Reinforcement Learning for Simulated Transaction Data Generation

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Financial crime detection systems face challenges in training predictive models due to the sensitivity and limited availability of real customer data, necessitating the use of simulated transaction data that resembles real data to effectively detect fraudulent activities.

Innovation Solution

A computer-implemented method using a reinforcement learning model with an intelligent agent, policy engine, and environment to simulate transaction data, iteratively adjusting policies based on feedback to achieve a high degree of similarity with standard customer transaction data, and identifying behavioral patterns, including fraudulent behavior.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If real customer data is used to train the predictive model, then the model accuracy is improved, but data availability is limited due to sensitivity concerns

Engineering Contradiction:

Improvemodel accuracyVSAvoiddata availability

Core Design Contradiction:

Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent creates simulated customer transaction data that copies the statistical characteristics and behavioral patterns of real customer data. The simulation generates synthetic transaction records, account balances, and customer behaviors that mirror real banking data distributions, enabling model training without using actual sensitive customer information.

Inventive Principle:

Principle #26Copying

Solution Approach 2:

The patent transforms real customer data characteristics into simulated data by adjusting and randomizing parameters such as transaction amounts, frequencies, account types, and customer demographics. The simulation varies these parameters within statistically valid ranges to generate diverse training scenarios while preserving the underlying behavioral patterns of real customers.

Inventive Principle:

Principle #35Parameter changes

2Quantity of substance

If simulated customer data is used to train the predictive model, then data availability is improved, but the realism and quality of training data may be reduced

Engineering Contradiction:

Improvedata availabilityVSAvoidtraining data quality

Core Design Contradiction:

Quantity of substanceVSMeasurement precision

Solution Approach 1:

The patent employs reinforcement learning where the simulation receives feedback from the predictive model's performance. The simulated data generation process is iteratively refined based on how well the trained model detects fraud, with the environment providing rewards or penalties that guide the policy engine to generate more realistic fraudulent and legitimate transaction patterns.

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The patent performs preliminary analysis of real customer transaction data to establish statistical baselines, behavioral patterns, and fraud characteristics before generating simulated data. This preliminary characterization ensures that the simulated data inherits the essential features of real banking transactions, including transaction amount distributions, frequency patterns, and customer behavior profiles.

Inventive Principle:

Principle #10Preliminary action

3Productivity

If more simulated transaction data is generated, then the predictive model training is improved, but the complexity of the simulation system increases

Engineering Contradiction:

Improvetraining data generationVSAvoidsimulation system complexity

Core Design Contradiction:

ProductivityVSDevice complexity

Solution Approach 1:

The patent divides the simulation system into distinct modular components: an intelligent agent that generates transaction actions, an environment that models banking operations and customer behaviors, and a policy engine that learns from feedback. Each module handles specific aspects of data generation independently, allowing the system to scale data production without proportionally increasing overall complexity.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The simulation system uses reinforcement learning to automatically refine its own data generation process. The policy engine self-adjusts the simulation parameters and transaction patterns based on feedback from the predictive model performance, eliminating the need for manual tuning and reducing operational complexity while maintaining high productivity in generating realistic training data.

Inventive Principle:

Principle #25Self-service

Data Source

PatentUS11461793B2Identification of behavioral pattern of simulated transaction data

Publication Date: 2022.10.04 INTERNATIONAL BUSINESS MACHINE CORPORATION

AI summary

Embodiments can provide a method for identifying a behavioral pattern from simulated transaction data, the method including: simulating transaction data using a reinforcement learning model; and identifying a behavioral pattern from the simulated transaction data. The step of simulating transaction data further includes: providing standard customer transaction data representing a group of customers having similar transaction characteristics as a goal; and performing a plurality of iterations to simulate the standard customer transaction data, wherein the plurality of iterations is performed until a degree of similarity of simulated customer transaction data relative to the standard customer transaction data is higher than a first predefined threshold. In each iteration, the method includes conducting an action including a plurality of simulated transactions; comparing the action with the goal; providing feedback associated with the action based on a degree of similarity relative to the goal; and adjusting a policy based on the feedback.