movie_recommendation_with_LightFM_person.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Movie Recommendation for Single Users"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "This notebook uses the users and movie datasets without friends grouping yet. Here, we test two different models, manually created Siamese Neural Network model and the existing model called LightFM. We compare the results of both and choose the one that suits best our website.\n",
    "For movie recommendation for friends, please look at another notebook."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import pandas as pd\n",
    "import json\n",
    "import datetime, time\n",
    "from sklearn.model_selection import train_test_split\n",
    "import itertools\n",
    "import os\n",
    "import zipfile\n",
    "\n",
    "import numpy as np\n",
    "\n",
    "import requests\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "import scipy.sparse as sp\n",
    "from sklearn.metrics import roc_auc_score\n",
    "\n",
    "\n",
    "from tensorflow.keras import backend as K\n",
    "from tensorflow.keras.models import Model\n",
    "from tensorflow.keras.layers import Embedding, Flatten, Input, Layer\n",
    "from tensorflow.keras.optimizers import Adam\n",
    "\n",
    "from tensorflow.keras.utils import plot_model"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## (1) Read data"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## (1.1) Movies dataframe"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(1315, 22)\n"
     ]
    },
    {
     "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>movie_id_ml</th>\n",
       "      <th>title</th>\n",
       "      <th>release</th>\n",
       "      <th>unknown</th>\n",
       "      <th>action</th>\n",
       "      <th>adventure</th>\n",
       "      <th>animation</th>\n",
       "      <th>childrens</th>\n",
       "      <th>comedy</th>\n",
       "      <th>crime</th>\n",
       "      <th>...</th>\n",
       "      <th>fantasy</th>\n",
       "      <th>noir</th>\n",
       "      <th>horror</th>\n",
       "      <th>musical</th>\n",
       "      <th>mystery</th>\n",
       "      <th>romance</th>\n",
       "      <th>scifi</th>\n",
       "      <th>thriller</th>\n",
       "      <th>war</th>\n",
       "      <th>western</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>toy story</td>\n",
       "      <td>1995</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>2</td>\n",
       "      <td>goldeneye</td>\n",
       "      <td>1995</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>3</td>\n",
       "      <td>four rooms</td>\n",
       "      <td>1995</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>4</td>\n",
       "      <td>get shorty</td>\n",
       "      <td>1995</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>5</td>\n",
       "      <td>copycat</td>\n",
       "      <td>1995</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>5 rows × 22 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "   movie_id_ml       title  release  unknown  action  adventure  animation  \\\n",
       "0            1   toy story     1995        0       0          0          1   \n",
       "1            2   goldeneye     1995        0       1          1          0   \n",
       "2            3  four rooms     1995        0       0          0          0   \n",
       "3            4  get shorty     1995        0       1          0          0   \n",
       "4            5     copycat     1995        0       0          0          0   \n",
       "\n",
       "   childrens  comedy  crime  ...  fantasy  noir  horror  musical  mystery  \\\n",
       "0          1       1      0  ...        0     0       0        0        0   \n",
       "1          0       0      0  ...        0     0       0        0        0   \n",
       "2          0       0      0  ...        0     0       0        0        0   \n",
       "3          0       1      0  ...        0     0       0        0        0   \n",
       "4          0       0      1  ...        0     0       0        0        0   \n",
       "\n",
       "   romance  scifi  thriller  war  western  \n",
       "0        0      0         0    0        0  \n",
       "1        0      0         1    0        0  \n",
       "2        0      0         1    0        0  \n",
       "3        0      0         0    0        0  \n",
       "4        0      0         1    0        0  \n",
       "\n",
       "[5 rows x 22 columns]"
      ]
     },
     "execution_count": 2,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df_movies = pd.read_csv(\"data/movies_cast_company.csv\", encoding='utf8')\n",
    "df_movies[\"cast\"] = df_movies[\"cast\"].apply(lambda x: json.loads(x))\n",
    "df_movies[\"company\"] = df_movies[\"company\"].apply(lambda x: json.loads(x))\n",
    "\n",
    "df_movies = df_movies.drop([\"url\"]+list(df_movies.columns[-4:]), axis=1)\n",
    "\n",
    "print(df_movies.shape)\n",
    "df_movies.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "def string2ts(string, fmt=\"%Y-%m-%d %H:%M:%S\"):\n",
    "    dt = datetime.datetime.strptime(string, fmt)\n",
    "    t_tuple = dt.timetuple()\n",
    "    return int(time.mktime(t_tuple))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## (2.2) Ratings dataframe"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(100000, 4)\n"
     ]
    },
    {
     "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>user_id</th>\n",
       "      <th>movie_id_ml</th>\n",
       "      <th>rating</th>\n",
       "      <th>rating_timestamp</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>196</td>\n",
       "      <td>242</td>\n",
       "      <td>3</td>\n",
       "      <td>881247349</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>186</td>\n",
       "      <td>302</td>\n",
       "      <td>3</td>\n",
       "      <td>891710542</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>22</td>\n",
       "      <td>377</td>\n",
       "      <td>1</td>\n",
       "      <td>878883516</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>244</td>\n",
       "      <td>51</td>\n",
       "      <td>2</td>\n",
       "      <td>880603323</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>166</td>\n",
       "      <td>346</td>\n",
       "      <td>1</td>\n",
       "      <td>886393996</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "   user_id  movie_id_ml  rating  rating_timestamp\n",
       "0      196          242       3         881247349\n",
       "1      186          302       3         891710542\n",
       "2       22          377       1         878883516\n",
       "3      244           51       2         880603323\n",
       "4      166          346       1         886393996"
