Reinforcement Learning Agent for ML Label Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models, particularly in fraud detection, suffer from inaccuracies due to false positive and false negative predictions, which are difficult to eliminate and require significant network resources, especially when training datasets contain incorrect labels.
Innovation Solution
Utilize reinforcement learning to enhance a deep learning model by generating a second training dataset with improved labeling accuracy through a reinforcement learning agent (RLA), followed by training and testing to generate a reward parameter, thereby increasing the detection rate of correct predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning models are trained on initial training datasets, then the models can perform predictions, but the prediction accuracy is reduced due to false positive and false negative predictions from incorrect labels in the training data
Solution Approach 1:
A reinforcement learning agent is introduced as an intermediary between the initial training dataset and the deep learning model. The agent processes the initial dataset, corrects incorrect labels through reinforcement learning, and generates an improved training dataset with higher label accuracy, thereby resolving the contradiction between prediction accuracy and label reliability
Solution Approach 2:
The reinforcement learning agent uses feedback mechanisms to iteratively improve label accuracy. The agent receives feedback from the deep learning model's predictions and adjusts its labeling corrections accordingly, creating a closed-loop system that continuously enhances both label accuracy and prediction accuracy
2Device complexity
If traditional machine learning models are used with initial training datasets, then the training process is simpler, but significant network resources are required to eliminate incorrect predictions
Solution Approach 1:
The reinforcement learning agent performs preliminary action by correcting labels in the training dataset before the deep learning model is trained. This preprocessing step eliminates incorrect predictions in advance, reducing the need for post-training resource-intensive correction processes and optimizing network resource utilization
3Ease of manufacture
If the initial training dataset contains incorrect labels, then data collection is easier, but the detection rate of correct predictions is reduced
Solution Approach 1:
The manual mechanical process of meticulous data labeling is replaced with an automated reinforcement learning system. The RLA automatically identifies and corrects incorrect labels in the training dataset, maintaining ease of data collection while significantly improving the detection rate through automated label refinement
Data Source
AI summary
Provided are methods that include receiving an initial training dataset, where each data instance of the initial training dataset has a label and a first percentage of the plurality of data instances are labeled correctly, providing the initial training dataset as an input to a reinforcement learning agent (RLA) machine learning model to generate a second training dataset, where a second percentage of data instances of the second training dataset are labeled correctly and the second percentage is greater than the first percentage, training a deep learning model using the second training dataset to provide a trained deep learning model, testing the trained deep learning model using a testing dataset to generate a resultant dataset that has a detection rate, which is an indication of a number of data instances that were predicted correctly, and generating a reward parameter. Systems and computer program products are also disclosed.