Acquirer-Level Fraud Detection Model Using Event-Aware Loss
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional fraud detection systems in payment transactions are primarily issuer-specific and focus on historical data of cardholders, failing to effectively predict fraudulent transactions at the acquirer level, leading to increased monetary losses and reputational damage.
Innovation Solution
A computer-implemented method that accesses historical transaction data to determine acquirer and transaction features, generates latent representations, and trains fraud and acquirer classifiers using a multi-component event-aware loss function, which incorporates domain-specific knowledge to optimize fraud predictions and classify transactions in real-time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If conventional issuer-specific fraud detection systems are used, then historical cardholder data can be analyzed, but acquirer-level fraud prediction capability is lost
Solution Approach 1:
The system creates a universal fraud detection model that serves both issuer-specific and acquirer-level needs. The acquirer-server specific model is trained to handle multiple functions: detecting fraud patterns specific to individual acquirers, analyzing merchant risk profiles, and providing predictions that can be applied across different issuers. This multi-functional approach recovers the lost acquirer-level prediction capability while maintaining broad applicability.
Solution Approach 2:
The fraud detection system is segmented into distinct components: issuer-specific models, acquirer-server specific models, and merchant-specific models. Each segment operates independently with specialized training data and features, allowing the system to capture acquirer-level patterns without requiring complete system redesign. The segmented architecture enables modular deployment and maintains information capability across different levels.
2Reliability
If real-time fraud detection is implemented, then fraudulent transactions can be blocked immediately, but false alarms increase
Solution Approach 1:
The system applies local quality by creating acquirer-server specific models that are tailored to the unique characteristics, risk profiles, and transaction patterns of individual acquirers and their associated merchants. Instead of applying a uniform detection threshold or model across all transactions, the system adapts the detection criteria locally to each acquirer-merchant context, improving reliability while reducing false alarms through context-aware prediction.
Solution Approach 2:
The system incorporates feedback mechanisms where prediction outcomes are continuously monitored and used to refine the acquirer-server specific models. By analyzing actual fraud cases and false alarms, the system adjusts its prediction algorithms and thresholds, progressively improving detection accuracy while learning to distinguish true fraud from legitimate transactions that might trigger false alarms.
3Reliability
If comprehensive historical data analysis is performed, then fraud patterns can be identified, but real-time inference speed decreases
Solution Approach 1:
The system performs preliminary action by pre-training acquirer-server specific models offline using comprehensive historical transaction data and fraud patterns. During real-time operation, the pre-trained models can quickly infer fraud risk without needing to re-analyze the entire historical dataset for each transaction. This preliminary processing of patterns and features enables both comprehensive pattern recognition and fast real-time inference.
Solution Approach 2:
The system extracts essential fraud patterns and features from comprehensive historical data during the offline training phase, separating the complex pattern recognition task from real-time transaction processing. By extracting and storing the learned patterns in the pre-trained model, the system retains the ability to detect sophisticated fraud patterns while achieving fast inference speeds during actual transaction processing.
Data Source
AI summary
Embodiments provide methods and systems for training a transaction monitoring model based on a multi-component event-aware loss function. The method performed by a server system includes accessing historical transaction data of payment transactions associated with an acquirer server. Method includes determining acquirer features associated with the acquirer server and transaction features associated with an individual payment transaction based on the historical transaction data. Method includes generating, via an embedding layer, a latent representation corresponding to the individual payment transaction. Method includes training a fraud classifier and an acquirer classifier based on the latent representation and the multi-component event-aware loss function. Method includes computing the multi-component event-aware loss function based on execution of the fraud classifier and the acquirer classifier. Moreover, method includes updating network parameters of the fraud classifier, the acquirer classifier, and the embedding layer based on the multi-component event-aware loss function.