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df_ratings = pd.read_csv(\"data/ratings.csv\")\n",
    "\n",
    "df_ratings.rating_timestamp = df_ratings.rating_timestamp.apply(lambda x: string2ts(x))\n",
    "\n",
    "print(df_ratings.shape)\n",
    "df_ratings.head()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## (2.3) Users dataframe"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(943, 4)\n"
     ]
    },
    {
     "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>user_id</th>\n",
       "      <th>user_age</th>\n",
       "      <th>user_gender</th>\n",
       "      <th>user_occupation</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>24</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>2</td>\n",
       "      <td>53</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>3</td>\n",
       "      <td>23</td>\n",
       "      <td>0</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>4</td>\n",
       "      <td>24</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>5</td>\n",
       "      <td>33</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "   user_id  user_age  user_gender  user_occupation\n",
       "0        1        24            0                0\n",
       "1        2        53            1                1\n",
       "2        3        23            0                2\n",
       "3        4        24            0                0\n",
       "4        5        33            1                1"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df_users = pd.read_csv(\"data/users.csv\")\n",
    "\n",
    "# TODO: just temporary, later remove\n",
    "additional_rows = [\"user_zipcode\"]\n",
    "df_users = df_users.drop(additional_rows, axis=1)\n",
    "\n",
    "num2occupation = dict(enumerate(df_users.user_occupation.unique()))\n",
    "occupation2num = {y:x for x,y in num2occupation.items()}\n",
    "num2gender = dict(enumerate(df_users.user_gender.unique()))\n",
    "gender2num = {y:x for x,y in num2gender.items()}\n",
    "df_users.user_occupation = df_users.user_occupation.apply(lambda x: occupation2num[x])\n",
    "df_users.user_gender = df_users.user_gender.apply(lambda x: gender2num[x])\n",
    "\n",
    "\n",
    "print(df_users.shape)\n",
    "df_users.head()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## (2.4) Meging dataframes into one "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(90851, 28)\n"
     ]
    },
    {
     "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>movie_id_ml</th>\n",
       "      <th>title</th>\n",
       "      <th>release</th>\n",
       "      <th>unknown</th>\n",
       "      <th>action</th>\n",
       "      <th>adventure</th>\n",
       "      <th>animation</th>\n",
       "      <th>childrens</th>\n",
       "      <th>comedy</th>\n",
       "      <th>crime</th>\n",
       "      <th>...</th>\n",
       "      <th>scifi</th>\n",
       "      <th>thriller</th>\n",
       "      <th>war</th>\n",
       "      <th>western</th>\n",
       "      <th>user_id</th>\n",
       "      <th>rating</th>\n",
       "      <th>rating_timestamp</th>\n",
       "      <th>user_age</th>\n",
       "      <th>user_gender</th>\n",
       "      <th>user_occupation</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>toy story</td>\n",
       "      <td>1995</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>308</td>\n",
       "      <td>4</td>\n",
       "      <td>887732932</td>\n",
       "      <td>60</td>\n",
       "      <td>0</td>\n",
       "      <td>18</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>4</td>\n",
       "      <td>get shorty</td>\n",
       "      <td>1995</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>308</td>\n",
       "      <td>5</td>\n",
       "      <td>887734290</td>\n",
       "      <td>60</td>\n",
       "      <td>0</td>\n",
       "      <td>18</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>5</td>\n",
       "      <td>copycat</td>\n",
       "      <td>1995</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>308</td>\n",
       "      <td>4</td>\n",
       "      <td>887736008</td>\n",
       "      <td>60</td>\n",
       "      <td>0</td>\n",
       "      <td>18</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>7</td>\n",
       "      <td>twelve monkeys</td>\n",
       "      <td>1995</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>308</td>\n",
       "      <td>4</td>\n",
       "      <td>887735247</td>\n",
       "      <td>60</td>\n",
       "      <td>0</td>\n",
       "      <td>18</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>8</td>\n",
       "      <td>babe</td>\n",
       "      <td>1995</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>308</td>\n",
       "      <td>5</td>\n",
       "      <td>887733096</td>\n",
       "      <td>60</td>\n",
       "      <td>0</td>\n",
       "      <td>18</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>5 rows × 28 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "   movie_id_ml           title  release  unknown  action  adventure  \\\n",
       "0            1       toy story     1995        0       0          0   \n",
       "1            4      get shorty     1995        0       1          0   \n",
       "2            5         copycat     1995        0       0          0   \n",
       "3            7  twelve monkeys     1995        0       0          0   \n",
       "4            8            babe     1995        0       0          0   \n",
       "\n",
       "   animation  childrens  comedy  crime  ...  scifi  thriller  war  western  \\\n",
       "0          1          1       1      0  ...      0         0    0        0   \n",
       "1          0          0       1      0  ...      0         0    0        0   \n",
       "2          0          0       0      1  ...      0         1    0        0   \n",
       "3          0          0       0      0  ...      1         0    0        0   \n",
       "4          0          1       1      0  ...      0         0    0        0   \n",
       "\n",
       "   user_id  rating  rating_timestamp  user_age  user_gender  user_occupation  \n",
       "0      308       4         887732932        60            0               18  \n",
       "1      308       5         887734290        60            0               18  \n",
       "2      308       4         887736008        60            0               18  \n",
       "3      308       4         887735247        60            0               18  \n",
       "4      308       5         887733096        60            0               18  \n",
       "\n",
       "[5 rows x 28 columns]"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df = pd.merge(df_movies, df_ratings, on=\"movie_id_ml\")\n",
    "df = pd.merge(df, df_users, on=\"user_id\")\n",
    "\n",
