Fraud Detection using No-Code Machine Learning Models: Benefits, Source and Data Sufficiency

Fraud is a criminal behavior with the intention of making a gain (usually financial) at the expense of others. In today’s world, fraud is not just something people hear about; it is everywhere. In daily life, most of us receive phishing phone calls and emails regularly, and credit card fraud is not a new phenomenon either. In business, fraud can be internal (e.g., a purchasing officer makes a personal gain for signing a contract with a particular vendor) or external (e.g., an insurance policyholder submits a fraudulent claim). It would be resource-intensive to monitor every employee or check every business transaction to identify fraud. How can a business identify fraud efficiently using a machine learning model?

Read this article to understand

Benefits of Machine Learning in Fraud Detection

In the past, fraud detection was typically performed by experienced personnel, as it requires a good understanding of what is normal and what is unusual. Often, a simple business rule is not sufficient to identify fraud. There could be somewhat odd combinations of factors, or the abnormality becomes more apparent when compared with peers. For example, a credit cardholder had a transaction in Paris and another transaction in New York within three hours. Individually, each transaction looked fine, but when the location factor is taken into account, there appears to be some fraudulent activity. Another example is in disability insurance, where claimants for a health care provider (say, a physiotherapist) on average have 50% more visits than other providers. This abnormality may suggest unnecessary health care services.

As there could be many different stakeholders and numerous business transactions, detecting fraud manually would be very resource-intensive. A machine learning model can evaluate various factors and assign a probability of fraud for each employee, vendor, claimant, or business transaction. A predictive fraud scoring model using machine learning can allocate resources more efficiently and be more effective in detecting fraud.

By leveraging insights drawn from a good model, management can prioritize resources and have fraud investigators focus on the high-potential group of fraudulent behavior. Fraud investigators may target specific groups of high-potential individuals or transactions through more specific and detailed investigation strategies. This targeted investigation strategy improves the effectiveness of detecting fraud and thus reduces financial losses.

Source and Examples of Frauds

Before creating a model to detect fraud, it is important to first understand the source of fraud. This depends on the industry.

Example 1 – Insurance Industry

Various common types of fraud are presented below. 

(ⅰ) Underwriting Frauds

Potential or existing policyholders may misrepresent some information to lower their risk profile and hence pay a lower premium. For example, a person does not disclose that he is a smoker when purchasing health insurance so that he can be considered low risk for lung cancer.

(ⅱ) Claim Frauds

Existing policyholders may misrepresent some information when they file a claim in order to increase their claim payment. For example, a claimant might overstate the loss from a house robbery when filing a home insurance claim.

(ⅲ) Vendor Frauds

Many insurance companies have various types of vendors. For example, in the workers’ compensation line, health care providers such as physiotherapists could be vendors providing treatments for injured workers. A physiotherapist could design a treatment plan with excessive visits in order to earn more provider fees.

Vendor frauds can also be found in other industries. Basically, any company that outsources some of its services to a vendor or hires a vendor to perform certain functions is exposed to vendor fraud risk. This could be building developers outsourcing work to contractors in the construction industry, or suppliers providing raw materials to producers in the manufacturing industry.

(ⅳ) Internal Frauds

Internal frauds are often in the form of personal gain at the expense of the company. For example, a procurement officer could contract to a particular information technology partner to receive a kickback.

Internal frauds can also be found in other industries. In fact, internal frauds are largely related to people risk. Therefore, as long as a company has employees, there is a probability of internal fraud. In the earlier example, any company in any industry with a procurement department could be subject to internal fraud risk.

Example 2 – Banking and Credit Card Industry

In the banking and credit card industry, other than vendor frauds and internal frauds, the following are very common frauds specific to this industry.

(ⅰ) Credit Card Frauds

Credit card frauds are very common, especially for online transactions. This normally happens when credit card information is lost due to computer hacking, phishing emails or text messages, stolen credit cards, etc.

Usually, fraudulent transactions exhibit unusual purchasing behaviors such as buying a product with a large amount in a foreign country while all other transactions around the same time are within the home country. A machine learning model could be built to identify potential fraudulent transactions. After filtering out high-potential fraudulent transactions, further investigation can be done. This would increase the success rate of identifying fraud. Read more about credit card fraud

(ⅱ) Identity Frauds

As described by the name, identity frauds involve the use of other people’s identities to access their accounts or create new accounts with stolen identity information. The purposes of identity frauds include account takeover, money laundering, unauthorized money transfer, etc.

Like credit card frauds, identity fraudulent transactions would also exhibit unusual behavior. For example, large and frequent money transfers, unusual transaction timing (e.g., after midnight), etc.

(ⅲ) Mortgage Application Fraud

Mortgage fraud involves the misrepresentation of information in mortgage applications. The outcome is often the default of mortgage payments. This normally involves an overstatement of assets and income, and an understatement of liabilities.

