elliptic-dataset-gnn.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "77a82adc-adfa-460f-aa87-6089b4d1d494",
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "import copy\n",
    "import torch\n",
    "import warnings\n",
    "import numpy as np\n",
    "import pandas as pd\n",
    "import networkx as nx\n",
    "import seaborn as sns\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "from sklearn.metrics import confusion_matrix, classification_report\n",
    "from sklearn.model_selection import train_test_split\n",
    "\n",
    "from torch_geometric.utils import to_networkx\n",
    "from torch_geometric.data import Data, DataLoader\n",
    "\n",
    "import torch.nn.functional as F\n",
    "from torch.nn import Linear, Dropout\n",
    "from torch_geometric.nn import GCNConv, GATv2Conv\n",
    "\n",
    "warnings.filterwarnings('ignore')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "457a8a47-f59c-4d31-a47b-b2db14840761",
   "metadata": {},
   "outputs": [],
   "source": [
    "# from torch_geometric.datasets import EllipticBitcoinDataset\n",
    "\n",
    "# dataset = EllipticBitcoinDataset(root='./data/elliptic-bitcoin-dataset')"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8ad9e519-5f77-4ecb-aced-9bd83809a999",
   "metadata": {},
   "source": [
    "## Config"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "aaecec42-7b5a-43d7-852f-3536c2427942",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Using device: cpu\n"
     ]
    }
   ],
   "source": [
    "class Config:\n",
    "    seed = 0\n",
    "    learning_rate = 0.001\n",
    "    weight_decay = 1e-5\n",
    "    input_dim = 165\n",
    "    output_dim = 1\n",
    "    hidden_size = 128\n",
    "    num_epochs = 100\n",
    "    checkpoints_dir = './models/elliptic_gnn'\n",
    "    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n",
    "    \n",
    "print(\"Using device:\", Config.device)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "86f9d5db-57dd-48a8-b56f-7711c460a0bf",
   "metadata": {},
   "source": [
    "# Data loading/preparation"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3d6f81a8-240d-43b2-b5ad-f35e19160087",
   "metadata": {},
   "source": [
    "[Dataset link](https://www.kaggle.com/datasets/ellipticco/elliptic-data-set)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "e38e83fd-939b-45c8-bbfa-a4146358414c",
   "metadata": {},
   "outputs": [],
   "source": [
    "df_features = pd.read_csv('./data/elliptic_bitcoin_dataset/elliptic_txs_features.csv', header=None)\n",
    "df_edges = pd.read_csv(\"./data/elliptic_bitcoin_dataset/elliptic_txs_edgelist.csv\")\n",
    "df_classes =  pd.read_csv(\"./data/elliptic_bitcoin_dataset/elliptic_txs_classes.csv\")\n",
    "df_classes['class'] = df_classes['class'].map({'unknown': 2, '1': 1, '2': 0})"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "b088b5be-af28-4afe-bc71-23efcbec2f84",
   "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>0</th>\n",
       "      <th>1</th>\n",
       "      <th>2</th>\n",
       "      <th>3</th>\n",
       "      <th>4</th>\n",
       "      <th>5</th>\n",
       "      <th>6</th>\n",
       "      <th>7</th>\n",
       "      <th>8</th>\n",
       "      <th>9</th>\n",
       "      <th>...</th>\n",
       "      <th>157</th>\n",
       "      <th>158</th>\n",
       "      <th>159</th>\n",
       "      <th>160</th>\n",
       "      <th>161</th>\n",
       "      <th>162</th>\n",
       "      <th>163</th>\n",
       "      <th>164</th>\n",
       "      <th>165</th>\n",
       "      <th>166</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>230425980</td>\n",
       "      <td>1</td>\n",
       "      <td>-0.171469</td>\n",
       "      <td>-0.184668</td>\n",
       "      <td>-1.201369</td>\n",
       "      <td>-0.121970</td>\n",
       "      <td>-0.043875</td>\n",
       "      <td>-0.113002</td>\n",
       "      <td>-0.061584</td>\n",
       "      <td>-0.162097</td>\n",
       "      <td>...</td>\n",
       "      <td>-0.562153</td>\n",
       "      <td>-0.600999</td>\n",
       "      <td>1.461330</td>\n",
       "      <td>1.461369</td>\n",
       "      <td>0.018279</td>\n",
       "      <td>-0.087490</td>\n",
       "      <td>-0.131155</td>\n",
       "      <td>-0.097524</td>\n",
       "      <td>-0.120613</td>\n",
       "      <td>-0.119792</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>5530458</td>\n",
       "      <td>1</td>\n",
       "      <td>-0.171484</td>\n",
       "      <td>-0.184668</td>\n",
       "      <td>-1.201369</td>\n",
       "      <td>-0.121970</td>\n",
       "      <td>-0.043875</td>\n",
       "      <td>-0.113002</td>\n",
       "      <td>-0.061584</td>\n",
       "      <td>-0.162112</td>\n",
       "      <td>...</td>\n",
       "      <td>0.947382</td>\n",
