlpeWlmLz5s0aCUU1cye/BO9sOw9BAEYGN8CQAA+xIxEREdVJahegCRMmIC8vr8ry/Px8TJgwQSOhSH0KhYB3t53HnfwSNHWxxILBrcSOREREVGepXYAEQYBEIqmyPDU1FTY2NhoJRepbcyQBx69nwdRIilWjA2FmzOt+iIiIHkXl2+Dbtm0LiUQCiUSC3r17w9Dw37fK5XIkJibimWee0UpIerx/EnPwxYF4AMCiIX5o4mIlciIiIqK6TeUC9ODur/Pnz6N///6wtPz34lpjY2M0bNgQzz33nMYD0uPlFJZi6tZoKARgeFsPvBDkKXYkIiKiOk/lAjR//nwAQMOGDTFy5EiYmppqLRSpRqEQEPbLeWTIitHIyQIfDfWr9vQkERERVab2SNDjxo3TRg6qgfXHb+BQ/B0YG1Zc92Nhovb/TiIiIr2k9jemXC7Hl19+iV9++QXJyckoLS2t9HxOTo7GwtGjnUu+i0/3VVz3M39QS7RwsxY5ERERke5Q+y6whQsXYvny5RgxYgTy8vIwffp0DB8+HFKpFAsWLNBCRPqvwpJyTN0ajXKFgAFt3DC6vZfYkYiIiHSK2gVoy5Yt+Pbbb/Hee+/B0NAQL774ItavX4958+bh1KlT2shI//FlxFWk3i2Cp50Zlg5vzet+iIiI1KR2AcrIyEDr1q0BAJaWlspBEQcOHIi9e/dqNh1VcTEtDxtPJAIAPhrqBytTI5ETERER6R61C5CnpyfS09MBAL6+vjhw4AAA4MyZMzAxMVE7wOrVq+Hj4wNTU1MEBQXh2LFjKr3vxIkTMDQ0REBAgNqfqavkCgEf7IyFQgAGtnFDz2bOYkciIiLSSWoXoGHDhuHvv/8GALzzzjuYO3cumjRpgrFjx2LixIlqrWvbtm2YNm0aPvzwQ0RHR6Nr164IDQ1FcnLyY9+Xl5eHsWPHonfv3urG12k/nExCTGoerEwNMW9gS7HjEBER6SyJIAhCbVZw6tQpREZGwtfXF4MHD1brvR06dEBgYCDWrFmjXNaiRQsMHToUS5YseeT7Ro0ahSZNmsDAwAC7du3C+fPnVf5MmUwGGxsb5OXlwdpad+6cSs8rQt/lR1FQUo7FQ/3wckdvsSMRERE9NZr+/q71wDEdO3ZEx44d1X5faWkpoqKiMGvWrErL+/Xrh8jIyEe+b9OmTUhISMCPP/6IxYsXP/FzSkpKKs1eL5PJ1M5aFyzcfRkFJeVo62XLu76IiIhqSaUCtHv3bpVXqOpRoKysLMjlcri4uFRa7uLigoyMjGrfc+3aNcyaNQvHjh2rNBfZ4yxZsgQLFy5U6bV11V+Xb2PfpQwYSiVYMrw1pFLe9UVERFQbKrWIB/OAPYlEIoFcLlcrwH9v4X7UbPNyuRyjR4/GwoUL0bRpU5XXP3v2bEyfPl35WCaToUGDBmplFFNhSTnm774EAHilqw+au+rOaTsiIqK6SqUCpFAoNP7Bjo6OMDAwqHK0JzMzs8pRIQDIz8/H2bNnER0djSlTpihzCYIAQ0NDHDhwAL169aryPhMTkxrdnVZXfPXXVaTlVoz5807vJmLHISIiqhfUvgtMU4yNjREUFISIiIhKyyMiItCpU6cqr7e2tkZsbCzOnz+v/Jk8eTKaNWuG8+fPo0OHDk8r+lNTMeZPEoCKMX/MjTnXFxERkSao/Y26aNGixz4/b948ldc1ffp0jBkzBsHBwQgJCcG6deuQnJyMyZMnA6g4fZWWlobNmzdDKpXCz8+v0vudnZ1hampaZXl9IFcI+HBnLOT3p7vgmD9ERESao3YB2rlzZ6XHZWVlSExMhKGhIRo3bqxWARo5ciSys7OxaNEipKenw8/PD+Hh4fD2rrjFOz09/YljAtVXP566iQupebAyMcR8jvlDRESkUbUeBwiouLB4/PjxGDZsGMaMGaOJXFqjC+MAZeQVo8/yIygoKcdHQ/0whmP+EBGRntP097dGrgGytrbGokWLMHfuXE2sTu8t/OMSCkrKEdDAFi9xzB8iIiKN09hF0Lm5ucqJUanm/o67jT8vZsCAY/4QERFpjdrXAP3vf/+r9FgQBKSnp+OHH37AM888o7Fg+uheaTnm/V4x5s+rXXzQwq1unqIjIiLSdWoXoC+//LLSY6lUCicnJ4wbNw6zZ8/WWDB99NVf15CWWwQPWzO804dj/hAREWmL2gUoMTFRGzn03qVbedhwvGLbLuaYP0RERFol2kCI9C+5QsAHOy9WjPnT2g09m3PMHyIiIm1S+zBDcXExvv76axw6dAiZmZlVpsk4d+6cxsLpiy2nb+JCSi6sTAwxbxDH/CEiItI2tQvQxIkTERERgeeffx7t27evduJSUt1tWTE+3RcPAJj5TDO4WJuKnIiIiKj+U7sA7d27F+Hh4ejcubM28uidh8f8Gd2BAx4SERE9DWpfA+Th4QErKyttZNE7B6/cRnhsxZg/nwxrDQOO+UNERPRUqF2AvvjiC7z//vu4efOmNvLojXul5Zi7698xf1q6c8wfIiKip0XtU2DBwcEoLi5Go0aNYG5uDiMjo0rP5+TkaCxcfbaCY/4QERGJRu0C9OKLLyItLQ2ffPIJXFxceBF0DVy+JcP6+2P+LBrSimP+EBERPWVqf/NGRkbi5MmT8Pf310YevfDFgXjIFQJC/VzRu4WL2HGIiIj0jtrXADVv3hxFRUXayKIXrmcW4O8rmZBIgBn9m4kdh4iISC+pXYCWLl2KsLAwHD58GNnZ2ZDJZJV+6PE2HL8BAOjTwgWNnCxFTkNERKSf1D4F9mDG9969e1daLggCJBIJ5HK5ZpLVQ1kFJfjtXBoA4LVujUROQ0REpL/ULkCHDh3SRg69sPnkTZSWK+DfwBbB3nZixyEiItJbaheg7t27ayNHvVdUKscPJ5MAAK91bcS754iIiESkdgE6evToY5/v1q1bjcPUZ7+dS8Xde2XwtDND/1a884uIiEhMahegHj16VFn28NEMXgNUlUIhYMP9cX9e6eIDQwO1rz0nIiIiDVL7m/ju3buVfjIzM7Fv3z60a9cOBw4c0EZGnfdX3G0kZhXC2tQQI4IbiB2HiIhI76l9BMjGxqbKsr59+8LExATvvvsuoqKiNBKsPvn2WMWt7y919IaFCUd9JiIiEpvGzsU4OTkhPj5eU6urN6KT7+JM0l0YGUgwvlNDseMQERERanAEKCYmptJjQRCQnp6OpUuXcnqMaqw/VnHtz2B/D7hYm4qchoiIiIAaFKCAgABIJBIIglBpeceOHbFx40aNBasPUnLu4c+L6QCASd18RE5DRERED6hdgBITEys9lkqlcHJygqkpj27814bjiVAIQNcmjmjuai12HCIiIrpP7QLk7e2tjRz1Tt69MvxyNgUAp70gIiKqa1S+CPrgwYNo2bJltROe5uXloVWrVjh27JhGw+myLf/cxL1SOZq7WqGLr6PYcYiIiOghKhegr776CpMmTYK1ddVTOTY2Nnj99dexfPlyjYbTVaXlCnx3IgkAMInTXhAREdU5KhegCxcuKGeCr06/fv04BtB9uy/cQmZ+CVysTTDI313sOERERPQfKheg27dvw8jI6JHPGxoa4s6dOxoJpcsEQcD6+wMfju/kA2NDTntBRERU16j87ezh4YHY2NhHPh8TEwM3NzeNhNJlx65l4UpGPiyMDTC6g5fYcYiIiKgaKhegZ599FvPmzUNxcXGV54qKijB//nwMHDhQo+F00YNpL0a0awAbs0cfMSMiIiLxSIT/jmj4CLdv30ZgYCAMDAwwZcoUNGvWDBKJBHFxcVi1ahXkcjnOnTsHFxcXbWeuFZlMBhsbG+Tl5VV7QXdtXL4lw7P/OwapBDgyoyca2JtrdP1ERET6StPf3yqPA+Ti4oLIyEi88cYbmD17tnIkaIlEgv79+2P16tV1vvxo2/rjFUd/Qlu7sfwQERHVYWoNhOjt7Y3w8HDcvXsX169fhyAIaNKkCezs7LSVT2dk5BVj9/lbAIDXunLgQyIiorpM7ZGgAcDOzg7t2rXTdBad9l1kEsoVAto3tId/A1ux4xAREdFj8B5tDSgoKceW0zcBAJM47QUREVGdxwKkAb+cSUF+cTkaOVqgd3NnseMQERHRE7AA1VK5XIENxxMBAK909YFUymkviIiI6joWoFradykDablFsLcwxnOBnmLHISIiIhWwANWCIAj49mjFre9jOnrD1MhA5ERERESkChagWvgnMQcXUvNgYijFmBBvseMQERGRiliAauHbYxXX/gwP9ISjpYnIaYiIiEhVLEA1lHCnAH/F3QYAvNrVR+Q0REREpA4WoBp6cOdXnxYuaOxkKXIaIiIiUgcLUA1kF5Tgt6hUAMAkHv0hIiLSOSxANfDDqZsoKVfA39MG7X3sxY5DREREamIBUlNxmRybT1ZMe/Fq10aQSDjwIRERka5hAVLT33GZyCkshZuNKUL9XMWOQ0RERDXAAqSmHecqrv0ZHugBQwNuPiIiIl3Eb3A13MkvweGrdwAAw9py2gsiIiJdxQKkht0XbkGuEODfwBa+zrz1nYiISFexAKnhwa3vzwd6iJyEiIiIaoMFSEVx6TJcTpfByECCgW3cxY5DREREtcACpKKd0WkAgN7NXWBnYSxyGiIiIqoNFiAVlMsVygI0nKe/iIiIdB4LkAqOX8/CnfwS2JkboUczZ7HjEBERUS2xAKngt3MVR3+GBHjA2JCbjIiISNfx2/wJZMVlOHApAwBPfxEREdUXLEBP8GdsOkrKFfB1tkRrDxux4xAREZEGsAA9wW9RFae/ngv05MSnRERE9QQL0GMkZ9/DP0k5kEiAoW059g8REVF9wQL0GA9ufe/i6wg3GzOR0xAREZGmsAA9giAI2BH978zvREREVH+wAD1C1M27uJl9DxbGBujfylXsOERERKRBoheg1atXw8fHB6ampggKCsKxY8ce+dodO3agb9++cHJygrW1NUJCQrB//36t5PrtXMXRn9DWbjA3NtTKZxAREZE4RC1A27Ztw7Rp0/Dhhx8iOjoaXbt2RWhoKJKTk6t9/dGjR9G3b1+Eh4cjKioKPXv2xKBBgxAdHa3RXMVlcuyJSQfA019ERET1kUQQBEGsD+/QoQMCAwOxZs0a5bIWLVpg6NChWLJkiUrraNWqFUaOHIl58+ap9HqZTAYbGxvk5eXB2tq62tfsibmFKT9Fw8PWDMdm9oRUytvfiYiIxKTK97c6RDsCVFpaiqioKPTr16/S8n79+iEyMlKldSgUCuTn58Pe3v6RrykpKYFMJqv08yS/RVWc/hrW1oPlh4iIqB4SrQBlZWVBLpfDxcWl0nIXFxdkZGSotI4vvvgChYWFGDFixCNfs2TJEtjY2Ch/GjRo8Nh1ZuYX4+i1LADAMJ7+IiIiqpdEvwj6v6MrC4Kg0ojLW7duxYIFC7Bt2zY4Oz96hvbZs2cjLy9P+ZOSkvLY9e4+fwtyhYC2XrZo7GSp2m+CiIiIdIpotzc5OjrCwMCgytGezMzMKkeF/mvbtm145ZVXsH37dvTp0+exrzUxMYGJiYnKuR7M/D480FPl9xAREZFuEe0IkLGxMYKCghAREVFpeUREBDp16vTI923duhXjx4/HTz/9hAEDBmg00+VbMsSly2BsIMWgNm4aXTcRERHVHaIOcDN9+nSMGTMGwcHBCAkJwbp165CcnIzJkycDqDh9lZaWhs2bNwOoKD9jx47FihUr0LFjR+XRIzMzM9jY1H6m9h33x/7p3cIZtubGtV4fERER1U2iFqCRI0ciOzsbixYtQnp6Ovz8/BAeHg5vb28AQHp6eqUxgdauXYvy8nK89dZbeOutt5TLx40bh++++65WWcrlCuw6fwsAT38RERHVd6KOAySGR40jcCg+ExM2nYG9hTFOf9AbRgaiXx9ORERE99WbcYDqmgdj/wz2d2f5ISIiquf4TQ8gr6gMBy7fBgA8x9NfRERE9R4LEIDw2HSUlivQ1MUSfh61P6xGREREdRsLEP69+2t4oKdKgzASERGRbtP7AnQzuxBnku5CKgGGBnDqCyIiIn2g9wVox/2Rnzv7OsLVxlTkNERERPQ06HUBEgQBO6IrTn/x4mciIiL9odcF6OzNu0jJKYKFsQH6t3IVOw4RERE9JXpdgB6M/fNsazeYGRuInIaIiIieFr0tQMVlcuyNSQfAqS+IiIj0jd4WoINXMpFfUg4PWzN08LEXOw4RERE9RXpbgHZfqLj7a3igB6RSjv1DRESkT/S2AEVezwbA019ERET6SG8LkEIAAr1s4eNoIXYUIiIiesr0tgABPPpDRESkr/S2ABkZSDGojbvYMYiIiEgEeluAejZzhI25kdgxiIiISAR6W4AG+XPiUyIiIn2ltwWoSxNHsSMQERGRSPS2ABkZ6O1vnYiISO+xBRAREZHeYQEiIiIivcMCRERERHqHBYiIiIj0DgsQERER6R0WICIiItI7LEBERESkd1iAiIiISO+wABEREZHeYQEiIiIivcMCRERERHqHBYiIiIj0DgsQERER6R1DsQM8bYIgAABkMpnISYiIiEhVD763H3yP15beFaDs7GwAQIMGDUROQkREROrKzs6GjY1NrdejdwXI3t4eAJCcnKyRDajPZDIZGjRogJSUFFhbW4sdR6dxW2oGt6PmcFtqDrelZuTl5cHLy0v5PV5beleApNKKy55sbGy4I2qItbU1t6WGcFtqBrej5nBbag63pWY8+B6v9Xo0shYiIiIiHcICRERERHpH7wqQiYkJ5s+fDxMTE7Gj6DxuS83httQMbkfN4bbUHG5LzdD0dpQImrqfjIiIiEhH6N0RICIiIiIWICIiItI7LEBERESkd1iAiIiISO/oXQFavXo1fHx8YGpqiqCgIBw7dkzsSDpnwYIFkEgklX5cXV3FjlXnHT16FIMGDYK7uzskEgl27dpV6XlBELBgwQK4u7vDzMwMPXr0wKVLl8QJW8c9aVuOHz++yj7asWNHccLWYUuWLEG7du1gZWUFZ2dnDB06FPHx8ZVew/1SNapsS+6XqlmzZg3atGmjHDgyJCQEf/75p/J5Te2TelWAtm3bhmnTpuHDDz9EdHQ0unbtitDQUCQnJ4sdTee0atUK6enpyp/Y2FixI9V5hYWF8Pf3x8qVK6t9/tNPP8Xy5cuxcuVKnDlzBq6urujbty/y8/OfctK670nbEgCeeeaZSvtoeHj4U0yoG44cOYK33noLp06dQkREBMrLy9GvXz8UFhYqX8P9UjWqbEuA+6UqPD09sXTpUpw9exZnz55Fr169MGTIEGXJ0dg+KeiR9u3bC5MnT660rHnz5sKsWbNESqSb5s+fL/j7+4sdQ6cBEHbu3Kl8rFAoBFdXV2Hp0qXKZcXFxYKNjY3wzTffiJBQd/x3WwqCIIwbN04YMmSIKHl0WWZmpgBAOHLkiCAI3C9r47/bUhC4X9aGnZ2dsH79eo3uk3pzBKi0tBRRUVHo169fpeX9+vVDZGSkSKl017Vr1+Du7g4fHx+MGjUKN27cEDuSTktMTERGRkal/dPExATdu3fn/llDhw8fhrOzM5o2bYpJkyYhMzNT7Eh1Xl5eHoB/J43mfllz/92WD3C/VI9cLsfPP/+MwsJChISEaHSf1JsClJWVBblcDhcXl0rLXVxckJGRIVIq3dShQwds3rwZ+/fvx7fffouMjAx06tQJ2dnZYkfTWQ/2Qe6fmhEaGootW7bg4MGD+OKLL3DmzBn06tULJSUlYkerswRBwPTp09GlSxf4+fkB4H5ZU9VtS4D7pTpiY2NhaWkJExMTTJ48GTt37kTLli01uk/q3WzwEomk0mNBEKoso8cLDQ1V/rp169YICQlB48aN8f3332P69OkiJtN93D81Y+TIkcpf+/n5ITg4GN7e3ti7dy+GDx8uYrK6a8qUKYiJicHx48erPMf9Uj2P2pbcL1XXrFkznD9/Hrm5ufjtt98wbtw4HDlyRPm8JvZJvTkC5OjoCAMDgyoNMTMzs0qTJPVYWFigdevWuHbtmthRdNaDu+i4f2qHm5sbvL29uY8+wttvv43du3fj0KFD8PT0VC7nfqm+R23L6nC/fDRjY2P4+voiODgYS5Ysgb+/P1asWKHRfVJvCpCxsTGCgoIQERFRaXlERAQ6deokUqr6oaSkBHFxcXBzcxM7is7y8fGBq6trpf2ztLQUR44c4f6pAdnZ2UhJSeE++h+CIGDKlCnYsWMHDh48CB8fn0rPc79U3ZO2ZXW4X6pOEASUlJRodp/U0AXaOuHnn38WjIyMhA0bNgiXL18Wpk2bJlhYWAhJSUliR9MpYWFhwuHDh4UbN24Ip06dEgYOHChYWVlxOz5Bfn6+EB0dLURHRwsAhOXLlwvR0dHCzZs3BUEQhKVLlwo2NjbCjh07hNjYWOHFF18U3NzcBJlMJnLyuudx2zI/P18ICwsTIiMjhcTEROHQoUNCSEiI4OHhwW35