Fraud detection is a critical challenge for businesses and organizations of all kinds. To ensure transactions are legitimate and protect against financial losses, business must be able to accurately identify fraudulent activity. But accuracy comes at a cost: if a system is too strict, it will flag too many legitimate transactions as fraudulent leading to unnecessary costs;  if it is too lenient, it will miss fraudulent activity and incur losses.

To address this challenge, machine learning algorithms (MLAs) are used to improve fraud detection. These algorithms can be trained to maximize the accuracy of fraud detection while minimizing the number of false positives. By optimizing the balance of precision and recall, machine learning algorithms can help organizations accurately identify fraudulent activity and protect financial resources.

‍

Here’s our methodology at Corgi Labs.

‍

Methodology

Core digital payment transaction data is captured from payment provider partners (Stripe, Adyen, Shopify, etc.) using a developer permissions account.

The ensemble machine learning approach consists of fraud prediction and proactive fraud prevention components.

For fraud prediction, we utilize a combination of RFE (recursive feature elimination) to identify effective clustering features, and optimized k-means clustering (k-means++ to generate robust centroids, modified inertia vs. cluster tradeoff calculation for enhance explainability) to identify transactions with high fraud probability. These methods allow for enhanced precision by isolating the target transactions cluster or downstream separation methods, while active counter-balancing for recall allows for status quo or better fraud prevention.

For proactive fraud prevention, we utilize tuned random forest to determine high correlation features for cluster separation, and a modified grid search to identify feature combination rules for proactive fraud prevention. These rules leverage existing or derived metadata features of the transactions, and are trivial to integrate with rule engines embedded in payment provider infrastructure. Leveraging existing infrastructure also enable computational load and direct cost reduction.

‍

Clustering Machine Learning Algorithm

Clustering machine learning algorithms are powerful tools for identifying high probability fraudulent transactions. By grouping similar transactions together, these algorithms can help organizations quickly identify fraudulent activity and take proactive measures to protect against losses.

‍

Hyperparameter Tuned Random Forest

Hyperparameter tuning can be used to further improve the accuracy of fraud detection by optimizing the random forest algorithm. By adjusting the parameters in the random forest, the algorithm can be tuned to identify the most important features for fraud prediction.

‍

Grid Search to Generate Proactive Prevention Rules

Finally, grid search can be used to generate proactive prevention rules for deployment on payment platforms. This allows organizations to take a proactive approach to fraud prevention and ensure that the most effective rules are in place to protect against losses.