       "      <td>0.673103</td>\n",
       "      <td>-0.979074</td>\n",
       "      <td>-0.978556</td>\n",
       "      <td>0.018279</td>\n",
       "      <td>-0.087490</td>\n",
       "      <td>-0.131155</td>\n",
       "      <td>-0.097524</td>\n",
       "      <td>-0.120613</td>\n",
       "      <td>-0.119792</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>232022460</td>\n",
       "      <td>1</td>\n",
       "      <td>-0.172107</td>\n",
       "      <td>-0.184668</td>\n",
       "      <td>-1.201369</td>\n",
       "      <td>-0.121970</td>\n",
       "      <td>-0.043875</td>\n",
       "      <td>-0.113002</td>\n",
       "      <td>-0.061584</td>\n",
       "      <td>-0.162749</td>\n",
       "      <td>...</td>\n",
       "      <td>0.670883</td>\n",
       "      <td>0.439728</td>\n",
       "      <td>-0.979074</td>\n",
       "      <td>-0.978556</td>\n",
       "      <td>-0.098889</td>\n",
       "      <td>-0.106715</td>\n",
       "      <td>-0.131155</td>\n",
       "      <td>-0.183671</td>\n",
       "      <td>-0.120613</td>\n",
       "      <td>-0.119792</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>232438397</td>\n",
       "      <td>1</td>\n",
       "      <td>0.163054</td>\n",
       "      <td>1.963790</td>\n",
       "      <td>-0.646376</td>\n",
       "      <td>12.409294</td>\n",
       "      <td>-0.063725</td>\n",
       "      <td>9.782742</td>\n",
       "      <td>12.414558</td>\n",
       "      <td>-0.163645</td>\n",
       "      <td>...</td>\n",
       "      <td>-0.577099</td>\n",
       "      <td>-0.613614</td>\n",
       "      <td>0.241128</td>\n",
       "      <td>0.241406</td>\n",
       "      <td>1.072793</td>\n",
       "      <td>0.085530</td>\n",
       "      <td>-0.131155</td>\n",
       "      <td>0.677799</td>\n",
       "      <td>-0.120613</td>\n",
       "      <td>-0.119792</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>230460314</td>\n",
       "      <td>1</td>\n",
       "      <td>1.011523</td>\n",
       "      <td>-0.081127</td>\n",
       "      <td>-1.201369</td>\n",
       "      <td>1.153668</td>\n",
       "      <td>0.333276</td>\n",
       "      <td>1.312656</td>\n",
       "      <td>-0.061584</td>\n",
       "      <td>-0.163523</td>\n",
       "      <td>...</td>\n",
       "      <td>-0.511871</td>\n",
       "      <td>-0.400422</td>\n",
       "      <td>0.517257</td>\n",
       "      <td>0.579382</td>\n",
       "      <td>0.018279</td>\n",
       "      <td>0.277775</td>\n",
       "      <td>0.326394</td>\n",
       "      <td>1.293750</td>\n",
       "      <td>0.178136</td>\n",
       "      <td>0.179117</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>5 rows × 167 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "         0    1         2         3         4          5         6    \\\n",
       "0  230425980    1 -0.171469 -0.184668 -1.201369  -0.121970 -0.043875   \n",
       "1    5530458    1 -0.171484 -0.184668 -1.201369  -0.121970 -0.043875   \n",
       "2  232022460    1 -0.172107 -0.184668 -1.201369  -0.121970 -0.043875   \n",
       "3  232438397    1  0.163054  1.963790 -0.646376  12.409294 -0.063725   \n",
       "4  230460314    1  1.011523 -0.081127 -1.201369   1.153668  0.333276   \n",
       "\n",
       "        7          8         9    ...       157       158       159       160  \\\n",
       "0 -0.113002  -0.061584 -0.162097  ... -0.562153 -0.600999  1.461330  1.461369   \n",
       "1 -0.113002  -0.061584 -0.162112  ...  0.947382  0.673103 -0.979074 -0.978556   \n",
       "2 -0.113002  -0.061584 -0.162749  ...  0.670883  0.439728 -0.979074 -0.978556   \n",
       "3  9.782742  12.414558 -0.163645  ... -0.577099 -0.613614  0.241128  0.241406   \n",
       "4  1.312656  -0.061584 -0.163523  ... -0.511871 -0.400422  0.517257  0.579382   \n",
       "\n",
       "        161       162       163       164       165       166  \n",
       "0  0.018279 -0.087490 -0.131155 -0.097524 -0.120613 -0.119792  \n",
       "1  0.018279 -0.087490 -0.131155 -0.097524 -0.120613 -0.119792  \n",
       "2 -0.098889 -0.106715 -0.131155 -0.183671 -0.120613 -0.119792  \n",
       "3  1.072793  0.085530 -0.131155  0.677799 -0.120613 -0.119792  \n",
       "4  0.018279  0.277775  0.326394  1.293750  0.178136  0.179117  \n",
       "\n",
       "[5 rows x 167 columns]"
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# here column 0 stands for node_id, column 1 is the time axis\n",
    "df_features.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "5de29d33-eca1-4c70-b0bb-10bb50b267e8",
   "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>txId1</th>\n",
       "      <th>txId2</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>230425980</td>\n",
       "      <td>5530458</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>232022460</td>\n",