H2+88YZgY2MjHD58WEhPT1f+3Lt3T/ka7peqedK25H6putmzZwtHjx4VEhMThZiYGOGDDz4QpFKpcODAAUEQNLdP6lUBEgRBWLVqleDt7S0YGxsLgYGBlW5RJNWMHDlScHNzE4yMjAR3d3dh+PDhwqVLl8SOVecdOnRIAFDlZ9y4cYIgVNxyPH/+fMHV1VUwMTERunXrJsTGxoobuo563La8d++e0K9fP8HJyUkwMjISvLy8hHHjxgnJyclix65zqtuGAIRNmzYpX8P9UjVP2pbcL1U3ceJE5fe0k5OT0Lt3b2X5EQTN7ZMSQRCEGh6RIiIiItJJenMNEBEREdEDLEBERESkd1iAiIiISO+wABEREZHeYQEiIiIivcMCRERERHqHBYiIiIj0DgsQERER6R0WICKqJCkpCRKJBOfPnxc7itKVK1fQsWNHmJqaIiAgQOw4RFQPsAAR1THjx4+HRCLB0qVLKy3ftWsXJBKJSKnENX/+fFhYWCA+Ph5///232HF0Vo8ePTBt2jSxYxDVCSxARHWQqakpli1bhrt374odRWNKS0tr/N6EhAR06dIF3t7ecHBw0GAqItJXLEBEdVCfPn3g6uqKJUuWPPI1CxYsqHI66KuvvkLDhg2Vj8ePH4+hQ4fik08+gYuLC2xtbbFw4UKUl5djxowZsLe3h6enJzZu3Fhl/VeuXEGnTp1gamqKVq1a4fDhw5Wev3z5Mp599llYWlrCxcUFY8aMQVZWlvL5Hj16YMqUKZg+fTocHR3Rt2/fan8fCoUCixYtgqenJ0xMTBAQEIB9+/Ypn5dIJIiKisKiRYsgkUiwYMGCR65n2bJl8PX1hYmJCby8vPDxxx8rn4+NjUWvXr1gZmYGBwcHvPbaaygoKKjVtnpwuvDnn39+7LY6cuQI2rdvDxMTE7i5uWHWrFkoLy+vtK2mTp2KmTNnwt7eHq6urlV+n3l5eXjttdfg7OwMa2tr9OrVCxcuXFA+/2B/+OGHH9CwYUPY2Nhg1KhRyM/PV/7+jhw5ghUrVkAikUAikSApKQl3797FSy+9BCcnJ5iZmaFJkybYtGlTtduYqD5hASKqgwwMDPDJJ5/g66+/Rmpqaq3WdfDgQdy6dQtHjx7F8uXLsWDBAgwcOBB2dnY4ffo0Jk+ejMmTJyMlJaXS+2bMmIGwsDBER0ejU6dOGDx4MLKzswEA6enp6N69OwICAnD27Fns27cPt2/fxogRIyqt4/vvv4ehoSFOnDiBtWvXVptvxYoV+OKLL/D5558jJiYG/fv3x+DBg3Ht2jXlZ7Vq1QphYWFIT0/He++9V+16Zs+ejWXLlmHu3Lm4fPkyfvrpJ7i4uAAA7t27h2eeeQZ2dnY4c+YMtm/fjr/++gtTpkzR+rZKS0vDs88+i3bt2uHChQtYs2YNNmzYgMWLF1fZVhYWFjh9+jQ+/fRTLFq0CBEREQAAQRAwYMAAZGRkIDw8HFFRUQgMDETv3r2Rk5OjXEdCQgJ27dqFPXv2YM+ePThy5IjyVOqKFSsQEhKCSZMmIT09Henp6WjQoIFye/3555+Ii4vDmjVr4OjoWO02JqpXNDZ/PRFpxLhx44QhQ4YIgiAIHTt2FCZOnCgIgiDs3LlTePiP7Pz58wV/f/9K7/3yyy8Fb2/vSuvy9vYW5HK5clmzZs2Erl27Kh+Xl5cLFhYWwtatWwVBEITExEQBgLB06VLla8rKygRPT09h2bJlgiAIwty5c4V+/fpV+uyUlBQBgBAfHy8IgiB0795dCAgIeOLv193dXfj4448rLWvXrp3w5ptvKh/7+/sL8+fPf+Q6ZDKZYGJiInz77bfVPr9u3TrBzs5OKCgoUC7bu3evIJVKhYyMDEEQtLetPvjgA6FZs2aCQqFQvmbVqlWCpaWl8rO6d+8udOnSpco2eP/99wVBEIS///5bsLa2FoqLiyu9pnHjxsLatWsFQajYH8zNzQWZTKZ8fsaMGUKHDh2Uj7t37y688847ldYxaNAgYcKECdVuN6L6jEeAiOqwZcuW4fvvv8fly5drvI5WrVpBKv33j7qLiwtat26tfGxgYAAHBwdkZmZWel9ISIjy14aGhggODkZcXBwAICoqCocOHYKlpaXyp3nz5gAqjkI8EBwc/NhsMpkMt27dQufOnSst79y5s/KzVBEXF4eSkhL07t37kc/7+/vDwsKi0mcoFArEx8crl2ljW8XFxSEkJKTSBeydO3dGQUFBpaN7bdq0qbRONzc35edERUWhoKAADg4OlbZ5YmJipe3dsGFDWFlZVbuOR3njjTfw888/IyAgADNnzkRkZORjX09UXxiKHYCIHq1bt27o378/PvjgA4wfP77Sc1KpFIIgVFpWVlZWZR1GRkaVHkskkmqXKRSKJ+Z58CWuUCgwaNAgLFu2rMpr3NzclL9+uHCost4HBEFQ6443MzOzxz7/uPU9vFwb26q6z37w/+1Jn/3gcxQKBdzc3KpcWwQAtra2Kq3jUUJDQ3Hz5k3s3bsXf/31F3r37o233noLn3/++eN/g0Q6jkeAiOq4pUuX4o8//qjyL3MnJydkZGRUKkGaHLvn1KlTyl+Xl5cjKipKeZQnMDAQly5dQsOGDeHr61vpR9XSAwDW1tZwd3fH8ePHKy2PjIxEixYtVF5PkyZNYGZm9shb5Fu2bInz58+jsLBQuezEiROQSqVo2rSpyp/zKI/bVi1btkRkZGSl/0+RkZGwsrKCh4eHSusPDAxERkYGDA0Nq2xvda7XMTY2hlwur7LcyckJ48ePx48//oivvvoK69atU3mdRLqKBYiojmvdujVeeuklfP3115WW9+jRA3fu3MGnn36KhIQErFq1Cn/++afGPnfVqlXYuXMnrly5grfeegt3797FxIkTAQBvvfUWcnJy8OKLL+Kff/7BjRs3cODAAUycOLHaL9jHmTFjBpYtW4Zt27YhPj4es2bNwvnz5/HOO++ovA5TU1O8//77mDlzJjZv3oyEhAScOnUKGzZsAAC89NJLMDU1xbhx43Dx4kUcOnQIb7/9NsaMGaO8ULo2Hret3nzzTaSkpODtt9/GlStX8Pvvv2P+/PmYPn16pdNtj9OnTx+EhIRg6NCh2L9/P5KSkhAZGYk5c+bg7NmzKuds2LAhTp8+jaSkJGRlZUGhUGDevHn4/fffcf36dVy6dAl79uxRq3wS6SoWICId8NFHH1U53dWiRQusXr0aq1atgr+/P/75559H3iFVE0uXLsWyZcvg7++PY8eO4ffff1cebXB3d8eJEycgl8vRv39/+Pn54Z133oGNjY3KX+oPTJ06FWFhYQgLC0Pr1q2xb98+7N69G02aNFFrPXPnzkVYWBjmzZuHFi1aYOTIkcrrX8zNzbF//37k5OSgXbt2eP7559G7d2+sXLlSrc94lMdtKw8PD4SHh+Off/6Bv78/Jk+ejFdeeQVz5sxRef0SiQTh4eHo1q0bJk6ciKZNm2LUqFFISkpSq8C99957MDAwQMuWLeHk5ITk5GQYGxtj9uzZaNOmDbp16wYDAwP8/PPPam8DIl0jEf77tyoREakkKSkJPj4+iI6O5hQdRDqGR4CIiIhI77AAERERkd7hKTAiIiLSOzwCRERERHqHBYiIiIj0DgsQERER6R0WICIiItI7LEBERESkd1iAiIiISO+wABEREZHeYQEiIiIivfN/XjLH6Oo0HRoAAAAASUVORK5CYII=",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "## how many features are needed to explain 90% of the variance?\n",
    "print(\"We need \" + str(len(PCAcalc_expl_var(x_train=vif_x_train,target_expl_variance=0.9)[0])) + \" features (out of \" + str(vif_x_train.shape[1]) + \" available) to explain 90 percent of the variance after VIF dimensionality reduction\")\n",
    "\n",
    "vif_x_train_scaled = apply_standard_scaler(vif_x_train)[0]\n",
    "pca = PCA().fit(vif_x_train_scaled)\n",
    "\n",
