ML Fraud Detection via Contextual Risk Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current loss prevention solutions in retail environments fail to distinguish between human errors and fraudulent activities, leading to a high number of false positives and overlooking actual fraudulent transactions, as they treat all abnormal actions with the same severity.
Innovation Solution
A system and method that utilize Machine-Learning Models to analyze transaction events, calculate the probability of target actions within the context of the entire transaction, and provide a risk score to differentiate between risky and legitimate actions, incorporating features such as payment types, repetitive items, and item weights.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If simplistic rules are used to detect fraud by treating all abnormal actions with the same severity, then the detection system is easy to implement, but the measurement precision of fraud detection deteriorates due to high false positives
Solution Approach 1:
The patent changes the parameter of risk scoring from uniform treatment to differentiated scoring based on action type, item category, and transaction context. By introducing multiple scoring dimensions (base risk score modified by action-specific factors, item-specific factors, and transaction-specific factors), the system achieves precise fraud detection while maintaining implementation through automated parameter application.
Solution Approach 2:
The patent applies local quality by treating different abnormal actions differently based on their specific characteristics. Instead of uniform treatment, the system assigns different risk weights to voiding actions versus price overrides versus returns, and further modulates these based on item categories (e.g., electronics vs. produce) and transaction contexts, achieving precise local differentiation.
2Reliability
If the risk score is elevated with the increase in counter value or amount when abnormal actions are detected, then the detection sensitivity is improved, but the false positive rate increases marking legitimate actions as suspicious
Solution Approach 1:
The patent modifies the risk score calculation by introducing contextual parameters that adjust the impact of abnormal actions. The base risk score is modified by action-specific factors (e.g., voiding 1 item vs. 10 items), item-specific factors (e.g., expensive items vs. inexpensive items), and transaction-specific factors, creating a nuanced scoring system that reduces false positives while maintaining sensitivity.
Solution Approach 2:
The patent applies local quality differentially by considering the specific context of each abnormal action. The system evaluates whether voiding 10 items is more suspicious than voiding 2 items by considering item categories, transaction patterns, and customer behavior history, thereby reducing false positives while maintaining detection sensitivity for actual fraud.
3Reliability
If existing solutions audit transactions based on abnormal actions, then the fraud detection coverage is increased, but the productivity of legitimate transactions deteriorates due to time-consuming audits of false positives
Solution Approach 1:
The patent changes the audit parameter from binary (audit or not) to probabilistic (risk score threshold). By calculating a nuanced risk score that incorporates multiple factors, the system can confidently identify low-risk transactions that should not be audited, while maintaining comprehensive coverage for high-risk transactions, thereby improving productivity without sacrificing detection coverage.
Solution Approach 2:
The system incorporates feedback mechanisms where the risk score calculation continuously refines its assessment based on transaction patterns, customer behavior history, and outcome data. This feedback loop enables the system to learn from past decisions and improve its accuracy over time, reducing false positives and optimizing audit selection to maintain both coverage and efficiency.
Data Source
AI summary
Risky actions versus non-risky actions in a transaction are identified and a fraud score associated with probabilities of the risky actions is updated accordingly for purposes of determining whether the transaction is likely or not likely to be associated with fraud. A machine-learning model is trained to predict the risky actions versus non-risky actions of a transaction based on the transaction features as a whole and compare the predicted action labels of risky and non-risky versus the actual actions taken in the transaction to calculate probabilities of risky actions taken and output a risk or fraud score based thereon. Higher probabilities correlate with lower risk scores and vice versus.