    "print(df.shape)\n",
    "df.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "id2movie = dict(enumerate(df.movie_id_ml.unique()))\n",
    "movie2id = {y:x for x,y in id2movie.items()}\n",
    "\n",
    "id2user = dict(enumerate(df.user_id.unique()))\n",
    "user2id = {y:x for x,y in id2user.items()}"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "df[\"iid\"] = df.apply(lambda x: movie2id[x.movie_id_ml], axis=1)\n",
    "df[\"uid\"] = df.apply(lambda x: user2id[x.user_id], axis=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "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>movie_id_ml</th>\n",
       "      <th>title</th>\n",
       "      <th>release</th>\n",
       "      <th>unknown</th>\n",
       "      <th>action</th>\n",
       "      <th>adventure</th>\n",
       "      <th>animation</th>\n",
       "      <th>childrens</th>\n",
       "      <th>comedy</th>\n",
       "      <th>crime</th>\n",
       "      <th>...</th>\n",
       "      <th>war</th>\n",
       "      <th>western</th>\n",
       "      <th>user_id</th>\n",
       "      <th>rating</th>\n",
       "      <th>rating_timestamp</th>\n",
       "      <th>user_age</th>\n",
       "      <th>user_gender</th>\n",
       "      <th>user_occupation</th>\n",
       "      <th>iid</th>\n",
       "      <th>uid</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>toy story</td>\n",
       "      <td>1995</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>308</td>\n",
       "      <td>4</td>\n",
       "      <td>887732932</td>\n",
       "      <td>60</td>\n",
       "      <td>0</td>\n",
       "      <td>18</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>4</td>\n",
       "      <td>get shorty</td>\n",
       "      <td>1995</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>308</td>\n",
       "      <td>5</td>\n",
       "      <td>887734290</td>\n",
       "      <td>60</td>\n",
       "      <td>0</td>\n",
       "      <td>18</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>5</td>\n",
       "      <td>copycat</td>\n",
       "      <td>1995</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>308</td>\n",
       "      <td>4</td>\n",
       "      <td>887736008</td>\n",
       "      <td>60</td>\n",
       "      <td>0</td>\n",
       "      <td>18</td>\n",
       "      <td>2</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>7</td>\n",
       "      <td>twelve monkeys</td>\n",
       "      <td>1995</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>308</td>\n",
       "      <td>4</td>\n",
       "      <td>887735247</td>\n",
       "      <td>60</td>\n",
       "      <td>0</td>\n",
       "      <td>18</td>\n",
       "      <td>3</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>8</td>\n",
       "      <td>babe</td>\n",
       "      <td>1995</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>308</td>\n",
       "      <td>5</td>\n",
       "      <td>887733096</td>\n",
       "      <td>60</td>\n",
       "      <td>0</td>\n",
       "      <td>18</td>\n",
       "      <td>4</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>90846</td>\n",
       "      <td>748</td>\n",
       "      <td>the saint</td>\n",
       "      <td>1997</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>729</td>\n",
       "      <td>4</td>\n",
       "      <td>893279438</td>\n",
       "      <td>19</td>\n",
       "      <td>0</td>\n",
       "      <td>5</td>\n",
       "      <td>348</td>\n",
       "      <td>942</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>90847</td>\n",
       "      <td>751</td>\n",
       "      <td>tomorrow never dies</td>\n",
       "      <td>1997</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>729</td>\n",
       "      <td>3</td>\n",
       "      <td>893279138</td>\n",
       "      <td>19</td>\n",
       "      <td>0</td>\n",
       "      <td>5</td>\n",
       "      <td>447</td>\n",
       "      <td>942</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>90848</td>\n",
       "      <td>879</td>\n",
       "      <td>the peacemaker</td>\n",
       "      <td>1997</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>729</td>\n",
       "      <td>3</td>\n",
       "      <td>893279099</td>\n",
       "      <td>19</td>\n",
       "      <td>0</td>\n",
       "      <td>5</td>\n",
       "      <td>659</td>\n",
       "      <td>942</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>90849</td>\n",
       "      <td>894</td>\n",
       "      <td>home alone 3</td>\n",
       "      <td>1997</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>729</td>\n",
       "      <td>1</td>\n",
       "      <td>893279311</td>\n",
       "      <td>19</td>\n",
       "      <td>0</td>\n",
       "      <td>5</td>\n",
       "      <td>875</td>\n",
       "      <td>942</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>90850</td>\n",
       "      <td>901</td>\n",
       "      <td>mr. magoo</td>\n",
       "      <td>1997</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>729</td>\n",
       "      <td>1</td>\n",
       "      <td>893279291</td>\n",
       "      <td>19</td>\n",
       "      <td>0</td>\n",
       "      <td>5</td>\n",
       "      <td>878</td>\n",
       "      <td>942</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>90851 rows × 30 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "       movie_id_ml                title  release  unknown  action  adventure  \\\n",
       "0                1            toy story     1995        0       0          0   \n",
       "1                4           get shorty     1995        0       1          0   \n",
       "2                5              copycat     1995        0       0          0   \n",
       "3                7       twelve monkeys     1995        0       0          0   \n",
       "4                8                 babe     1995        0       0          0   \n",
       "...            ...                  ...      ...      ...     ...        ...   \n",
       "90846          748            the saint     1997        0       1          0   \n",
       "90847          751  tomorrow never dies     1997        0       1          0   \n",
       "90848          879       the peacemaker     1997        0       1          0   \n",
       "90849          894         home alone 3     1997        0       0          0   \n",
       "90850          901            mr. magoo     1997        0       0          0   \n",
       "\n",
       "       animation  childrens  comedy  crime  ...  war  western  user_id  \\\n",
       "0              1          1       1      0  ...    0        0      308   \n",
       "1              0          0       1      0  ...    0        0      308   \n",
       "2              0          0       0      1  ...    0        0      308   \n",