    "plt.plot(np.cumsum(pca.explained_variance_ratio_))\n",
    "plt.xlim(0,30,1)\n",
    "plt.xlabel('Number of components')\n",
    "plt.ylabel('Cumulative explained variance')\n",
    "plt.show()"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Dimensionality reduction with PCA"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "pca_xtrain,pca_xtest,pca_ytrain,pca_ytest = build_train_test(dataset=train_df,target_feature='actual_total_delivery_duration',test_size=0.2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [],
   "source": [
    "xtrain_pca,xtest_pca = PCAcalc_dim_reduction(\n",
    "    x_train=pca_xtrain,\n",
    "    x_test=pca_xtest,\n",
    "    desired_components=len(PCAcalc_expl_var(x_train=pca_xtrain,target_expl_variance=0.9)[0])\n",
    ")"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Model functions, model selection and first round results"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [],
   "source": [
    "## applicator function for ML methods\n",
    "\n",
    "def applyMethod_make_regression(x_train,y_train,x_test,y_test,model):\n",
    "    try:\n",
    "        model.fit(x_train,y_train)\n",
    "        ## train error\n",
    "        y_hat_train = model.predict(x_train)\n",
    "        train_RMSE = mean_squared_error(y_true=y_train,y_pred=y_hat_train,squared=False)\n",
    "        train_MAE = mean_absolute_error(y_true=y_train,y_pred=y_hat_train)\n",
    "        ## test error\n",
    "        y_hat_test = model.predict(x_test)\n",
    "        test_RMSE = mean_squared_error(y_true=y_test,y_pred=y_hat_test,squared=False)\n",
    "        test_MAE = mean_absolute_error(y_true=y_test,y_pred=y_hat_test)\n",
    "        eval_stats = {\n",
    "            'train_RMSE:':train_RMSE,\n",
    "            # gives context to the RMSE; the lower the ratio the better\n",
    "            'train_RMSE_to_mean_ratio':train_RMSE / (np.mean(y_train)),\n",
    "            'train_MAE:':train_MAE,\n",
    "            'test_RMSE:':test_RMSE,\n",
    "            # gives context to the RMSE; the lower the ratio the better\n",
    "            'test_RMSE_to_mean_ratio':test_RMSE / (np.mean(y_test)),\n",
    "            'test_MAE:':test_MAE\n",
    "        }\n",
    "        trained_model = model\n",
    "    except Exception as ex:\n",
    "        raise ex\n",
    "    return trained_model,y_hat_test,eval_stats"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "def ApplyMethod_autoML(x_train,x_test,y_train,y_test,time_budget=int,metric=str,estimator_list=list,task=str):\n",
    "    auto_ml = AutoML()\n",
    "    class xgbGamma(XGBoostEstimator):\n",
    "        def __init__(self, task=\"regression\", **config):\n",
    "            super().__init__(objective='reg:gamma', **config)\n",
    "    auto_ml.add_learner(learner_name=\"xgb_gamma\",learner_class=xgbGamma)\n",
    "    try:\n",
    "        auto_ml.fit(X_train=x_train,y_train=y_train,time_budget=time_budget,metric=metric,estimator_list=estimator_list,task=task,seed=42,verbose=0)\n",
    "        ## test error\n",
    "        y_hat_test = auto_ml.predict(x_test)\n",
    "        test_RMSE = mean_squared_error(y_true=y_test,y_pred=y_hat_test,squared=False)\n",
    "        test_MAE = mean_absolute_error(y_true=y_test,y_pred=y_hat_test)\n",
    "        ## feature importances\n",
    "        feature_importances_df = pd.DataFrame({'features':auto_ml.feature_names_in_,'importance':auto_ml.feature_importances_})\n",
    "        ## train error\n",
    "        y_hat_train = auto_ml.predict(x_train)\n",
    "        train_RMSE = mean_squared_error(y_true=y_train,y_pred=y_hat_train,squared=False)\n",
    "        train_MAE = mean_absolute_error(y_true=y_train,y_pred=y_hat_train)\n",
    "        eval_stats = {\n",
    "            'train_RMSE:':train_RMSE,\n",
    "            # gives context to the RMSE; the lower the ratio the better\n",
    "            'train_RMSE_to_mean_ratio':train_RMSE / (np.mean(y_train)),\n",
    "            'train_MAE:':train_MAE,\n",
    "            'test_RMSE:':test_RMSE,\n",
    "            # gives context to the RMSE; the lower the ratio the better\n",
    "            'test_RMSE_to_mean_ratio':test_RMSE / (np.mean(y_test)),\n",