How a Fraud Investigator Detect Possible Frauds Using No-Code Machine Learning Model

1. Problem Framing

When using Machine Learning, the technical details are not the initial concerns. Instead, the most crucial aspects are to:

• Understand what kind of frauds you are trying to detect
• Gather historical data related to the type of frauds in question (e.g. to detect claim frauds, all the claim data and historical claim frauds in formation would be needed)

2. Dataset Size & Quality

Relevance of Predictors:

• Data used for predicting probabilities should be relevant and helpful in making accurate predictions.
• Business understanding and intuition play a role in judging the relevance of predictors.
• Some predictors may not be obvious initially but including them in the modelling process can confirm their usefulness after running the model.
• Feature engineering, such as transforming data, can lead to valuable predictions. For instance, while “date of birth” may not directly predict outcomes, deriving “age” from it can be useful in modelling.

Data Cleanliness:

The dataset should be clean, with minimal data issues (e.g., typos, inconsistent data types, missing data).

Data Quantity:

More data in the modeling process can lead to a model with better predictive power and more valuable business insights. Also, as a rule of thumb, ensure that the number of training data in your dataset to be at least ten times more than the number of trainable parameters in your model. This helps prevent overfitting and allows the model to learn robust patterns.

Model Complexity:

While complex models may seem appealing, simple models often perform better on large datasets. These simpler models generalize well and avoid unnecessary overfitting. In contrast, fancy models with intricate architectures may struggle when data is limited.

Remember, striking the right balance between data size and model complexity is crucial for achieving optimal performance. For more understanding on data construction, some beginner users may find the free Google course helpful.

3. User-friendly Tool

Traditionally, machine learning modeling has been considered a skillset exclusive to data scientists. However, nowadays, no-code applications empower fraud investigator to follow point-and-click steps, quickly build machine learning models, and assess their efficiency using Excel spreadsheets. Fraud detection strategy can now be designed more conveniently by leveraging machine learning.

While various tools are available in the market for building a machine learning model, one can utilize an easy-to-use Excel add-in called ‘Aidy’ for initial research. Powered by Google Cloud, Aidy enables rapid and reliable modelling for machine learning projects. You can obtain the free starter pack of Aidy and follow the case study to test it out without any coding requirements and use the function of data feature engineering and data cleanliness. Read more about how insurers detect and prevent claim fraud using machine learning tool.

Insufficient Fraud Data – Problems and Remedies

One common problem with fraud data is the number of frauds in historical data. Fraud is not typical behavior and therefore historical instances are usually limited. As such, even if the overall data quantity is sufficient (i.e., a lot of data records), the dataset could be highly unbalanced. That is, there could be a disproportionate number of fraudulent records versus non-fraudulent records. This could largely affect the performance of the resulting model. Take an extreme example: if 99% of the records are non-fraudulent and only 1% are fraudulent, there would be a tendency to have a model with a prediction result of non-fraudulent most of the time, if not all the time. The reason is that a model always predicting non-fraudulent would be about 99% accurate. However, such a model would not be able to predict any frauds and thus would have no predictive power in detecting fraud. A simple rule of thumb is to have at least 10%-20% of data points represented by fraudulent records in the training dataset.

To address this issue, the following are common ways to bring balance to the dataset used for modeling purposes:

1. Undersampling

Simplistically speaking, undersampling is to randomly remove non-fraudulent data records to make the dataset more “balanced.” For example, let’s say the raw dataset has 100,000 records where only 5,000 (or 5%) of them are fraudulent records. To bring this dataset to be more balanced (i.e., 20% fraudulent), 75,000 of the 95,000 non-fraudulent records would be randomly removed. That is, the resulting dataset for modeling purposes would have 20,000 non-fraudulent records (i.e., 80%) and 5,000 fraudulent records (i.e., 20%).

An obvious disadvantage of this method is that the data size could be significantly reduced. When the dataset is highly unbalanced, after undersampling, the resulting dataset for modeling might not have sufficient data points to produce a robust predictive model.

2. Oversampling

The idea of oversampling is to increase the number of fraudulent records by “creating” some. One way to do this is by adding slightly modified copies of existing fraudulent records or introducing new synthetic fraudulent data. These additional data points should be constructed carefully so that they truly represent fraudulent behavior if they were real.

For a highly unbalanced dataset, e.g., only 1% fraudulent, this method would be difficult to execute and subject to challenge as one would need to create a lot more new and non-existing fraudulent records than real fraudulent data points. Essentially, the resulting model for detecting fraud would be built mostly on non-real fraudulent data.

3. Hybrid

This method is a combination of undersampling and oversampling. It can strike a balance between removing too many data records and adding too many non-existing data points.

Conclusion

This blog article presents various types of fraud and connects machine learning modeling to fraud detection.

A machine learning model can help create a more efficient and effective fraud detection strategy, as it ranks the likelihood of fraudulent records and provides the probability of fraud for each record. This reduces the administrative burden on the fraud detection team, as more likely fraudulent records are identified and investigated first.