       "      <td>232438397</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>230460314</td>\n",
       "      <td>230459870</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>230333930</td>\n",
       "      <td>230595899</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>232013274</td>\n",
       "      <td>232029206</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "       txId1      txId2\n",
       "0  230425980    5530458\n",
       "1  232022460  232438397\n",
       "2  230460314  230459870\n",
       "3  230333930  230595899\n",
       "4  232013274  232029206"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df_edges.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "0d0f195d-a34b-437d-bd2e-d9154b022c6f",
   "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>txId</th>\n",
       "      <th>class</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>230425980</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>5530458</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>232022460</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>232438397</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>230460314</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "        txId  class\n",
       "0  230425980      2\n",
       "1    5530458      2\n",
       "2  232022460      2\n",
       "3  232438397      0\n",
       "4  230460314      2"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df_classes.head()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "e15f4f7c-1059-4370-89aa-9bdd7a2c0ac9",
   "metadata": {},
   "source": [
    "0 - legitimate<br>\n",
    "1 - fraud<br>\n",
    "2 - unknown class<br>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "0204623c-1108-4fc5-b230-a8d1e97ab421",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "2    157205\n",
       "0     42019\n",
       "1      4545\n",
       "Name: class, dtype: int64"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df_classes['class'].value_counts()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "id": "13daa363-ce0f-46d0-8a64-c1ba209ad31f",
   "metadata": {},
   "outputs": [],
   "source": [
    "# merging node features DF with classes DF\n",
    "df_merge = df_features.merge(df_classes, how='left', right_on=\"txId\", left_on=0)\n",
    "df_merge = df_merge.sort_values(0).reset_index(drop=True)\n",
    "\n",
    "# extracting classified/non-classified nodes\n",
    "classified = df_merge.loc[df_merge['class'].loc[df_merge['class']!=2].index].drop('txId', axis=1)\n",
    "unclassified = df_merge.loc[df_merge['class'].loc[df_merge['class']==2].index].drop('txId', axis=1)\n",
    "\n",
    "# extracting classified/non-classified edges\n",
    "classified_edges = df_edges.loc[df_edges['txId1'].isin(classified[0]) & df_edges['txId2'].isin(classified[0])]\n",
    "unclassifed_edges = df_edges.loc[df_edges['txId1'].isin(unclassified[0]) | df_edges['txId2'].isin(unclassified[0])]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "1713a1cc-10ed-469c-8311-13a37d1a2ce0",
   "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>0</th>\n",
       "      <th>1</th>\n",
       "      <th>2</th>\n",
       "      <th>3</th>\n",
       "      <th>4</th>\n",
       "      <th>5</th>\n",
       "      <th>6</th>\n",
       "      <th>7</th>\n",
       "      <th>8</th>\n",
       "      <th>9</th>\n",
       "      <th>...</th>\n",
       "      <th>159</th>\n",
       "      <th>160</th>\n",
       "      <th>161</th>\n",
       "      <th>162</th>\n",
       "      <th>163</th>\n",
       "      <th>164</th>\n",
       "      <th>165</th>\n",
       "      <th>166</th>\n",
       "      <th>txId</th>\n",
       "      <th>class</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>1076</td>\n",
       "      <td>48</td>\n",
       "      <td>-0.168500</td>\n",
       "      <td>0.270909</td>\n",
       "      <td>-0.091383</td>\n",
       "      <td>-0.046932</td>\n",
       "      <td>-0.043875</td>\n",
       "      <td>-0.029140</td>\n",
       "      <td>-0.061584</td>\n",
       "      <td>-0.163591</td>\n",
       "      <td>...</td>\n",
       "      <td>1.461330</td>\n",
       "      <td>1.461369</td>\n",
       "      <td>0.018279</td>\n",
       "      <td>0.470019</td>\n",
       "      <td>1.216796</td>\n",
       "      <td>1.151607</td>\n",
       "      <td>1.519700</td>\n",
       "      <td>1.521399</td>\n",
       "      <td>1076</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>2534</td>\n",
       "      <td>6</td>\n",
       "      <td>-0.170834</td>\n",
       "      <td>-0.131425</td>\n",
       "      <td>1.018602</td>\n",
       "      <td>0.028105</td>\n",
       "      <td>0.055376</td>\n",
       "      <td>0.054722</td>\n",
       "      <td>-0.061584</td>\n",