    "            'test_MAE:':test_MAE\n",
    "        }\n",
    "        trained_model = auto_ml\n",
    "    except Exception as ex:\n",
    "        raise ex\n",
    "    return trained_model,feature_importances_df,eval_stats"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [],
   "source": [
    "#TODO: #7 fix, finish cross validation function -- DONE\n",
    "def evalMethod_cross_validate(model,model_params=dict,dataset=pd.DataFrame,target_feature=str,n_splits=int):\n",
    "## evaluates a model using k-fold cross-validation\n",
    "    try:\n",
    "        X = np.array(dataset.drop(columns=str(target_feature)))\n",
    "        y = np.array(dataset[[str(target_feature)]])\n",
    "        cv = KFold(n_splits=n_splits,shuffle=True,random_state=42)\n",
    "        scores = cross_val_score(estimator=model,fit_params=model_params,X=X,y=y,scoring='neg_root_mean_squared_error',cv=cv,n_jobs=-1)\n",
    "    except Exception as ex:\n",
    "        raise ex\n",
    "    return scores"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### PCA versus VIF dimensionality reduction:\n",
    "\n",
    "The PCA featureset took 4.6 seconds to produce, and needed 2 minutes 14.8 seconds to train a vanilla xgboost regression that produced the following performance statistics:\n",
    "\n",
    "    {'train_RMSE:': 796.2581,\n",
    "    'train_RMSE_to_mean_ratio': 0.2789955436479982,\n",
    "    'train_MAE:': 575.54443,\n",
    "    'test_RMSE:': 1045.1416,\n",
    "    'test_RMSE_to_mean_ratio': 0.3666982263503457,\n",
    "    'test_MAE:': 648.3273}\n",
    "\n",
    "The VIF featureset took 5 minutes 19.2 seconds to produce, and needed 7.8 seconds to train a vanilla xgboost regression that produced the following performance statistics:\n",
    "\n",
    "    {'train_RMSE:': 862.523,\n",
    "    'train_RMSE_to_mean_ratio': 0.302213655111118,\n",
    "    'train_MAE:': 612.2747,\n",
    "    'test_RMSE:': 1037.2075,\n",
    "    'test_RMSE_to_mean_ratio': 0.3639144755129205,\n",
    "    'test_MAE:': 652.5748}\n",
    "\n",
    "For comparison, FLAML needed needed 14 minutes, 15.6 seconds to and xgboost regression model on a dataset without any dimensionality reduction. That model produced the following performance statistics:\n",
    "\n",
    "    {'train_RMSE:': 1262.1234,\n",
    "    'train_RMSE_to_mean_ratio': 0.4422269612931676,\n",
    "    'train_MAE:': 624.24036,\n",
    "    'test_RMSE:': 1005.10626,\n",
    "    'test_RMSE_to_mean_ratio': 0.3526514331588428,\n",
    "    'test_MAE:': 623.9275}\n",
    "\n",
    "(Update with results of cross-validation)\n",
    "\n",
    "In short, the null hypothesis is true in that dimensionality reduction in this instance does not improve the forecasting accuracy."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 74,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'train_RMSE:': 1094.7580574599147,\n",
       " 'train_RMSE_to_mean_ratio': 0.38358493637162616,\n",
       " 'train_MAE:': 736.6306754416084,\n",
       " 'test_RMSE:': 1195.7121863919986,\n",
       " 'test_RMSE_to_mean_ratio': 0.4195273992729103,\n",
       " 'test_MAE:': 739.3343840308352}"
      ]
     },
     "execution_count": 74,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pca_regressions = ApplyMethod_autoML(\n",
    "    x_train=xtrain_pca,\n",
    "    x_test=xtest_pca,\n",
    "    y_train=pca_ytrain,\n",
    "    y_test=pca_ytest,\n",
    "    time_budget=60,\n",
    "    metric='mse',\n",
    "    estimator_list=['rf','extra_tree','lgbm','xgboost'],\n",
    "    task='regression'\n",
    "    )\n",
    "pca_regressions[2]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "## vif xgbregressor\n",
    "trained_xgbregressor = ApplyMethod_autoML(\n",
    "    x_train=vif_x_train,\n",
    "    x_test=vif_x_test,\n",
    "    y_test=vif_y_test,\n",
    "    y_train=vif_y_train,\n",
    "    time_budget=60,\n",
    "    metric='mse',\n",
    "    estimator_list=['rf','extra_tree','lgbm','xgboost'],\n",
    "    task='regression'\n",
    "    )\n",
    "trained_xgbregressor[2]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 66,
   "metadata": {},
   "outputs": [],
   "source": [
    "## train-test split with no dimensionality reduction\n",