       "3              0          0       0      0  ...    0        0      308   \n",
       "4              0          1       1      0  ...    0        0      308   \n",
       "...          ...        ...     ...    ...  ...  ...      ...      ...   \n",
       "90846          0          0       0      0  ...    0        0      729   \n",
       "90847          0          0       0      0  ...    0        0      729   \n",
       "90848          0          0       0      0  ...    1        0      729   \n",
       "90849          0          1       1      0  ...    0        0      729   \n",
       "90850          0          0       1      0  ...    0        0      729   \n",
       "\n",
       "       rating  rating_timestamp  user_age  user_gender  user_occupation  iid  \\\n",
       "0           4         887732932        60            0               18    0   \n",
       "1           5         887734290        60            0               18    1   \n",
       "2           4         887736008        60            0               18    2   \n",
       "3           4         887735247        60            0               18    3   \n",
       "4           5         887733096        60            0               18    4   \n",
       "...       ...               ...       ...          ...              ...  ...   \n",
       "90846       4         893279438        19            0                5  348   \n",
       "90847       3         893279138        19            0                5  447   \n",
       "90848       3         893279099        19            0                5  659   \n",
       "90849       1         893279311        19            0                5  875   \n",
       "90850       1         893279291        19            0                5  878   \n",
       "\n",
       "       uid  \n",
       "0        0  \n",
       "1        0  \n",
       "2        0  \n",
       "3        0  \n",
       "4        0  \n",
       "...    ...  \n",
       "90846  942  \n",
       "90847  942  \n",
       "90848  942  \n",
       "90849  942  \n",
       "90850  942  \n",
       "\n",
       "[90851 rows x 30 columns]"
      ]
     },
     "execution_count": 9,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## (2.5) Remove random user that will be used for prediction"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "# TODO; remove this line when real user\n",
    "new_user_id = 942 #user_movie_ratings_training[0][0]\n",
    "df = df[df.uid != new_user_id]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "df = df.drop_duplicates()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Number of users features: 942\n",
      "Number of movies features: 1280\n"
     ]
    }
   ],
   "source": [
    "# shape [n_users, n_user_features]\n",
    "df_users = df[[\"uid\", \"user_age\", \"user_gender\", \"user_occupation\"]].drop_duplicates()\n",
    "print(f\"Number of users features: {df_users.shape[0]}\")\n",
    "\n",
    "df_movies = df[[\"iid\"]+list(df.columns[3:22])].drop_duplicates()\n",
    "print(f\"Number of movies features: {df_movies.shape[0]}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## (2) Simulate our user input"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "First, `user_movie_ratings_training` is used to simulate initial user input that we will ask him to rate some random movies. Later, we will try to recommend movies to him, but we know what is his preference thanks to `user_movie_ratings_test`. Both lists are extracted from the existed dataset."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "columns = [\"user_age\", \"user_gender\", \"user_occupation\",  \"uid\", \"iid\", \"rating\", \"rating_timestamp\"]\n",
    "\n",
    "# user info\n",
    "user_age = 19\n",
    "user_gender = 0\n",
    "user_occupation = 5\n",
    "\n",
    "# ratings\n",
    "user_movie_ratings_training = [[942, 891, 4, 893279438],\n",
    "                               [942, 130, 2, 893279061],\n",
    "                               [942, 133, 2, 893279138],\n",
    "                               [942, 488, 4, 893279438],\n",
    "                               [942, 854, 3, 893279004],\n",
    "                               [942, 137, 3, 893279438],\n",
    "                               [942, 141, 4, 893279437],\n",
    "                               [942, 609, 3, 893279438],\n",
    "                               [942, 612, 4, 893279438],\n",
    "                               [942, 857, 1, 893279173]]\n",
    "\n",
    "user_movie_ratings_test = [[942, 343, 1, 893278968],\n",
    "                           [942, 862, 5, 893279437],\n",
    "                           [942, 1124, 4, 893279437],\n",
    "                           [942, 918, 4, 893279438],\n",
    "                           [942, 438, 2, 893278949],\n",
    "                           [942, 348, 4, 893279438],\n",
    "                           [942, 447, 3, 893279138],\n",
    "                           [942, 659, 3, 893279099],\n",
    "                           [942, 875, 1, 893279311],\n",
    "                           [942, 878, 1, 893279291]]\n",
    "\n",
    "new_user_id = 942 #user_movie_ratings_training[0][0]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Append the new information to existing dataframes"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Number of users features: 943\n",
      "Number of movies features: 1280\n"
     ]
    }
   ],
   "source": [
    "df_users = df_users.append({\"uid\": new_user_id, \"user_age\":user_age, \"user_gender\":user_gender, \"user_occupation\":user_occupation}, ignore_index=True)\n",
    "print(f\"Number of users features: {df_users.shape[0]}\")\n",
    "print(f\"Number of movies features: {df_movies.shape[0]}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Prepare data for the training"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "# create new user dataframe with training data\n",
    "\n",
    "data_new_user_training = []\n",
    "for x in user_movie_ratings_training:\n",
    "    data_new_user_training.append([user_age, user_gender, user_occupation] + x)\n",
    "\n",
    "data_new_user_test = []\n",
    "for x in user_movie_ratings_test:\n",
    "    data_new_user_test.append([user_age, user_gender, user_occupation] + x)\n",
    "\n",
    "# user initial input that will be given to him to rate it before recommendation\n",
    "df_new_user_train = pd.DataFrame(data_new_user_training, columns=columns)\n",
    "# the input that will be checked if recommendation works fine\n",
    "df_new_user_test = pd.DataFrame(data_new_user_test, columns=columns)\n",
    "\n",
    "# training and test data\n",
    "train_idx, test_idx = train_test_split(range(df.shape[0]), test_size=0.2, random_state=42)\n",
    "\n",