       "      <td>-0.163572</td>\n",
       "      <td>...</td>\n",
       "      <td>0.955101</td>\n",
       "      <td>0.459257</td>\n",
       "      <td>-0.098889</td>\n",
       "      <td>-0.087490</td>\n",
       "      <td>-0.099080</td>\n",
       "      <td>-0.122137</td>\n",
       "      <td>-0.379970</td>\n",
       "      <td>-0.379288</td>\n",
       "      <td>2534</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>3181</td>\n",
       "      <td>34</td>\n",
       "      <td>1.305212</td>\n",
       "      <td>-0.210553</td>\n",
       "      <td>-1.756361</td>\n",
       "      <td>-0.121970</td>\n",
       "      <td>97.300650</td>\n",
       "      <td>-0.113002</td>\n",
       "      <td>-0.061584</td>\n",
       "      <td>1.348765</td>\n",
       "      <td>...</td>\n",
       "      <td>0.059948</td>\n",
       "      <td>0.113967</td>\n",
       "      <td>-0.098889</td>\n",
       "      <td>1.969527</td>\n",
       "      <td>0.037532</td>\n",
       "      <td>-0.131010</td>\n",
       "      <td>0.006994</td>\n",
       "      <td>0.017772</td>\n",
       "      <td>3181</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>3321</td>\n",
       "      <td>1</td>\n",
       "      <td>-0.169615</td>\n",
       "      <td>-0.184668</td>\n",
       "      <td>-1.201369</td>\n",
       "      <td>-0.121970</td>\n",
       "      <td>-0.043875</td>\n",
       "      <td>-0.113002</td>\n",
       "      <td>-0.061584</td>\n",
       "      <td>-0.160199</td>\n",
       "      <td>...</td>\n",
       "      <td>0.241128</td>\n",
       "      <td>0.241406</td>\n",
       "      <td>-0.098889</td>\n",
       "      <td>-0.087490</td>\n",
       "      <td>-0.084674</td>\n",
       "      <td>-0.140597</td>\n",
       "      <td>1.519700</td>\n",
       "      <td>1.521399</td>\n",
       "      <td>3321</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>3889</td>\n",
       "      <td>48</td>\n",
       "      <td>-0.086232</td>\n",
       "      <td>-0.101835</td>\n",
       "      <td>-0.646376</td>\n",
       "      <td>-0.121970</td>\n",
       "      <td>17.046997</td>\n",
       "      <td>-0.113002</td>\n",
       "      <td>-0.061584</td>\n",
       "      <td>-0.074885</td>\n",
       "      <td>...</td>\n",
       "      <td>0.082065</td>\n",
       "      <td>0.114773</td>\n",
       "      <td>-0.098889</td>\n",
       "      <td>8.948005</td>\n",
       "      <td>1.024948</td>\n",
       "      <td>-0.009570</td>\n",
       "      <td>-0.080708</td>\n",
       "      <td>-0.123601</td>\n",
       "      <td>3889</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>5 rows × 169 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "      0   1         2         3         4         5          6         7  \\\n",
       "0  1076  48 -0.168500  0.270909 -0.091383 -0.046932  -0.043875 -0.029140   \n",
       "1  2534   6 -0.170834 -0.131425  1.018602  0.028105   0.055376  0.054722   \n",
       "2  3181  34  1.305212 -0.210553 -1.756361 -0.121970  97.300650 -0.113002   \n",
       "3  3321   1 -0.169615 -0.184668 -1.201369 -0.121970  -0.043875 -0.113002   \n",
       "4  3889  48 -0.086232 -0.101835 -0.646376 -0.121970  17.046997 -0.113002   \n",
       "\n",
       "          8         9  ...       159       160       161       162       163  \\\n",
       "0 -0.061584 -0.163591  ...  1.461330  1.461369  0.018279  0.470019  1.216796   \n",
       "1 -0.061584 -0.163572  ...  0.955101  0.459257 -0.098889 -0.087490 -0.099080   \n",
       "2 -0.061584  1.348765  ...  0.059948  0.113967 -0.098889  1.969527  0.037532   \n",
       "3 -0.061584 -0.160199  ...  0.241128  0.241406 -0.098889 -0.087490 -0.084674   \n",
       "4 -0.061584 -0.074885  ...  0.082065  0.114773 -0.098889  8.948005  1.024948   \n",
       "\n",
       "        164       165       166  txId  class  \n",
       "0  1.151607  1.519700  1.521399  1076      2  \n",
       "1 -0.122137 -0.379970 -0.379288  2534      0  \n",
       "2 -0.131010  0.006994  0.017772  3181      0  \n",
       "3 -0.140597  1.519700  1.521399  3321      2  \n",
       "4 -0.009570 -0.080708 -0.123601  3889      2  \n",
       "\n",
       "[5 rows x 169 columns]"
      ]
     },
     "execution_count": 9,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df_merge.head()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "9d8e428d-a83d-4ee2-83cc-c978d6aa902c",
   "metadata": {},
   "source": [
    "## Preparing edges"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "be290324-29dc-4c2c-ac56-ad86597d90a2",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Total amount of edges in DAG: torch.Size([2, 234355])\n"
     ]
    }
   ],
   "source": [
    "# mapping nodes to indices\n",
    "nodes = df_merge[0].values\n",
    "map_id = {j:i for i,j in enumerate(nodes)}\n",