    "base_xtrain,base_xtest,base_ytrain,base_ytest = build_train_test(dataset=train_df,target_feature='actual_total_delivery_duration',test_size=0.2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 67,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'train_RMSE:': 926.54297,\n",
       " 'train_RMSE_to_mean_ratio': 0.32483146316497724,\n",
       " 'train_MAE:': 623.2688,\n",
       " 'test_RMSE:': 991.1644,\n",
       " 'test_RMSE_to_mean_ratio': 0.34764720852738124,\n",
       " 'test_MAE:': 627.4639}"
      ]
     },
     "execution_count": 67,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "auto_xgb_no_reduction = ApplyMethod_autoML(\n",
    "    x_train=base_xtrain,\n",
    "    x_test=base_xtest,\n",
    "    y_train=base_ytrain,\n",
    "    y_test=base_ytest,\n",
    "    time_budget=60,\n",
    "    metric='mse',\n",
    "    estimator_list=['xgb_gamma','xgboost'],\n",
    "    task='regression'\n",
    ")\n",
    "auto_xgb_no_reduction[2]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 136,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'baseline_MAE:': 783.5442750429107, 'baseline_RMSE:': 19249.89784985233}"
      ]
     },
     "execution_count": 136,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# does the first attempt at modeling improve on a baseline, average-derived model?\n",
    "createBaseline(historical_data=historical_data)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Is there a relationship between error and delivery duration?"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 78,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Pearson's r: 0.5741122441052466\n",
      "Spearman's rho: 0.11978694426832791\n",
      "Kendall's tau: 0.07788535681338618\n",
      "Linear regression p-value: 0.0\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<AxesSubplot:xlabel='abs_error', ylabel='actual_total_delivery_duration'>"
      ]
     },
     "execution_count": 78,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "model = auto_xgb_no_reduction[0]\n",
    "y_hat = model.predict(train_df.drop(columns=['actual_total_delivery_duration']))\n",
    "error_train_df = train_df.copy()\n",
    "error_train_df['yhat'] = y_hat.reshape(-1,1)\n",
    "error_train_df['abs_error'] = abs(error_train_df['actual_total_delivery_duration'] - error_train_df['yhat'])\n",
    "print(\"Pearson's r: \" + str(np.corrcoef(x=error_train_df['actual_total_delivery_duration'],y=error_train_df['abs_error'])[1,0]))\n",
    "print(\"Spearman's rho: \" + str(error_train_df['actual_total_delivery_duration'].corr(error_train_df['abs_error'],method='spearman')))\n",
    "print(\"Kendall's tau: \" + str(error_train_df['actual_total_delivery_duration'].corr(error_train_df['abs_error'],method='kendall')))\n",
    "print(\"Linear regression p-value: \" + str(linregress(error_train_df['actual_total_delivery_duration'],error_train_df['abs_error']).pvalue))\n",
    "error_train_df.plot.scatter(x='abs_error',y='actual_total_delivery_duration')"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Building a model to classify outliers\n",
    "\n",
    "Given that there is a linear relationship between actual total delivery duration and the absolute error for each observation/order, one could argue that there are features unavailable to us in this dataset that are confounding the features we can use.\n",
    "\n",
    "How might we go about removing outlier delivery durations while still providing an estimated delivery duration to the customer?\n",
    "\n",
    "Let's assume that from a customer's perspective, the importance of the accuracy of the delivery decreases with the estimated total delivery duration. At a certain point it might be more viable to simply inform a customer that their order will be delivered in at least two hour's time (or that the service is \"finding a Dasher...\"), rather than providing an actual estimate.\n",
    "\n",
    "If it is possible to build a model to classify an order as an outlier and then use that model as a filter to decide which orders receive genuine estimates, the delivery duration estimator could then only be trained on data we know will produce higher-quality results."
