Cash-Return Recognition Model Using Dual DQN Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for recognizing cash-return transactions, such as malicious or fraudulent returns, are inefficient and do not effectively address the risk of harming credit histories and increasing service charges, leading to poor disbursement rates and overdue payments.
Innovation Solution
A cash-return recognition model using two Deep Q-Networks (DQNs) is trained, where the first DQN predicts cash-return transactions based on transaction information and adjusts the second DQN's parameters to minimize the difference between predicted and actual Q-values, employing reinforcement learning and Q-learning algorithms to improve recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning, semi-supervised learning, or unsupervised learning methods are used for cash-return transaction recognition, then the system can process transactions, but the recognition efficiency and accuracy are insufficient
Solution Approach 1:
The patent divides the cash-return recognition task into two separate DQN models: the first DQN performs initial cash-return prediction based on transaction information, and the second DQN performs Q-value prediction based on the first DQN's output. This segmentation allows each model to specialize in a specific aspect, improving both recognition accuracy and processing efficiency compared to using a single general-purpose learning method.
Solution Approach 2:
The first DQN acts as an intermediary between the transaction information and the second DQN. It transforms raw transaction data into cash-return predictive values, which then serve as input for the second DQN's Q-value prediction. This intermediary layer enables the system to process complex transactions more efficiently while maintaining high accuracy.
2Reliability
If traditional recognition methods are used, then the system can operate, but it fails to provide sufficient caution for large-amount transactions leading to poor disbursement rates
Solution Approach 1:
The patent applies different processing qualities to different transaction types by using the first DQN to identify cash-return transactions and then applying enhanced Q-value prediction through the second DQN specifically to these identified transactions. This local quality approach ensures that large-amount and suspicious transactions receive more rigorous analysis, improving disbursement reliability without unnecessarily complicating the processing of normal transactions.
Solution Approach 2:
The system changes parameters dynamically by adjusting the depth of analysis based on transaction characteristics. For transactions identified as potential cash-return cases by the first DQN, the system activates the second DQN for additional Q-value prediction, effectively changing the processing parameters to provide more caution and scrutiny where needed, thereby improving overall reliability.
3Measurement precision
If comprehensive transaction analysis is performed to improve recognition accuracy, then more transactions can be identified, but time and resource overheads increase
Solution Approach 1:
The first DQN performs preliminary action by conducting an initial cash-return prediction on all transactions before they undergo further analysis. This preliminary filtering identifies potential cash-return transactions, allowing the system to then apply more time-consuming and resource-intensive second DQN analysis only to these identified cases, rather than performing comprehensive analysis on all transactions, thus reducing overall time and resource overheads.
Solution Approach 2:
The system applies partial action by using the second DQN for Q-value prediction only on transactions that the first DQN identifies as potential cash-return cases, rather than applying comprehensive analysis to all transactions. This selective approach maintains high identification accuracy for suspicious transactions while minimizing the time and resource overhead associated with processing large volumes of normal transactions.
Data Source
AI summary
Embodiments of the application provide a method and device for training a cash-return recognition model and a method and device for cash-return recognition, the training method comprises: acquiring first transaction information of a first transaction and a first cash-return label value of the first transaction; calculating a first cash-return predictive value for the first transaction, and a Q-value's label value corresponding to the first transaction information and the first cash-return predictive value; and training a Deep Q-Network (DQN) by adjusting parameters of the DQN such that an absolute value of a difference between a first Q-value's predictive value output by the trained DQN and the Q-value's label value is smaller than that between the first Q-value's predictive value output by the untrained DQN and the Q-value's label value, the first Q-value's predictive value obtained by inputting the first transaction information and the first cash-return predictive value to the DQN.