{
  "preprocessors": [
    {
      /* if columns are present, sort lines by columns */
      "type": "multicolumn"
    },
    {
      /*  OCR all pages containing 5 or fewer lines of text. 
          This ensures that Sensible OCRs page 2, which contains 
          images of text and charts */
      "type": "ocr",
      "pageLinesLimit": 5
    },
  ],
  "fields": [
    {
      /* extract the introduction as a page range
        between the headings "introduction" and "methodology".
        Without the Multicolumn preprocessor, the Document Range method
        would incorrectly extract a top portion of the first page.*/
      "id": "introduction",
      "anchor": {
        "match": {
          "pattern": "introduction$",
          "type": "regex",
          "flags": "i"
        }
      },
      "method": {
        "id": "documentRange",
        "includeAnchor": true,
        "stop": {
          "pattern": "(?:\\d+\\.\\s*)?(methodology||related work)",
          "type": "regex",
          "flags": "i",
          "minimumHeight": 0.16
        }
      }
    },
    {
      "method": {
        /* extracts the methodology text using an LLM prompt.
          Since this is an excerpted paper, the LLM
          accurately reports
          that the answer is incomplete. */
        "id": "queryGroup",
        "searchBySummarization": true,
        "queries": [
          {
            "id": "methodology",
            "type": "string",
            "description": "What is the complete methodology?"
          }
        ]
      }
    },
    {
      /* chain prompts by asking an LLM to answer questions about the
         extracted methodology field */
      "method": {
        "id": "queryGroup",
        /* use the  */
        "source_ids": [
          /* restrict queries to data in methodology field */
          "methodology"
        ],
        "queries": [
          {
            "id": "data_set_size",
            "type": "number",
            "description": "What is the size of the data set?"
          },
          {
            "id": "data_set_type",
            "description": "Is the data set real or simulated? Answer 'real' or 'simulated'"
          }
        ]
      }
    },
  ]
}

{
  "introduction": {
    "type": "string",
    "value": "1 INTRODUCTION Credit cards make up an annual $100 Billion global market. As credit cards become more and more popular in the digital age, they also come with their inherent risks. Credit card fraud is the act of a fraudulent credit card transaction in order for criminals to acquire money or goods and services not belonging to them. Criminals committing credit card fraud are rarely apprehended and therefore, automatic credit card fraud detection can be of great significance. The task of automation credit card fraud classification is not easy, since the data-sets for credit card transactions are severely imbalanced. An imbalanced data-set is one where the proportion of one class is significantly higher than the rest of the classes. Classification tasks on imbalanced data-sets are difficult since it tends to be possible to gain a high level of accuracy despite misclassifying most of the data points from the minority class. A plausible way of attempting to classify on an imbalanced data-set is to apply class weights such that each data point from the minority class is given more importance and each data point from the majority class is given less importance. This study will attempt to see how class weights can help tackle the task of credit card fraud classification. This will be done by comparing Binary Logistic Regression model and Random Forest Classifier model without weights to those with class weights. This study will aim to find if a weighted classification will be able to improve upon the unweighted classification of credit card fraud detection. 