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### What about a model that predicts order prep/in-store time?\n",
    "\n",
    "1. Is the amount of time an order spends in the restaurant (prep time) an important feature? \n",
    "2. Would the accuracy of a total delivery time model be improved by having an estimation of order preparation time for each order?\n",
    "3. Assuming the answers to the first two questions are true, is it worth keeping the store and category prepartion stats 'tables' I made during the feature engineering step?"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'train_RMSE:': 904.6304,\n",
       " 'train_RMSE_to_mean_ratio': 0.3169233164096178,\n",
       " 'train_MAE:': 42.500366,\n",
       " 'test_RMSE:': 349.1138,\n",
       " 'test_RMSE_to_mean_ratio': 0.12240001528748212,\n",
       " 'test_MAE:': 41.543285}"
      ]
     },
     "execution_count": 43,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "## first and foremost, is this model worth building? Is in-store time an important feature?\n",
    "hypoth_df = prep_time_df.join(train_df[['actual_total_delivery_duration']])\n",
    "# hypoth_feature_df = VIF_dim_reduction(feature_frame=hypoth_df,target_feature='actual_total_delivery_duration',vif_limit=20)[0]\n",
    "hypoth_xtrain,hypoth_xtest,hypoth_ytrain,hypoth_ytest = build_train_test(dataset=hypoth_df,target_feature='actual_total_delivery_duration',test_size=0.2)\n",
    "hypoth_feature_importances = ApplyMethod_autoML(\n",
    "    x_train=hypoth_xtrain,\n",
    "    x_test=hypoth_xtest,\n",
    "    y_train=hypoth_ytrain,\n",
    "    y_test=hypoth_ytest,\n",
    "    time_budget=60,\n",
    "    metric='mse',\n",
    "    estimator_list=['xgb_gamma','xgboost'],\n",
    "    task='regression'\n",
    ")\n",
    "hypoth_feature_importances[2]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>features</th>\n",
       "      <th>importance</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>136</th>\n",
       "      <td>est_order_total_prep_time</td>\n",
       "      <td>0.309335</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>9</th>\n",
       "      <td>estimated_store_to_consumer_driving_duration</td>\n",
       "      <td>0.180432</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>10</th>\n",
       "      <td>est_time_non-prep</td>\n",
       "      <td>0.169372</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>14</th>\n",
       "      <td>onshift_to_outstanding</td>\n",
       "      <td>0.148681</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>13</th>\n",
       "      <td>busy_to_outstanding</td>\n",
       "      <td>0.122805</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>17</th>\n",
       "      <td>store_est_time_prep_per_item_median</td>\n",
       "      <td>0.014416</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>8</th>\n",
       "      <td>estimated_order_place_duration</td>\n",
       "      <td>0.012803</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>25</th>\n",
       "      <td>store_est_median_total_prep_time</td>\n",
       "      <td>0.009709</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>19</th>\n",
       "      <td>store_est_time_prep_per_item_std</td>\n",
       "      <td>0.007571</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>16</th>\n",
       "      <td>store_est_time_prep_per_item_max</td>\n",
       "      <td>0.004082</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>61</th>\n",
       "      <td>clean_store_primary_category__indian</td>\n",
       "      <td>0.003340</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>18</th>\n",
       "      <td>store_est_time_prep_per_item_mean</td>\n",
       "      <td>0.002422</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>109</th>\n",
       "      <td>order_protocol__6.0</td>\n",
       "      <td>0.001605</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>26</th>\n",
       "      <td>store_est_mean_total_prep_time</td>\n",
       "      <td>0.001563</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>15</th>\n",
       "      <td>store_est_time_prep_per_item_min</td>\n",
       "      <td>0.001519</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                                         features  importance\n",
       "136                     est_order_total_prep_time    0.309335\n",
       "9    estimated_store_to_consumer_driving_duration    0.180432\n",
       "10                              est_time_non-prep    0.169372\n",
       "14                         onshift_to_outstanding    0.148681\n",
       "13                            busy_to_outstanding    0.122805\n",
       "17            store_est_time_prep_per_item_median    0.014416\n",
       "8                  estimated_order_place_duration    0.012803\n",
       "25               store_est_median_total_prep_time    0.009709\n",
       "19               store_est_time_prep_per_item_std    0.007571\n",
       "16               store_est_time_prep_per_item_max    0.004082\n",
       "61           clean_store_primary_category__indian    0.003340\n",
       "18              store_est_time_prep_per_item_mean    0.002422\n",
       "109                           order_protocol__6.0    0.001605\n",
       "26                 store_est_mean_total_prep_time    0.001563\n",
       "15               store_est_time_prep_per_item_min    0.001519"
      ]
     },
     "execution_count": 44,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "hypoth_feature_importances[1].sort_values(by='importance',ascending=False).head(15)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Per the performance scores and feature importance table above, providing a delivery time estimation model with an order prep time estimation greatly improves the quality of estimations. Since the feature est_order_total_prep_time is leaky (it depends on the value we are trying to predict), we will need to estimate it instead."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {},
   "outputs": [],
   "source": [
    "## produce estimates with mergable indices\n",
    "## todo: #6 fix appending of new columns -- done\n",
    "\n",
    "def predict_prep_time(prep_xtrain,prep_xtest,prep_ytrain,prep_ytest,time_budget=int,metric=str,estimator_list=list,prep_est_features_df=pd.DataFrame,delivery_est_features_df=pd.DataFrame) -> pd.DataFrame:\n",
    "    prep_est_features = prep_est_features_df.drop(columns=['est_order_total_prep_time'])\n",
    "    prep_time_model = ApplyMethod_autoML(\n",
    "    x_train=prep_xtrain,\n",
    "    x_test=prep_xtest,\n",
    "    y_train=prep_ytrain,\n",
    "    y_test=prep_ytest,\n",
    "    time_budget=time_budget,\n",
    "    metric=metric,\n",
    "    estimator_list=estimator_list,\n",
    "    task='regression'\n",
    "    )\n",
    "    try:\n",
    "       yhat = prep_time_model[0].predict(prep_est_features)\n",
    "       prep_est_features['pred_order_prep_time'] = yhat.reshape(-1,1)\n",
    "       #osgir = prep_xtrain[['pred_order_prep_time']]\n",
    "       delivery_est_features = delivery_est_features_df.join(prep_est_features[['pred_order_prep_time']])\n",
    "       delivery_est_features['est_composite_delivery'] = delivery_est_features['pred_order_prep_time'] + delivery_est_features['est_time_non-prep']\n",
    "    except Exception as ex:\n",
    "        raise ex\n",
    "    return delivery_est_features,prep_time_model[2]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {},
   "outputs": [],
   "source": [
    "## what if I did not use any dimensionality reduction?\n",
    "store_prep_split = build_train_test(dataset=prep_time_df,target_feature='est_order_total_prep_time',test_size=0.2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n",
      "check Spark installation...This line should appear only once.\n",
      "\n",
      "use_spark=True requires installation of PySpark. Please run pip install flaml[spark]\n",
      "        and check [here](https://spark.apache.org/docs/latest/api/python/getting_started/install.html)\n",
      "        for more details about installing Spark.\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "{'train_RMSE:': 1271.4131,\n",
       " 'train_RMSE_to_mean_ratio': 0.6356934286878115,\n",
       " 'train_MAE:': 620.153,\n",
       " 'test_RMSE:': 960.77496,\n",
       " 'test_RMSE_to_mean_ratio': 0.4806765506624559,\n",
       " 'test_MAE:': 622.0028}"
      ]
     },
     "execution_count": 44,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "deliv_est_df = predict_prep_time(\n",
    "    prep_xtrain=store_prep_split[0],\n",
    "    prep_xtest=store_prep_split[1],\n",
    "    prep_ytrain=store_prep_split[2],\n",