    "df_train = df.iloc[train_idx]\n",
    "df_train = pd.concat([df_train, df_new_user_train], sort=False)\n",
    "\n",
    "df_test = df.iloc[test_idx]\n",
    "df_test = pd.concat([df_test, df_new_user_test], sort=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "df_train = df_train[[\"uid\", \"iid\", \"rating\", \"rating_timestamp\"]]\n",
    "df_test = df_test[[\"uid\", \"iid\", \"rating\", \"rating_timestamp\"]]\n",
    "df_new_user_train = df_new_user_train[[\"uid\", \"iid\", \"rating\", \"rating_timestamp\"]]\n",
    "df_new_user_test = df_new_user_test[[\"uid\", \"iid\", \"rating\", \"rating_timestamp\"]]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## (3) Training the Siamese Neural Network"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Map user and movie IDs to consequitive numbers."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## (3.1) Data pre-processing"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Users number:  943 943\n",
      "Movies number:  1280 1280\n"
     ]
    }
   ],
   "source": [
    "uids = set(df_train.uid.unique()).union(set(df_test.uid.unique()))\n",
    "iids = set(df_train.iid.unique()).union(set(df_test.iid.unique()))\n",
    "\n",
    "rows = max(uids) + 1 \n",
    "cols = max(iids) + 1\n",
    "\n",
    "print(\"Users number: \", len(uids), rows)\n",
    "print(\"Movies number: \", len(iids), cols)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "def _build_interaction_matrix(rows, cols, data):\n",
    "\n",
    "    mat = sp.lil_matrix((rows, cols), dtype=np.int32)\n",
    "\n",
    "    for uid, iid, rating, timestamp in data:\n",
    "        # Let's assume only really good things are positives\n",
    "        if rating >= 4.0:\n",
    "            mat[uid, iid] = 1.0\n",
    "\n",
    "    return mat.tocoo()\n",
    "\n",
    "def get_triplets(mat):\n",
    "    return mat.row, mat.col, np.random.randint(mat.shape[1], size=len(mat.row))\n",
    "\n",
    "def create_sparse_matrix(df):\n",
    "    \"\"\"\n",
    "    Return (train_interactions, test_interactions).\n",
    "    \"\"\"\n",
    "    return _build_interaction_matrix(rows, cols, df.values.tolist())"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Using ROC curve from [here](https://developers.google.com/machine-learning/crash-course/classification/roc-and-auc) and [here](https://en.wikipedia.org/wiki/Receiver_operating_characteristic#Area_under_the_curve)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [],
   "source": [
    "def predict(model, uid, pids):\n",
    "    user_vector = model.get_layer('user_embedding').get_weights()[0][uid]\n",
    "    item_matrix = model.get_layer('item_embedding').get_weights()[0][pids]\n",
    "\n",
    "    scores = (np.dot(user_vector,\n",
    "                     item_matrix.T))\n",
    "\n",
    "    return scores\n",
    "\n",
    "\n",
    "def full_auc(model, ground_truth):\n",
    "    \"\"\"\n",
    "    Measure AUC for model and ground truth on all items.\n",
    "\n",
    "    Returns\n",
    "    -------\n",
    "    - float AUC\n",
    "    \"\"\"\n",
    "    ground_truth = ground_truth.tocsr()\n",
    "\n",
    "    no_users, no_items = ground_truth.shape\n",
    "\n",
    "    pid_array = np.arange(no_items, dtype=np.int32)\n",
    "\n",
    "    scores = []\n",
    "\n",
    "    for user_id, row in enumerate(ground_truth):\n",
    "\n",
    "        predictions = predict(model, user_id, pid_array)\n",
    "\n",
    "        true_pids = row.indices[row.data == 1]\n",
    "\n",
    "        grnd = np.zeros(no_items, dtype=np.int32)\n",
    "        grnd[true_pids] = 1\n",
    "\n",
    "        if len(true_pids):\n",
    "            scores.append(roc_auc_score(grnd, predictions))\n",
    "\n",
    "    return np.mean(scores)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## (3.2) Model Implementation"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Here, we add new custom layer called `TripletLossLayer`."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [],
   "source": [
    "class TripletLossLayer(Layer):\n",
    "    def call(self, inputs):\n",
    "        positive_item_latent, negative_item_latent, user_latent = inputs\n",
    "\n",
    "        # Bayesian Personalised Ranking (BPR) loss\n",
    "        loss = 1.0 - K.sigmoid(\n",
    "            K.sum(user_latent * positive_item_latent, axis=-1, keepdims=True) -\n",
    "            K.sum(user_latent * negative_item_latent, axis=-1, keepdims=True))\n",
    "\n",
    "        return loss"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [],
   "source": [
    "def identity_loss(y_true, y_pred):\n",
    "    return K.mean(y_pred - 0 * y_true)\n",
    "\n",
    "def build_model(num_users, num_items, latent_dim):\n",
    "\n",
    "    positive_item_input = Input((1, ), name='positive_item_input')\n",
    "    negative_item_input = Input((1, ), name='negative_item_input')\n",
    "\n",
    "    # Shared embedding layer for positive and negative items\n",
    "    item_embedding_layer = Embedding(num_items, latent_dim, name='item_embedding', input_length=1)\n",
    "    user_input = Input((1, ), name='user_input')\n",
    "\n",
    "    positive_item_embedding = Flatten()(item_embedding_layer(positive_item_input))\n",
    "    negative_item_embedding = Flatten()(item_embedding_layer(negative_item_input))\n",
    "    user_embedding = Flatten()(Embedding(num_users, latent_dim, name='user_embedding', input_length=1)(user_input))\n",
    "    \n",
    "    loss = TripletLossLayer()((positive_item_embedding, negative_item_embedding, user_embedding))\n",
    "    \n",
    "    model = Model([positive_item_input, negative_item_input, user_input], loss)\n",
    "    \n",
    "    model.compile(loss=identity_loss, optimizer=Adam())\n",
    "    \n",
    "    return model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [],
   "source": [
    "# train.todense()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Bellow, we show the visualization of layers used in this Siamese Neural Network:"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "![](model_plot.png)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## (3.3) Training"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"model\"\n",
      "__________________________________________________________________________________________________\n",
      "Layer (type)                    Output Shape         Param #     Connected to                     \n",
      "==================================================================================================\n",
      "positive_item_input (InputLayer [(None, 1)]          0                                            \n",
      "__________________________________________________________________________________________________\n",
