AI Fraud Detection System Reducing False Positives
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing fraud detection systems in retail rely heavily on rule-based filters, which can generate excessive false positives, fail to detect new types of suspicious activity, and require manual re-tuning, making them inefficient and inflexible.
Innovation Solution
An AI-based system that combines rule-based systems with unsupervised machine learning models, specifically 1-class Support Vector Machines (SVM) and random forests, to identify and score anomalies in customer and cashier activities, allowing for adaptive fraud detection without human intervention.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If rule-based filters are used to detect fraud, then known types of suspicious activity can be identified, but the system generates excessive false positives and flags too many risk targets for investigation
Solution Approach 1:
The patent replaces the mechanical rule-based filtering system with an AI-based unsupervised machine learning system. Instead of using predefined rules that generate excessive false positives, the system uses anomaly detection algorithms (such as isolation forests, one-class SVMs, or autoencoders) to learn normal behavior patterns from historical data and automatically identify true anomalies, significantly reducing the number of false positive risk targets while maintaining high detection accuracy
Solution Approach 2:
The system changes the detection parameters from static rule thresholds to dynamic anomaly scores generated by machine learning models. These models continuously learn from new data and adapt their detection parameters automatically, allowing the system to distinguish between normal variations and true fraudulent behavior, thereby reducing unnecessary investigations while catching actual fraud
2Reliability
If rule-based systems are used for fraud detection, then existing fraud patterns can be detected, but new types of suspicious activity are missed
Solution Approach 1:
The patent implements a dynamic system where the fraud detection model continuously learns and adapts to new fraud patterns. Unlike static rules that must be manually updated, the unsupervised machine learning system automatically adjusts its understanding of normal behavior and detects emerging anomalies, enabling it to identify new types of fraud as they appear without requiring manual reconfiguration
Solution Approach 2:
The system performs self-learning and self-adaptation through unsupervised machine learning algorithms that automatically identify patterns in data without human intervention. The model continuously trains on new transaction data, automatically updating its internal representations of normal behavior and detecting deviations, thereby enabling autonomous detection of both known and novel fraud types without requiring manual rule updates
3Ease of manufacture
If rule-based filters are deployed, then fraud detection can be implemented, but manual re-tuning is required when business or environment changes
Solution Approach 1:
The system implements self-service through automated machine learning pipelines that continuously train, evaluate, and deploy updated detection models without human intervention. The system automatically adapts to business and environmental changes by learning from new data patterns, eliminating the need for manual re-tuning while maintaining high detection performance across changing conditions
Data Source
AI summary
Embodiments detect fraud of risk targets that include both customer accounts and cashiers. Embodiments receive historical point of sale (“POS”) data and divide the POS data into store groupings. Embodiments create a first aggregation of the POS data corresponding to the customer accounts and a second aggregation of the POS data corresponding to the cashiers. Embodiments calculate first features corresponding to the customer accounts and second features corresponding to the cashiers. Embodiments filter the risk targets based on rules and separate the filtered risk targets into a plurality of data ranges. For each combination of store groupings and data ranges, embodiments train an unsupervised machine learning model. Embodiments then apply the unsupervised machine learning models after the training to generate first anomaly scores for each of the customer accounts and cashiers.