    "    prep_ytest=store_prep_split[3],\n",
    "    time_budget=120,\n",
    "    metric='r2',\n",
    "    estimator_list=['xgb_gamma'],\n",
    "    prep_est_features_df=prep_time_df,\n",
    "    delivery_est_features_df=train_df\n",
    ")\n",
    "deliv_est_df[1]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "metadata": {},
   "outputs": [],
   "source": [
    "deliv_xtrain,deliv_xtest,deliv_ytrain,deliv_ytest = build_train_test(dataset=deliv_est_df[0],target_feature='actual_total_delivery_duration',test_size=0.2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 72,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'train_RMSE:': 907.9414091211183,\n",
       " 'train_RMSE_to_mean_ratio': 0.3180832875822893,\n",
       " 'train_MAE:': 590.8988058029486,\n",
       " 'test_RMSE:': 949.8923890480916,\n",
       " 'test_RMSE_to_mean_ratio': 0.3330342224360306,\n",
       " 'test_MAE:': 614.9219966432594}"
      ]
     },
     "execution_count": 72,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "deliv_prep_model = ApplyMethod_autoML(\n",
    "    x_train=deliv_xtrain,\n",
    "    x_test=deliv_xtest,\n",
    "    y_train=deliv_ytrain,\n",
    "    y_test=deliv_ytest,\n",
    "    time_budget=120,\n",
    "    metric='r2',\n",
    "    estimator_list=['rf','extra_tree','lgbm','xgboost','xgb_gamma'],\n",
    "    task='regression'\n",
    ")\n",
    "deliv_prep_model[2]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 47,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "-1105.4249097279348"
      ]
     },
     "execution_count": 47,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "np.average(\n",
    "    evalMethod_cross_validate(model=AutoML(),model_params={'time_budget':120,'metric':'r2','estimator_list':['rf','xgboost','lgbm','extra_tree'],'task':'regression','seed':42},dataset=deliv_est_df[0],target_feature='actual_total_delivery_duration',n_splits=10)\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 73,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>features</th>\n",
       "      <th>importance</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>137</th>\n",
       "      <td>est_composite_delivery</td>\n",
       "      <td>0.650622</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>136</th>\n",
       "      <td>pred_order_prep_time</td>\n",
       "      <td>0.313782</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>14</th>\n",
       "      <td>onshift_to_outstanding</td>\n",
       "      <td>0.014919</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>10</th>\n",
       "      <td>est_time_non-prep</td>\n",
       "      <td>0.002618</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>29</th>\n",
       "      <td>clean_store_primary_category__alcohol</td>\n",
       "      <td>0.002037</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>115</th>\n",
       "      <td>created_hour_of_day__4</td>\n",
       "      <td>0.001956</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>13</th>\n",
       "      <td>busy_to_outstanding</td>\n",
       "      <td>0.001752</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>26</th>\n",
       "      <td>store_est_mean_total_prep_time</td>\n",
       "      <td>0.001612</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>18</th>\n",
       "      <td>store_est_time_prep_per_item_mean</td>\n",
       "      <td>0.001471</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>6</th>\n",
       "      <td>total_busy_dashers</td>\n",
       "      <td>0.001439</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>25</th>\n",
       "      <td>store_est_median_total_prep_time</td>\n",
       "      <td>0.001323</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>22</th>\n",
       "      <td>category_est_time_prep_per_item_median</td>\n",
       "      <td>0.000798</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>7</th>\n",
       "      <td>total_outstanding_orders</td>\n",
       "      <td>0.000659</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>5</th>\n",
       "      <td>total_onshift_dashers</td>\n",
       "      <td>0.000652</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>93</th>\n",
       "      <td>clean_store_primary_category__thai</td>\n",
       "      <td>0.000649</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                                   features  importance\n",
       "137                  est_composite_delivery    0.650622\n",
       "136                    pred_order_prep_time    0.313782\n",
       "14                   onshift_to_outstanding    0.014919\n",
       "10                        est_time_non-prep    0.002618\n",
       "29    clean_store_primary_category__alcohol    0.002037\n",
       "115                  created_hour_of_day__4    0.001956\n",
       "13                      busy_to_outstanding    0.001752\n",
       "26           store_est_mean_total_prep_time    0.001612\n",
       "18        store_est_time_prep_per_item_mean    0.001471\n",
       "6                        total_busy_dashers    0.001439\n",
       "25         store_est_median_total_prep_time    0.001323\n",
       "22   category_est_time_prep_per_item_median    0.000798\n",
       "7                  total_outstanding_orders    0.000659\n",
       "5                     total_onshift_dashers    0.000652\n",
       "93       clean_store_primary_category__thai    0.000649"
      ]
     },
     "execution_count": 73,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "deliv_prep_model[1].sort_values(by='importance',ascending=False).head(15)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Results thusfar:\n",
    "\n",
    "Before I go any farther, it is worth asking if all this modeling is even worth the effort. The function `create_baseline` uses basic averages and outlier removal to provide a delivery duration value for each observation. The performance results for that method are as follows:\n",
    "\n",
    "    {'baseline_MAE:': 783.5442750429107,\n",
    "    'baseline_RMSE:': 19249.89784985233}\n",
    "\n",
    "In short: yes, building models is worth the time, effort and expense for this project. The profession of data science survives.\n",
    "\n",
    "The most successful model I have built so far requires no dimensionality reduction and and uses an intermediary model to predict the amount of time each order spends in-store. Per the feature importance plot above, the output of that model - `pred_order_prep_time` - is an immensely important feature for estimating total delivery time. To answer question three from the headline above, features from the function `addFeatures_store_prep_stats` make up six of the top fifteen features for the final model, so that featureset should be retained.\n",
    "\n",
    "The prep time featureset needs 21.6 seconds to train a vanilla xgboost regression model, and when the results of that model are combined with the total delivery time featureset only 7.9 seconds are required to train a vanilla xgboost regression model. That final model produces the following performance statistics:\n",
    "\n",
    "    {'train_RMSE:': 585.9649,\n",
    "    'train_MAE:': 437.55215,\n",
    "    'test_RMSE:': 720.2951,\n",
    "    'test_MAE:': 516.90875}\n",
    "\n",
    "(update with cross validation)\n",
    "\n",
    "Comparison of performance statistics with the best single model approach:\n",
    "\n",
    "    {'single_model_train_RMSE:': 796.1849 (-26.4034%),\n",
    "    'single_model_train_MAE:': 561.24567 (-22.0391%),\n",
    "    'single_model_test_RMSE:': 1014.22095 (-28.9805%),\n",
    "    'single_model_test_MAE:': 608.876 (-15.1044%)}\n",
    "\n",
    "For further comparison, the best RMSE from Stratascratch's 'offical solution' for this dataset is 986.6912510458277 using the same holdout/test size. User raihanmasud on GitHub developed a gradient-boosted tree model with calculated paramaters (L1, learning rate, etc.) and achieved a cross-validated RMSE of 1030.6837 from a holdout/test size of 0.33. My model's performance statistics using that sample size are as follows:\n",
    "\n",
    "    {'train_RMSE:': 634.61224,\n",
    "    'train_MAE:': 469.97223,\n",
    "    'test_RMSE:': 1353.4868,\n",
    "    'test_MAE:': 513.71735}\n",
    "\n",
    "That significantly outperforms my model, so the next step is to do some parameter tuning and model selection."
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3.9.7 ('base')",
   "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": "3.9.15"
  },
  "orig_nbformat": 4,
  "vscode": {
   "interpreter": {
    "hash": "96d1af5f869624c072d9e03e4f7fcffed44526ca720f1a36a76cd678c41e0e80"
   }
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}