    "\n",
    "# mapping edges to indices\n",
    "edges = df_edges.copy()\n",
    "edges.txId1 = edges.txId1.map(map_id)\n",
    "edges.txId2 = edges.txId2.map(map_id)\n",
    "edges = edges.astype(int)\n",
    "\n",
    "edge_index = np.array(edges.values).T\n",
    "edge_index = torch.tensor(edge_index, dtype=torch.long).contiguous()\n",
    "\n",
    "# weights for the edges are equal in case of model without attention\n",
    "weights = torch.tensor([1] * edge_index.shape[1] , dtype=torch.float32)\n",
    "\n",
    "print(\"Total amount of edges in DAG:\", edge_index.shape)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "eead6786-7d2e-4ae0-a5b9-b87200001d9c",
   "metadata": {},
   "source": [
    "## Preparing nodes"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "88255e48-cd74-421c-9f9a-2312b49f0b5d",
   "metadata": {},
   "outputs": [],
   "source": [
    "# maping node ids to corresponding indexes\n",
    "node_features = df_merge.drop(['txId'], axis=1).copy()\n",
    "node_features[0] = node_features[0].map(map_id)\n",
    "\n",
    "classified_idx = node_features['class'].loc[node_features['class']!=2].index\n",
    "unclassified_idx = node_features['class'].loc[node_features['class']==2].index\n",
    "\n",
    "# replace unkown class with 0, to avoid having 3 classes, this data/labels never used in training\n",
    "node_features['class'] = node_features['class'].replace(2, 0) \n",
    "\n",
    "labels = node_features['class'].values\n",
    "\n",
    "# drop indeces, class and temporal axes \n",
    "node_features = torch.tensor(np.array(node_features.drop([0, 'class', 1], axis=1).values, dtype=np.float32), dtype=torch.float32)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d9fa3926-a7ae-461a-89de-42254ceef1ef",
   "metadata": {},
   "source": [
    "## PyG Dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "d864b8f6-8fbd-405d-a207-b44fed20baeb",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Number of nodes: 203769\n",
      "Number of node features: 165\n",
      "Number of edges: 234355\n",
      "Number of edge features: 165\n",
      "Average node degree: 1.15\n",
      "Number of classes: 2\n",
      "Has isolated nodes: False\n",
      "Has self loops: False\n",
      "Is directed: True\n"
     ]
    }
   ],
   "source": [
    "# converting data to PyGeometric graph data format\n",
    "elliptic_dataset = Data(x = node_features, \n",
    "                        edge_index = edge_index, \n",
    "                        edge_attr = weights,\n",
    "                        y = torch.tensor(labels, dtype=torch.float32)) \n",
    "\n",
    "print(f'Number of nodes: {elliptic_dataset.num_nodes}')\n",
    "print(f'Number of node features: {elliptic_dataset.num_features}')\n",
    "print(f'Number of edges: {elliptic_dataset.num_edges}')\n",
    "print(f'Number of edge features: {elliptic_dataset.num_features}')\n",
    "print(f'Average node degree: {elliptic_dataset.num_edges / elliptic_dataset.num_nodes:.2f}')\n",
    "print(f'Number of classes: {len(np.unique(elliptic_dataset.y))}')\n",
    "print(f'Has isolated nodes: {elliptic_dataset.has_isolated_nodes()}')\n",
    "print(f'Has self loops: {elliptic_dataset.has_self_loops()}')\n",
    "print(f'Is directed: {elliptic_dataset.is_directed()}')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "ba6d243f-1fb2-4c88-8bd8-544167054d49",
   "metadata": {},
   "outputs": [],
   "source": [
    "y_train = labels[classified_idx]\n",
    "\n",
    "# spliting train set and validation set\n",
    "_, _, _, _, train_idx, valid_idx = \\\n",
    "    train_test_split(node_features[classified_idx], \n",
    "                     y_train, \n",
    "                     classified_idx, \n",
    "                     test_size=0.15, \n",
    "                     random_state=Config.seed, \n",
    "                     stratify=y_train)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "404f0000-70bd-4963-96a6-07d53523feaa",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train dataset size: 39579\n",
      "Validation dataset size: 6985\n",
      "Test dataset size: 157205\n"
     ]
    }
   ],
   "source": [
    "elliptic_dataset.train_idx = torch.tensor(train_idx, dtype=torch.long)\n",
    "elliptic_dataset.val_idx = torch.tensor(valid_idx, dtype=torch.long)\n",
    "elliptic_dataset.test_idx = torch.tensor(unclassified_idx, dtype=torch.long)\n",
    "\n",
    "print(\"Train dataset size:\", elliptic_dataset.train_idx.shape[0])\n",
    "print(\"Validation dataset size:\", elliptic_dataset.val_idx.shape[0])\n",
    "print(\"Test dataset size:\", elliptic_dataset.test_idx.shape[0])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "id": "00bb9a5a-e93f-422a-bc66-c2f28709a26c",
   "metadata": {},
   "outputs": [],
   "source": [
    "# G = to_networkx(data_train, to_undirected=True)\n",