      "negative_item_input (InputLayer [(None, 1)]          0                                            \n",
      "__________________________________________________________________________________________________\n",
      "user_input (InputLayer)         [(None, 1)]          0                                            \n",
      "__________________________________________________________________________________________________\n",
      "item_embedding (Embedding)      (None, 1, 100)       128000      positive_item_input[0][0]        \n",
      "                                                                 negative_item_input[0][0]        \n",
      "__________________________________________________________________________________________________\n",
      "user_embedding (Embedding)      (None, 1, 100)       94300       user_input[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "flatten (Flatten)               (None, 100)          0           item_embedding[0][0]             \n",
      "__________________________________________________________________________________________________\n",
      "flatten_1 (Flatten)             (None, 100)          0           item_embedding[1][0]             \n",
      "__________________________________________________________________________________________________\n",
      "flatten_2 (Flatten)             (None, 100)          0           user_embedding[0][0]             \n",
      "__________________________________________________________________________________________________\n",
      "triplet_loss_layer (TripletLoss (None, 1)            0           flatten[0][0]                    \n",
      "                                                                 flatten_1[0][0]                  \n",
      "                                                                 flatten_2[0][0]                  \n",
      "==================================================================================================\n",
      "Total params: 222,300\n",
      "Trainable params: 222,300\n",
      "Non-trainable params: 0\n",
      "__________________________________________________________________________________________________\n",
      "None\n",
      "AUC before training 0.4990186633255141\n"
     ]
    }
   ],
   "source": [
    "latent_dim = 100\n",
    "num_epochs = 15\n",
    "\n",
    "# Read data\n",
    "train = create_sparse_matrix(df_train)\n",
    "test = create_sparse_matrix(df_test)\n",
    "num_users, num_items = train.shape\n",
    "\n",
    "# Prepare the test triplets\n",
    "test_uid, test_pid, test_nid = get_triplets(test)\n",
    "\n",
    "model = build_model(num_users, num_items, latent_dim)\n",
    "\n",
    "# Print the model structure\n",
    "print(model.summary())\n",
    "\n",
    "plot_model(model, to_file=\"model_plot.png\", show_shapes=True, show_layer_names=True)\n",
    "\n",
    "# Sanity check, should be around 0.5\n",
    "print(f'AUC before training {full_auc(model, test)}')\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 0\n",
      "Test AUC 0.8266112085581016\n",
      "Train AUC 0.894149467529688\n",
      "\n",
      "Epoch 1\n",
      "Test AUC 0.8544020668251694\n",
      "Train AUC 0.8740725657992398\n",
      "\n",
      "Epoch 2\n",
      "Test AUC 0.8640728427362321\n",
      "Train AUC 0.8838011922847954\n",
      "\n",
      "Epoch 3\n",
      "Test AUC 0.873426910297064\n",
      "Train AUC 0.8940342915924792\n",
      "\n",
      "Epoch 4\n",
      "Test AUC 0.8820197408800541\n",
      "Train AUC 0.9040221819254033\n",
      "\n",
      "Epoch 5\n",
      "Test AUC 0.8893199764002392\n",
      "Train AUC 0.912544553539495\n",
      "\n",
      "Epoch 6\n",
      "Test AUC 0.8949352693912704\n",
      "Train AUC 0.9196935299572806\n",
      "\n",
      "Epoch 7\n",
      "Test AUC 0.8993679180649193\n",
      "Train AUC 0.9254397866679173\n",
      "\n",
      "Epoch 8\n",
      "Test AUC 0.9022507920931286\n",
      "Train AUC 0.92971593091989\n",
      "\n",
      "Epoch 9\n",
      "Test AUC 0.9049499056885434\n",
      "Train AUC 0.9334791001683445\n",
      "\n",
      "Epoch 10\n",
      "Test AUC 0.9064818380288814\n",
      "Train AUC 0.9365343219061724\n",
      "\n",
      "Epoch 11\n",
      "Test AUC 0.9077681843229217\n",
      "Train AUC 0.939202323915051\n",
      "\n",
      "Epoch 12\n",
      "Test AUC 0.9089712950902361\n",
      "Train AUC 0.9415413584122592\n",
      "\n",
      "Epoch 13\n",
      "Test AUC 0.9099174552926286\n",
      "Train AUC 0.9435991108917016\n",
      "\n",
      "Epoch 14\n",
      "Test AUC 0.9108272757213374\n",
      "Train AUC 0.9455474809170649\n",
      "\n"
     ]
    }
   ],
   "source": [
    "accuracies_test = []\n",
    "accuracies_train = []\n",
    "losses = []\n",
    "\n",
    "for epoch in range(num_epochs):\n",
    "\n",
    "    print(f'Epoch {epoch}')\n",
    "\n",
    "    # Sample triplets from the training data\n",
    "    uid, pid, nid = get_triplets(train)\n",
    "\n",
    "    X = {\n",
    "        'user_input': uid,\n",
    "        'positive_item_input': pid,\n",
    "        'negative_item_input': nid\n",
    "    }\n",
    "\n",
    "    history = model.fit(X,\n",
    "                      np.ones(len(uid)),\n",
    "                      batch_size=64,\n",
    "                      epochs=1,\n",
    "                      verbose=0,\n",
    "                      shuffle=True)\n",
    "    loss = history.history[\"loss\"]\n",
    "    losses.append(loss)\n",
    "    \n",
    "    acc = full_auc(model, test)\n",
    "    accuracies_test.append(acc)\n",
    "    print(f'Test AUC {acc}')\n",
    "    acc = full_auc(model, train)\n",
    "    accuracies_train.append(acc)\n",
    "    print(f'Train AUC {acc}\\n')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1080x504 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(15,7))\n",
    "\n",
    "# Create count of the number of epochs\n",
    "epoch_count = range(1, num_epochs + 1)\n",
    "\n",
    "# Visualize loss history\n",
    "ax1.plot(epoch_count, losses, 'r--', label='Training')\n",
    "#ax1.plot(epoch_count, val_loss, 'b-', label='Validation')\n",
    "ax1.legend()\n",
    "ax1.set_xlabel('Epoch')\n",
    "ax1.set_ylabel('Loss')\n",
    "\n",
    "ax2.plot(epoch_count, accuracies_test, 'r-', label='Test')\n",
    "ax2.plot(epoch_count, accuracies_train, 'b-', label='Train')\n",
    "ax2.legend()\n",
    "ax2.set_xlabel('Epoch')\n",
    "ax2.set_ylabel('Accuracy')\n",
    "plt.show();"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## (3.4) Observe the recommendation results for our user"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "942\n",
      "[343, 862, 1124, 918, 438, 348, 447, 659, 875, 878]\n",
      "[0, 1, 1, 1, 0, 1, 0, 0, 0, 0]\n",
