Fraud Detection Model Using Dynamic Time Segments
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing fraud detection systems in transaction processing systems often fail to accurately distinguish between legitimate and fraudulent transactions, especially for parties with limited transaction history, leading to false positives and false negatives, due to their reliance on party-specific training data and fixed time periods that neglect timing relationships between transactions.
Innovation Solution
The method involves organizing transaction data into variable-length time segments, generating ragged tensors, and using max pooling to reduce these into representative vectors, which are then used to train a fraud detection model that considers timing relationships and historical activity patterns, including those from distant past transactions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If party-specific training data is used to train fraud detection models, then prediction accuracy for that specific party is improved, but prediction accuracy for other parties deteriorates
Solution Approach 1:
The patent segments transaction data by dividing it into multiple time periods (e.g., recent transactions vs. historical transactions) and trains separate fraud detection models for each segment. This allows the system to capture different fraud patterns at different time scales while maintaining adaptability across multiple parties through the unified multi-period modeling framework.
2Device complexity
If transaction data from fixed time periods is used for training, then model training is simplified, but timing relationships between transactions are lost
Solution Approach 1:
The patent employs dynamic time period segmentation where transaction data is divided into multiple variable time periods (e.g., recent 7 days, recent 30 days, historical data) rather than fixed uniform intervals. This dynamic segmentation preserves the temporal relationships and timing patterns between transactions while organizing data in a structured manner that facilitates model training.
3Device complexity
If all historical transactions are considered equally, then data processing is simplified, but the influence of distant past transactions cannot be properly captured
Solution Approach 1:
The patent applies local quality by assigning different weights or importance levels to different time periods. Recent transactions are given higher weight as they are more indicative of current fraud patterns, while historical transactions provide contextual baseline information. This differential weighting ensures that distant past transactions contribute appropriately to the model without dominating recent fraud indicators.
Data Source
AI summary
Certain aspects of the present disclosure provide techniques for detecting fraudulent transactions in a transaction processing system. An example method generally includes receiving a request to process a transaction. An input data set including a vector representing the transaction and a plurality of vectors representing historical transactions is generated. The input data set is divided into a plurality of ragged tensors corresponding to non-overlapping time segments of variable length and having a plurality of vectors associated with dates within each time segment A reduced input data set is generated by generating, for each respective ragged tensor of the plurality of ragged tensors, a respective representative vector using max pooling over vectors in the ragged tensor. A fraudulent transaction score is generated based on the reduced input data set using a fraud detection model. The transaction is processed based, at least in part, on the fraudulent transaction score.