    "# plt.figure(figsize=(18,18))\n",
    "# plt.axis('off')\n",
    "# nx.draw_networkx(G,\n",
    "#                 pos=nx.spring_layout(G, seed=0),\n",
    "#                 with_labels=False,\n",
    "#                 node_size=50,\n",
    "#                 node_color=data.y,\n",
    "#                 width=2,\n",
    "#                 edge_color=\"grey\"\n",
    "#                 )\n",
    "# plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "aa70deba-dc72-44d5-a4c8-c2201e59b240",
   "metadata": {},
   "source": [
    "## Models"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "id": "8093b715-99eb-492d-bc0b-fd8c3cf19d45",
   "metadata": {},
   "outputs": [],
   "source": [
    "class GCN(torch.nn.Module):\n",
    "    \"\"\"Graph Convolutional Network\"\"\"\n",
    "    def __init__(self, dim_in, dim_h, dim_out):\n",
    "        super(GCN, self).__init__()\n",
    "        \n",
    "        self.gcn1 = GCNConv(dim_in, dim_h)\n",
    "        self.gcn2 = GCNConv(dim_h, dim_out)\n",
    "        \n",
    "    def forward(self, x, edge_index):\n",
    "        h = self.gcn1(x, edge_index)\n",
    "        h = torch.relu(h)\n",
    "        h = F.dropout(h, p=0.6, training=self.training)\n",
    "        out = self.gcn2(h, edge_index)\n",
    "        return out\n",
    "    \n",
    "class GAT(torch.nn.Module):\n",
    "    \"\"\"Graph Attention Network\"\"\"\n",
    "    def __init__(self, dim_in, dim_h, dim_out, heads=8):\n",
    "        super(GAT, self).__init__()\n",
    "        self.gat1 = GATv2Conv(dim_in, dim_h, heads=heads, dropout=0.6)\n",
    "        self.gat2 = GATv2Conv(dim_h*heads, dim_out, concat=False, heads=1, dropout=0.6)\n",
    "\n",
    "    def forward(self, x, edge_index):\n",
    "        h = F.dropout(x, p=0.6, training=self.training)\n",
    "        h = self.gat1(h, edge_index)\n",
    "        h = F.elu(h)\n",
    "        h = F.dropout(h, p=0.6, training=self.training)\n",
    "        out = self.gat2(h, edge_index)\n",
    "        return out\n",
    "    \n",
    "def accuracy(y_pred, y_test, prediction_threshold=0.5):\n",
    "    y_pred_label = (torch.sigmoid(y_pred) > prediction_threshold).float()*1\n",
    "\n",
    "    correct_results_sum = (y_pred_label == y_test).sum().float()\n",
    "    acc = correct_results_sum/y_test.shape[0]\n",
    "\n",
    "    return acc"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "2169b21f-d0ae-481a-be01-73253825a5fc",
   "metadata": {
    "tags": []
   },
   "source": [
    "## Train/test helpers"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "id": "6fef8088-b372-45c5-9f13-f92593fe2693",
   "metadata": {},
   "outputs": [],
   "source": [
    "def train_evaluate(model, data, criterion, optimizer, *args):\n",
    "    num_epochs = args[0]\n",
    "    checkpoints_dir = args[1]\n",
    "    model_filename = args[2]\n",
    "\n",
    "    best_model_wts = copy.deepcopy(model.state_dict())\n",
    "\n",
    "    best_loss = 10e10\n",
    "\n",
    "    if not os.path.exists(checkpoints_dir):\n",
    "        os.makedirs(checkpoints_dir)\n",
    "\n",
    "    model.train()\n",
    "    for epoch in range(num_epochs+1):\n",
    "        # Training\n",
    "        optimizer.zero_grad()\n",
    "        out = model(data.x, data.edge_index)\n",
    "        loss = criterion(out[data.train_idx], data.y[data.train_idx].unsqueeze(1))\n",
    "        acc = accuracy(out[data.train_idx], data.y[data.train_idx].unsqueeze(1), prediction_threshold=0.5)\n",
    "        loss.backward()\n",
    "        optimizer.step()\n",
    "\n",
    "        # Validation\n",
    "        val_loss = criterion(out[data.val_idx], data.y[data.val_idx].unsqueeze(1))\n",
    "        val_acc = accuracy(out[data.val_idx], data.y[data.val_idx].unsqueeze(1), prediction_threshold=0.5)\n",
    "\n",
    "        if(epoch % 10 == 0):\n",
    "            print(f'Epoch {epoch:>3} | Train Loss: {loss:.3f} | Train Acc: '\n",
    "                  f'{acc*100:>6.2f}% | Val Loss: {val_loss:.2f} | '\n",
    "                  f'Val Acc: {val_acc*100:.2f}%')\n",
    "        \n",
    "            if val_loss < best_loss:\n",
    "                best_loss = val_loss\n",
    "                print(\"Saving model for best loss\")\n",
    "                checkpoint = {\n",
    "                    'state_dict': best_model_wts\n",
    "                }\n",
    "                torch.save(checkpoint, os.path.join(checkpoints_dir, model_filename))\n",
    "                best_model_wts = copy.deepcopy(model.state_dict())\n",
    "\n",
    "    return model\n",
    "\n",
    "def test(model, data):\n",
    "    model.eval()\n",
    "    out = model(data.x, data.edge_index) \n",
    "    preds = ((torch.sigmoid(out) > 0.5).float()*1).squeeze(1)\n",