      "[1, 1, 0, 1, 1, 1, 1, 1, 0, 0]\n",
      "0.5\n"
     ]
    }
   ],
   "source": [
    "user_id = new_user_id\n",
    "print(user_id)\n",
    "movie_ids = [343, 862, 1124, 918, 438, 348, 447, 659, 875, 878]\n",
    "print(movie_ids)\n",
    "\n",
    "g = [0, 1, 1, 1, 0, 1, 0, 0, 0, 0]\n",
    "print(g)\n",
    "# = [1, 1, 0, 0, 1, 1]\n",
    "# p is supposed to be the score of the class with greater label (if score is less than 0.5 with 1 decimal, then it predicted 0 otherwise it predicts 1)\n",
    "p = predict(model, user_id, movie_ids)\n",
    "print([1 if np.round(i)>0.5 else 0 for i in p])\n",
    "print(roc_auc_score(g, p))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [],
   "source": [
    "def predict_movies(model, user_id):\n",
    "    movie_ids = list(df.iid.unique())\n",
    "    prediction = predict(model, user_id, movie_ids)\n",
    "    rated_high = [1 if np.round(i)>0.5 else 0 for i in prediction]\n",
    "    good_movie_ids = np.where(np.array(rated_high) == 1)[0]  \n",
    "    return list(good_movie_ids)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [],
   "source": [
    "predicted_good_movies = predict_movies(model, user_id)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [],
   "source": [
    "user_ratings = user_movie_ratings_training + user_movie_ratings_test\n",
    "user_ratings_positive = []\n",
    "user_ratings_negative = []\n",
    "for ur in user_ratings:\n",
    "    if ur[2] >= 4:\n",
    "        user_ratings_positive.append(ur[1])\n",
    "    else:\n",
    "        user_ratings_negative.append(ur[1])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "891\n",
      "488\n",
      "141\n",
      "612\n",
      "862\n",
      "1124 PREDICTED WRONG: SHOULD BE GOOD MOVIE FOR THIS USER\n",
      "918\n",
      "348\n"
     ]
    }
   ],
   "source": [
    "for m in user_ratings_positive:\n",
    "    print(m if m in predicted_good_movies else f\"{m} PREDICTED WRONG: SHOULD BE GOOD MOVIE FOR THIS USER\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "130 PREDICTED WRONG: SHOULD BE BAD MOVIE FOR THIS USER\n",
      "133 PREDICTED WRONG: SHOULD BE BAD MOVIE FOR THIS USER\n",
      "854 PREDICTED WRONG: SHOULD BE BAD MOVIE FOR THIS USER\n",
      "137 PREDICTED WRONG: SHOULD BE BAD MOVIE FOR THIS USER\n",
      "609 PREDICTED WRONG: SHOULD BE BAD MOVIE FOR THIS USER\n",
      "857 PREDICTED WRONG: SHOULD BE BAD MOVIE FOR THIS USER\n",
      "343 PREDICTED WRONG: SHOULD BE BAD MOVIE FOR THIS USER\n",
      "438 PREDICTED WRONG: SHOULD BE BAD MOVIE FOR THIS USER\n",
      "447 PREDICTED WRONG: SHOULD BE BAD MOVIE FOR THIS USER\n",
      "659 PREDICTED WRONG: SHOULD BE BAD MOVIE FOR THIS USER\n",
      "875\n",
      "878\n"
     ]
    }
   ],
   "source": [
    "for m in user_ratings_negative:\n",
    "    print(m if m not in predicted_good_movies else f\"{m} PREDICTED WRONG: SHOULD BE BAD MOVIE FOR THIS USER\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Accuracy for new user: 84.01%\n"
     ]
    }
   ],
   "source": [
    "user_sparse_matrix = create_sparse_matrix(df_new_user_test)\n",
    "acc = full_auc(model, user_sparse_matrix)\n",
    "print(f\"Accuracy for new user: {acc*100:.2f}%\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Accuracy is different with every run, but when this was written, the accuracy was **93.1%** for this user is not bad, could be better! On the test set, accuracy on average is **90.9%**."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {},
   "outputs": [],
   "source": [
    "def predict_top_k_movies(model, user_id, k):\n",
    "    n_users, n_movies = train.shape\n",
    "    prediction = predict(model, user_id, np.arange(n_movies))\n",
    "    movie_ids = df.iid.unique()\n",
    "    return movie_ids[np.argsort(-prediction)][:k]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "User 942\n",
      "     Known positives:\n",
      "        891 | good will hunting\n",
      "        488 | air force one\n",
      "        141 | murder at 1600\n",
      "        612 | the game\n",
      "     Recommended:\n",
      "        488 | air force one\n",
      "        137 | titanic\n",
      "        486 | contact\n",
      "        52 | the godfather\n",
      "        17 | star wars\n",
      "        130 | scream\n",
      "        404 | the english patient\n",
      "        0 | toy story\n",
      "        133 | liar liar\n",
      "        891 | good will hunting\n"
     ]
    }
   ],
   "source": [
    "known_positives = df_train[(df_train.iid.isin(train.tocsr()[user_id].indices))&(df_train.uid==user_id)][\"iid\"]\n",
    "\n",
    "k = 10\n",
    "top_movies = predict_top_k_movies(model, user_id, k)\n",
    "\n",
    "print(\"User %s\" % user_id)\n",
    "print(\"     Known positives:\")\n",
    "for x in known_positives[:k]:\n",
    "    print(f\"        {df[df.iid==x]['iid'].iloc[0]} | {df[df.iid==x]['title'].iloc[0]}\" )\n",
    "    \n",
    "print(\"     Recommended:\")\n",
    "for x in top_movies:\n",
    "    print(f\"        {df[df.iid==x]['iid'].iloc[0]} | {df[df.iid==x]['title'].iloc[0]}\" )"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## (4) LightFM model"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Much faster implementation that is similar doing of the code implemented above."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {},
   "outputs": [],
   "source": [
    "from lightfm import LightFM\n",
    "from lightfm.datasets import fetch_movielens\n",
    "from lightfm.evaluation import precision_at_k\n",
    "from lightfm.evaluation import auc_score"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {},
   "outputs": [],
   "source": [
    "def my2csr(df):\n",
    "    return sp.csr_matrix(df.values)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {},
   "outputs": [],
   "source": [
    "train = create_sparse_matrix(df_train)#, mat_type=\"ratings\")\n",
    "test = create_sparse_matrix(df_test)#, mat_type=\"ratings\")\n",
    "\n",
    "# shape [n_users, n_user_features]\n",
    "user_features = my2csr(df_users)\n",
    "item_features = my2csr(df_movies)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Precision: train 0.50, test 0.09.\n",
      "AUC: train 0.94, test 0.93.\n"
     ]
    }
   ],
   "source": [
    "use_features = False\n",
    "\n",
    "loss_type = \"warp\"  # \"bpr\"\n",
    "\n",
    "model = LightFM(learning_rate=0.05, loss=loss_type, max_sampled=100)\n",
    "\n",
    "if use_features:\n",
    "    model.fit_partial(train, epochs=20, user_features=user_features, item_features=item_features)\n",
    "    #model.fit(train, epochs=50, user_features=user_features, item_features=item_features)\n",
    "    train_precision = precision_at_k(model, train, k=10, user_features=user_features, item_features=item_features).mean()\n",
