Transaction Review System Using Decision Tree for Fraud Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for fraud detection in online transactions are inefficient, leading to labor-intensive manual reviews, incorrect flagging of legitimate orders as fraudulent, and missed fraudulent orders, resulting in increased costs and potential losses for merchants.
Innovation Solution
A data processing method using a hierarchical decision tree based on transaction data and historical reviews to determine the likelihood of a transaction being accepted or rejected, reducing the need for manual review by providing a likelihood value to aid reviewers in making decisions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If automated screening is used to identify fraudulent orders, then fraud detection capability is improved, but manual review workload increases
Solution Approach 1:
The system creates a digital twin or replica of the human reviewer's decision-making process through machine learning models trained on historical review data. This digital copy can automatically evaluate transactions and predict reviewer decisions, thereby reducing the need for actual manual reviews while maintaining consistent fraud detection standards.
Solution Approach 2:
The patent introduces an automated screening system as an intermediary between the transaction flow and manual review. This intermediary layer pre-evaluates transactions, filters out clearly fraudulent or clearly legitimate orders, and only presents ambiguous cases to human reviewers, thereby significantly reducing manual review workload.
2Reliability
If more orders are sent for manual review to ensure fraud detection accuracy, then fraud detection reliability is improved, but labor costs increase
Solution Approach 1:
The system applies partial automation by using automated screening for all transactions but relying on it partially - only human reviewers handle cases where the automated system is uncertain or where the stakes are highest. This partial action approach maintains high accuracy while avoiding the excessive cost of having humans review every single transaction.
Solution Approach 2:
The patent dynamically adjusts the threshold for sending orders to manual review based on various parameters such as transaction amount, customer history, and detected fraud patterns. By changing these parameters, the system optimizes the balance between detection accuracy and labor cost, sending fewer orders for manual review when confidence is high and more when uncertainty is high.
3Measurement precision
If manual review is performed on all orders, then detection precision is improved, but processing time increases
Solution Approach 1:
The patent segments the transaction review process into multiple stages: automated preliminary screening, risk scoring, and selective manual review. This segmentation allows the system to apply different levels of scrutiny to different transactions based on their risk profile, maintaining high detection precision for fraudulent orders while quickly processing low-risk transactions without manual intervention.
Solution Approach 2:
The system performs preliminary automated screening and risk assessment before manual review is considered. This preliminary action filters out many clearly legitimate or clearly fraudulent transactions, so that when manual review is needed, the reviewers are already working with pre-analyzed cases, reducing their processing time while maintaining detection precision.
Data Source
AI summary
In an embodiment, a data processing method comprises obtaining a plurality of first transaction data items for a proposed online credit card purchase transaction that has been recommended for review; obtaining a plurality of second transaction data items for a set of similar past online credit card purchase transactions, wherein each member of the set has one or more transaction feature values that are similar to the transaction data items of the proposed online credit card purchase transaction, and a decision value specifying whether the member was accepted or rejected by a reviewer; obtaining a stored data model of features, feature values, transaction acceptance decisions and rejection decisions of the reviewer based at least in part on the set, determining, based on applying the first transaction data items to the stored data model and a subsequent query to the database among more recent transactions that were not included during model construction, a likelihood value of a particular decision of whether the proposed online credit card purchase transaction would be accepted or rejected by the reviewer of the merchant; causing the likelihood value to be displayed; wherein the method is performed by one or more computing devices.