2 METHODOLOGY 2.1 Data Credit Card transactions data tend to contain confidential information and therefore, they are scarcely available to the public. Hence, the data-set for this study is one created from a simulator, PaySim, that simulates credit card transactions. The simulator utilizes a private data-set to generate a simulated data-set that mimics ordinary credit card transactions along with fraudulent activity. The data-set contains 3,223,223 synthetic transactions. Table 1. Fields of the Data-set Field Data Type Description type of transaction STRING type amount of the transaction DOUBLE amount account balance before the transaction oldbalanceOrg DOUBLE account balance after the transaction newbalanceOrig DOUBLE account balance of recipient before the transaction. oldbalanceDest DOUBLE account balance of recipient after the transaction. newbalanceDest DOUBLE if the transaction is fraudulent or not isFraud Binary Table 2. Frequency and average amount of Valid and Fraudulent Transactions Number of Transactions Category Average Amount Valid $ 156,675.40 3,220,396 Fraudulent 2,827 $ 1,309,250.00 TRANSFER CASH IN TRANSFER CASH_OUT DEBIT PAYMENT CASH_OUT (b) Proportions of fraudulent (a) Proportions of transaction by type transaction by type Fig. 1. Breakdown of proportion of all and fraudulent transactions by type 2.1.1 Exploratory Data Analysis. From table 2 we can see the imbalanced nature of data-set and how on average fraudulent transactions amount is more than 8 times the amount of valid transactions. Figure 1 also demonstrates that how fraudulent transactions are only of type ʼTRANSFERʼ and ʼCASH OUTʼ. Finally, Figure 2 illustrates how the distribution of fraudulent transactions tend to have amount higher than those of valid transactions. 2.2 Pre-processing The data needs to be pre-processed because it contains categorical variables and Machine Learning algorithms use quantitative variables to discriminate between classes. The categorical variable in this data-set that will be used is type. The first transformation applied to the field type is the StringIndexer. The StringIndexer maps the string variable to indices belonging to the set 0, . . . , 𝑛𝑢𝑚𝑏𝑒𝑟𝑂 𝑓 𝑈 𝑛𝑖𝑞𝑢𝑒𝑉 𝑎𝑙𝑢𝑒𝑠 - 1. The order of the index assignment is ordered by frequency of that value in the data-set. Furthermore, the string-indexed variable is now transformed into a one-hot vector using the OneHotEncoder. This creates a vector of length 𝑛 − 1 (where 𝑛 is the number of unique categorical values) where all values are 0.0 except for at the index location which has a value of 1.0. Finally, all the feature variables are placed into one vector which represents the features for the respective transaction. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected]. © 2021 Association for Computing Machinery. Manuscript submitted to ACM"
  },
  "methodology": {
    "value": "2 METHODOLOGY\n2.1 Data\nCredit Card transactions data tend to contain confidential information and therefore, they are scarcely available to the public. Hence, the data-set for this study is one created from a simulator, PaySim, that simulates credit card transactions. The simulator utilizes a private data-set to generate a simulated data-set that mimics ordinary credit card transactions along with fraudulent activity. The data-set contains 3,223,223 synthetic transactions.\n2.1.1 Exploratory Data Analysis.\nFrom table 2 we can see the imbalanced nature of data-set and how on average fraudulent transactions amount is more than 8 times the amount of valid transactions. Figure 1 also demonstrates that how fraudulent transactions are only of type ʼTRANSFERʼ and ʼCASH OUTʼ. Finally, Figure 2 illustrates how the distribution of fraudulent transactions tend to have amount higher than those of valid transactions. 2.2 Pre-processing The data needs to be pre-processed because it contains categorical variables and Machine Learning algorithms use quantitative variables to discriminate between classes. The categorical variable in this data-set that will be used is type. The first transformation applied to the field type is the StringIndexer. The StringIndexer maps the string variable to indices belonging to the set 0, . . . , 𝑛𝑢𝑚𝑏𝑒𝑟𝑂 𝑓 𝑈 𝑛𝑖𝑞𝑢𝑒𝑉 𝑎𝑙𝑢𝑒𝑠 - 1. The order of the index assignment is ordered by frequency of that value in the data-set. Furthermore, the string-indexed variable is now transformed into a one-hot vector using the OneHotEncoder. This creates a vector of length 𝑛 - 1 (where 𝑛 is the number of unique categorical values) where all values are 0.0 except for at the index location which has a value of 1.0. Finally, all the feature variables are placed into one vector which represents the features for the respective transaction.",
    "type": "string",
    "confidenceSignal": "answer_may_be_incomplete"
  },
  "data_set_size": {
    "source": "3,223,223",
    "value": 3223223,
    "type": "number",
    "confidenceSignal": "not_supported"
  },
  "data_set_type": {
    "value": "simulated",
    "type": "string",
    "confidenceSignal": "not_supported"
  }
}