    "    return preds"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "2d9be37b-cf67-429d-ae36-1c5e7439130b",
   "metadata": {},
   "source": [
    "## Train GCN"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "id": "cbcd33d4-4b19-4963-9058-0e9654b065f4",
   "metadata": {},
   "outputs": [],
   "source": [
    "gcn_model = GCN(Config.input_dim, Config.hidden_size, Config.output_dim).to(Config.device)\n",
    "data_train = elliptic_dataset.to(Config.device)\n",
    "\n",
    "optimizer = torch.optim.Adam(gcn_model.parameters(), lr=Config.learning_rate, weight_decay=Config.weight_decay)\n",
    "scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min')\n",
    "criterion = torch.nn.BCEWithLogitsLoss()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "id": "a1394af1-e3b0-4b46-9b30-e88b744d2757",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch   0 | Train Loss: 0.759 | Train Acc:  62.16% | Val Loss: 0.73 | Val Acc: 64.07%\n",
      "Saving model for best loss\n",
      "Epoch  10 | Train Loss: 0.307 | Train Acc:  86.43% | Val Loss: 0.30 | Val Acc: 87.16%\n",
      "Saving model for best loss\n",
      "Epoch  20 | Train Loss: 0.258 | Train Acc:  89.52% | Val Loss: 0.25 | Val Acc: 89.61%\n",
      "Saving model for best loss\n",
      "Epoch  30 | Train Loss: 0.244 | Train Acc:  90.49% | Val Loss: 0.24 | Val Acc: 90.32%\n",
      "Saving model for best loss\n",
      "Epoch  40 | Train Loss: 0.230 | Train Acc:  91.32% | Val Loss: 0.22 | Val Acc: 91.40%\n",
      "Saving model for best loss\n",
      "Epoch  50 | Train Loss: 0.219 | Train Acc:  91.85% | Val Loss: 0.22 | Val Acc: 91.77%\n",
      "Saving model for best loss\n",
      "Epoch  60 | Train Loss: 0.214 | Train Acc:  92.35% | Val Loss: 0.21 | Val Acc: 92.61%\n",
      "Saving model for best loss\n",
      "Epoch  70 | Train Loss: 0.210 | Train Acc:  92.60% | Val Loss: 0.21 | Val Acc: 92.80%\n",
      "Saving model for best loss\n",
      "Epoch  80 | Train Loss: 0.201 | Train Acc:  92.86% | Val Loss: 0.20 | Val Acc: 92.81%\n",
      "Saving model for best loss\n",
      "Epoch  90 | Train Loss: 0.195 | Train Acc:  93.15% | Val Loss: 0.20 | Val Acc: 92.81%\n",
      "Saving model for best loss\n",
      "Epoch 100 | Train Loss: 0.194 | Train Acc:  93.25% | Val Loss: 0.19 | Val Acc: 93.53%\n",
      "Saving model for best loss\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "GCN(\n",
       "  (gcn1): GCNConv(165, 128)\n",
       "  (gcn2): GCNConv(128, 1)\n",
       ")"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "train_evaluate(gcn_model,\n",
    "              data_train,\n",
    "              criterion,\n",
    "              optimizer,\n",
    "              Config.num_epochs,\n",
    "              Config.checkpoints_dir,\n",
    "              'gcn_best_model.pth.tar')"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0b6ae7ae-9f2f-44fc-816e-2a406edb8912",
   "metadata": {},
   "source": [
    "## Test GCN"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "id": "041569d5-b24e-491b-ab62-096e014b7bc5",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x432 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "              precision    recall  f1-score   support\n",
      "\n",
      "       licit       0.94      0.99      0.97      6303\n",
      "     illicit       0.85      0.45      0.59       682\n",
      "\n",
      "    accuracy                           0.94      6985\n",
      "   macro avg       0.90      0.72      0.78      6985\n",
      "weighted avg       0.93      0.94      0.93      6985\n",
      "\n",
      "Test data fraud cases, percentage: 6.191278903342769 %\n"
     ]
    }
   ],
   "source": [
    "gcn_model.load_state_dict(torch.load(os.path.join(Config.checkpoints_dir, 'gcn_best_model.pth.tar'))['state_dict'])\n",
    "\n",
    "y_test_preds = test(gcn_model, data_train)\n",
    "\n",
    "# confusion matrix on validation data\n",
    "conf_mat = confusion_matrix(data_train.y[data_train.val_idx].detach().cpu().numpy(), y_test_preds[valid_idx])\n",
    "\n",
    "plt.subplots(figsize=(6,6))\n",
    "sns.set(font_scale=1.4)\n",
    "sns.heatmap(conf_mat, annot=True, fmt=\".0f\", annot_kws={\"size\": 16}, cbar=False)\n",
    "plt.xlabel('Target (true) Class'); plt.ylabel('Output (predicted) class'); plt.title('Confusion Matrix')\n",
    "plt.show();\n",
    "\n",
    "print(classification_report(data_train.y[data_train.val_idx].detach().cpu().numpy(),\n",
    "                            y_test_preds[valid_idx],\n",
    "                            target_names=['licit', 'illicit']))\n",
    "\n",
    "\n",
    "print(f\"Test data fraud cases, percentage: {y_test_preds[data_train.test_idx].detach().cpu().numpy().sum() / len(data_train.y[data_train.test_idx]) *100} %\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "b9672996-7de9-4f4e-8746-7cc23422d41d",
