International laboratory Octapharma used to manually check each IBAN and bank details change request. After a potential fraud attempt, the company turned to Trustpair. By using the machine learning fraud detection solution, any future fraud attempts would be outlined. Also, the digital processes for checking bank details offer better collaboration with the purchasing department.
Read on to find out more about fraud detection models, machine learning, and how Trustpair’s algorithm can best help your business prevent vendor fraud… Request a demo now!
What is machine learning?
Machine learning is a type of artificial intelligence. It means that a system can learn and perform autonomously without clear instructions.
It uses historical data sets to learn patterns. Machine learning methods can range from recommendations of products to buy or films to watch all the way to detecting invoice fraud in a business.
There are four types of machine learning:
- Supervised learning – input data is labeled and classified as good or bad. The machine makes predictions based on information. However, it cannot detect new fraud that it didn’t learn
- Unsupervised learning – doesn’t need much data input or involvement from humans. The algorithm uses unstructured and unlabelled data to find patterns and regularities
- Semi-supervised learning – uses both labeled and unlabelled data
- Reinforcement learning – taught from its own experiences, always learning and adapting using feedback to move forward. The computer needs lots of data to do this
How does machine learning work in fraud detection?
There are several steps for how machine learning works in practice. Fraud detection software based on machine learning already does these for you.
- Insert data
- Determine fraud signals
- Prepare the algorithm
- Deploy the model
- Test and iterate
- Fraud risk score generated
Let’s go into more detail…
1. Insert data
By inputting data into the machine, it has a baseline to work from. When using supervised learning, the data has to be classed as ‘good’ (non-fraudulent customers) or ‘bad’ (fraudulent customers or those with chargeback claims) data.
2. Determine fraud signals
Fraud red flags will be used as triggers to indicate when fraudulent activity may be occurring.
Fraud signals and features can be split into categories and we have provided an example for each:
- Identification – such as the number of devices a customer has used
- Geographical location – where in the world the vendor is being paid from
- Order history – how many orders a customer made in the time after opening the account
- Network history – how many phone numbers are within the network
- Methods of payment – vendor name and billing name aligning
3. Prepare the algorithm
The algorithm can learn from a business’s own dataset to make predictions. The training period is useful because the system is not using live customer data, it is historical. So, any false positives in training won’t impact the customers.
4. Deploy the model
Following these steps, there is a model that is tailored to your business and ready to be put into practice.
5. Test and iterate
As part of the testing process, give the algorithm some data that is new to the model but that the business knows the outcomes of (fraudulent activity). Therefore, in the test the algorithm will either pick up on the fraud or fail to do so, and then we know whether it’s working. It will also give an idea of the performance and accuracy of the model.
6. Set the right fraud risk score threshold
Once the algorithm is in action and you have ensured its precision, fraud risk scores are created for every transaction. They are between 1 – 100 and the closer to 100, the higher the risk of fraud.
If it’s a low-risk score, the transaction should be approved. However, if the risk score is high, then the purchase should be reviewed by the customer or the bank to ensure there is no fraud.
It’s often on your business to set these risk scores accordingly to find the right balance. If it’s too low, there may be lots of false negatives (fraud that goes unnoticed). However, if it’s too high there can be lots of false positives (non-fraud transactions that are blocked) as this impacts customer satisfaction.
It is worth adding that instead of a risk score, Trustpair’s process offers a risk status that can be positive, negative, or unconfirmed. Some software solutions don’t provide scores as it can be difficult for companies to know what to do with them.
Fraud detection models vs traditional fraud detection measures
Traditional fraud detection measures were only able to detect fraud that it has been programmed to detect.
For example, if a trigger is set to check over five digit vendor payments, if that transaction is made, the trigger will go off.
However, the traditional measures cannot recognize fraud that it hasn’t been programmed to. Also, it cannot adapt to new methods of committing fraud.
On the other hand, machine learning fraud detection models can learn from their experiences and therefore challenge new fraud methods. They are flexible and adaptable to new patterns of fraud.
Machine learning systems are also more cost-effective because it is automated and often doesn’t need manual input or reviews from fraud analysts. Their time can be spent elsewhere.
Examples of machine learning for fraud detection
Credit card fraud
Machine learning based algorithms can outline credit card fraud attempts via fraud signals. Two of the fraud signals may be triggered if a fraudster makes several high-value financial transactions on a new account. Also, the customer and billing name may not line up.
Invoice fraud
The machine learning algorithm can inspect invoices and pick up on duplicate invoices or a credit card bank account number of a supplier that has been involved in suspicious transactions and it can flag it up. Learn more about invoice fraud here!
Internal fraud
The models can recognize patterns internally within the accounts payable team. Say that an employee within your business is approving an unusually high number of invoices or approving invoices to a new, unverified supplier, that could be a method of internal fraud.
That is because the unverified supplier could be a fake vendor with account information that aligns with the employee’s own account details.
Account takeover
You may have logged in to your work email on a different computer or laptop in a different location to where you normally do and had to confirm your identity via a code sent to your mobile phone.
It is machine learning algorithms that does this and helps in the case of an account takeover where someone does have your work email password and is trying to log in to your account from another country.
Identity verification
It is important that your business is able to check to see that your suppliers are who they say they are, that they are legitimate, and that they are not fraudsters.
A machine learning model can help identify this. A platform like Trustpair uses machine learning to detect any suspicious data or status change that raises fraud red flags.
For example, Trustpair checks that the identity of the vendor corresponds to its bank account details thanks to its machine learning based algorithm.
Recap
A machine learning system can be incredibly useful in a business’s efforts to combat fraud. Now you know the steps to take to set up a machine learning model and how it can help with identity verification, preventing invoice fraud, internal fraud, and account takeovers. Fraud detection model Trustpair uses machine learning to spot any dubious activity or status change to raise an alert.