    "    test_precision = precision_at_k(model, test, k=10, train_interactions=train, user_features=user_features, item_features=item_features).mean()\n",
    "\n",
    "    train_auc = auc_score(model, train, user_features=user_features, item_features=item_features).mean()\n",
    "    test_auc = auc_score(model, test, train_interactions=train, user_features=user_features, item_features=item_features).mean()\n",
    "\n",
    "    print('Precision: train %.2f, test %.2f.' % (train_precision, test_precision))\n",
    "    print('AUC: train %.2f, test %.2f.' % (train_auc, test_auc))\n",
    "else:\n",
    "    model.fit_partial(train, epochs=20)\n",
    "    \n",
    "    train_precision = precision_at_k(model, train, k=10).mean()\n",
    "    test_precision = precision_at_k(model, test, k=10).mean()\n",
    "\n",
    "    train_auc = auc_score(model, train).mean()\n",
    "    test_auc = auc_score(model, test, train_interactions=train).mean()\n",
    "\n",
    "    print('Precision: train %.2f, test %.2f.' % (train_precision, test_precision))\n",
    "    print('AUC: train %.2f, test %.2f.' % (train_auc, test_auc))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {},
   "outputs": [],
   "source": [
    "user_ratings = user_movie_ratings_training + user_movie_ratings_test\n",
    "user_ratings_positive = []\n",
    "user_ratings_negative = []\n",
    "for ur in user_ratings:\n",
    "    if ur[2] >= 4:\n",
    "        user_ratings_positive.append(ur[1])\n",
    "    else:\n",
    "        user_ratings_negative.append(ur[1])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {},
   "outputs": [],
   "source": [
    "def predict_movies(model, user_id):\n",
    "    n_movies = train.shape[1]\n",
    "    if use_features:\n",
    "        prediction = model.predict(user_id, np.arange(n_movies), user_features=user_features, item_features=item_features)  #predict(model, user_id, movie_ids)\n",
    "    else:\n",
    "        prediction = model.predict(user_id, np.arange(n_movies))  #predict(model, user_id, movie_ids)\n",
    "    \n",
    "    rated_high = [1 if np.round(i)>0.5 else 0 for i in prediction]\n",
    "    good_movie_ids = np.where(np.array(rated_high) == 1)[0]  \n",
    "    return list(good_movie_ids)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {},
   "outputs": [],
   "source": [
    "predicted_good_movies = predict_movies(model, user_id)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "891\n",
      "488\n",
      "141\n",
      "612\n",
      "862 PREDICTED WRONG: SHOULD BE GOOD MOVIE FOR THIS USER\n",
      "1124 PREDICTED WRONG: SHOULD BE GOOD MOVIE FOR THIS USER\n",
      "918 PREDICTED WRONG: SHOULD BE GOOD MOVIE FOR THIS USER\n",
      "348\n"
     ]
    }
   ],
   "source": [
    "for m in user_ratings_positive:\n",
    "    print(m if m in predicted_good_movies else f\"{m} PREDICTED WRONG: SHOULD BE GOOD MOVIE FOR THIS USER\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "130 PREDICTED WRONG: SHOULD BE BAD MOVIE FOR THIS USER\n",
      "133 PREDICTED WRONG: SHOULD BE BAD MOVIE FOR THIS USER\n",
      "854 PREDICTED WRONG: SHOULD BE BAD MOVIE FOR THIS USER\n",
      "137 PREDICTED WRONG: SHOULD BE BAD MOVIE FOR THIS USER\n",
      "609 PREDICTED WRONG: SHOULD BE BAD MOVIE FOR THIS USER\n",
      "857\n",
      "343\n",
      "438 PREDICTED WRONG: SHOULD BE BAD MOVIE FOR THIS USER\n",
      "447 PREDICTED WRONG: SHOULD BE BAD MOVIE FOR THIS USER\n",
      "659 PREDICTED WRONG: SHOULD BE BAD MOVIE FOR THIS USER\n",
      "875\n",
      "878\n"
     ]
    }
   ],
   "source": [
    "\n",
    "for m in user_ratings_negative:\n",
    "    print(m if m not in predicted_good_movies else f\"{m} PREDICTED WRONG: SHOULD BE BAD MOVIE FOR THIS USER\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Accuracy for new user: 88.54%\n"
     ]
    }
   ],
   "source": [
    "user_sparse_matrix = create_sparse_matrix(df_new_user_test)\n",
    "if use_features:\n",
    "    acc = auc_score(model, user_sparse_matrix, user_features=user_features, item_features=item_features).mean()\n",
    "else:\n",
    "    acc = auc_score(model, user_sparse_matrix).mean()\n",
    "\n",
    "print(f\"Accuracy for new user: {acc*100:.2f}%\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We can see that this model performs preatty similar in accuracy as the Siamese NN above. Around **88.5%** accuracy. The main difference is that this one is way more faster. Therefore, we will use this one for all our future endevors."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 46,
   "metadata": {},
   "outputs": [],
   "source": [
    "def predict_top_k_movies(model, user_id, k):\n",
    "    n_users, n_movies = train.shape\n",
    "    if use_features:\n",
    "        prediction = model.predict(user_id, np.arange(n_movies), user_features=user_features, item_features=item_features)#predict(model, user_id, np.arange(n_movies))\n",
    "    else:\n",
    "        prediction = model.predict(user_id, np.arange(n_movies))\n",
    "    \n",
    "    movie_ids = np.arange(train.shape[1])\n",
    "    return movie_ids[np.argsort(-prediction)][:k]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 47,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "User 942\n",
      "     Known positives:\n",
      "        141 | murder at 1600\n",
      "        488 | air force one\n",
      "        612 | the game\n",
      "        891 | good will hunting\n",
      "     Recommended:\n",
      "        1279 | b. monkey\n",
      "        1278 | sliding doors\n",
      "        1277 | hush\n",
      "        1276 | nightwatch\n",
      "        1275 | hugo pool\n",
      "        1273 | tainted\n",
      "        1274 | entertaining angels: the dorothy day story\n",
      "        1272 | the next step\n",
      "        1271 | a brother's kiss\n",
      "        1269 | twisted\n"
     ]
    }
   ],
   "source": [
    "k = 10\n",
    "user_id = new_user_id\n",
    "movie_ids = np.arange(train.shape[1])\n",
    "\n",
    "n_users, n_items = train.shape\n",
    "\n",
    "known_positives = movie_ids[train.tocsr()[user_id].indices]\n",
    "\n",
    "scores = model.predict(user_id, np.arange(n_items), user_features=user_features, item_features=item_features)\n",
    "top_items = movie_ids[np.argsort(-scores)]\n",
    "\n",
    "print(\"User %s\" % user_id)\n",
    "print(\"     Known positives:\")\n",
    "\n",
    "for x in known_positives[:k]:\n",
    "    print(f\"        {df[df.iid==x]['iid'].iloc[0]} | {df[df.iid==x]['title'].iloc[0]}\" )\n",
    "    \n",
    "print(\"     Recommended:\")\n",
    "for x in top_items[:k]:\n",
    "    print(f\"        {df[df.iid==x]['iid'].iloc[0]} | {df[df.iid==x]['title'].iloc[0]}\" )"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.8"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}