   "metadata": {},
   "source": [
    "## Train GAT"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "id": "c45fd8a8-b28e-404a-a899-16090f87b6f4",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch   0 | Train Loss: 1.176 | Train Acc:  68.34% | Val Loss: 1.01 | Val Acc: 68.33%\n",
      "Saving model for best loss\n",
      "Epoch  10 | Train Loss: 0.509 | Train Acc:  88.63% | Val Loss: 0.48 | Val Acc: 88.70%\n",
      "Saving model for best loss\n",
      "Epoch  20 | Train Loss: 0.489 | Train Acc:  90.09% | Val Loss: 0.49 | Val Acc: 89.94%\n",
      "Epoch  30 | Train Loss: 0.465 | Train Acc:  89.87% | Val Loss: 0.48 | Val Acc: 89.76%\n",
      "Saving model for best loss\n",
      "Epoch  40 | Train Loss: 0.448 | Train Acc:  89.81% | Val Loss: 0.44 | Val Acc: 90.15%\n",
      "Saving model for best loss\n",
      "Epoch  50 | Train Loss: 0.445 | Train Acc:  90.04% | Val Loss: 0.44 | Val Acc: 89.89%\n",
      "Epoch  60 | Train Loss: 0.443 | Train Acc:  90.22% | Val Loss: 0.44 | Val Acc: 90.45%\n",
      "Epoch  70 | Train Loss: 0.439 | Train Acc:  90.38% | Val Loss: 0.43 | Val Acc: 90.16%\n",
      "Saving model for best loss\n",
      "Epoch  80 | Train Loss: 0.426 | Train Acc:  90.57% | Val Loss: 0.43 | Val Acc: 90.41%\n",
      "Saving model for best loss\n",
      "Epoch  90 | Train Loss: 0.423 | Train Acc:  90.72% | Val Loss: 0.42 | Val Acc: 90.38%\n",
      "Saving model for best loss\n",
      "Epoch 100 | Train Loss: 0.418 | Train Acc:  90.72% | Val Loss: 0.42 | Val Acc: 90.74%\n",
      "Saving model for best loss\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "GAT(\n",
       "  (gat1): GATv2Conv(165, 128, heads=8)\n",
       "  (gat2): GATv2Conv(1024, 1, heads=1)\n",
       ")"
      ]
     },
     "execution_count": 18,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "gat_model = GAT(Config.input_dim, Config.hidden_size, Config.output_dim).to(Config.device)\n",
    "data_train = elliptic_dataset.to(Config.device)\n",
    "\n",
    "optimizer = torch.optim.Adam(gat_model.parameters(), lr=Config.learning_rate, weight_decay=Config.weight_decay)\n",
    "scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min')\n",
    "criterion = torch.nn.BCEWithLogitsLoss()\n",
    "\n",
    "train_evaluate(gat_model,\n",
    "               data_train,\n",
    "               criterion,\n",
    "               optimizer,\n",
    "               Config.num_epochs,\n",
    "               Config.checkpoints_dir,\n",
    "               'gat_best_model.pth.tar')"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "4890a37b-6c09-43b7-9c15-99ea09756710",
   "metadata": {},
   "source": [
    "## Test GAT"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "id": "40b2fa67-0960-448c-b760-7e25a8ae10f2",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x432 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "              precision    recall  f1-score   support\n",
      "\n",
      "       licit       0.96      0.97      0.97      6303\n",
      "     illicit       0.68      0.67      0.68       682\n",
      "\n",
      "    accuracy                           0.94      6985\n",
      "   macro avg       0.82      0.82      0.82      6985\n",
      "weighted avg       0.94      0.94      0.94      6985\n",
      "\n",
      "Test data fraud cases, percentage: 18.133011036544637 %\n"
     ]
    }
   ],
   "source": [
    "gat_model.load_state_dict(torch.load(os.path.join(Config.checkpoints_dir, 'gat_best_model.pth.tar'))['state_dict'])\n",
    "\n",
    "y_test_preds = test(gat_model, data_train)\n",
    "\n",
    "# confusion matrix on validation data\n",
    "conf_mat = confusion_matrix(data_train.y[data_train.val_idx].detach().cpu().numpy(), y_test_preds[valid_idx])\n",
    "\n",
    "plt.subplots(figsize=(6,6))\n",
    "sns.set(font_scale=1.4)\n",
    "sns.heatmap(conf_mat, annot=True, fmt=\".0f\", annot_kws={\"size\": 16}, cbar=False)\n",
    "plt.xlabel('Target (true) Class'); plt.ylabel('Output (predicted) class'); plt.title('Confusion Matrix')\n",
    "plt.show();\n",
    "\n",
    "print(classification_report(data_train.y[data_train.val_idx].detach().cpu().numpy(),\n",
    "                            y_test_preds[valid_idx],\n",
    "                            target_names=['licit', 'illicit']))\n",
    "\n",
    "\n",
    "print(f\"Test data fraud cases, percentage: {y_test_preds[data_train.test_idx].detach().cpu().numpy().sum() / len(data_train.y[data_train.test_idx]) *100} %\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "055ba1fd-99a6-4eff-92be-f51db9693e9c",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "pytorch_metal",
   "language": "python",
   "name": "pytorch_metal"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.9.13"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}