Machine Learning Data Augmentation for Fair Minority Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models can make unfair determinations due to biases in learning data, particularly when data for minority attributes is scarce, leading to inaccurate classifications and ethical concerns.
Innovation Solution
An information processing apparatus and method that generates additional learning data using Adversarial Examples, specifically through the Fast Gradient Sign Method, to balance the dataset and reduce bias, ensuring fair determinations across all attributes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If learning data is collected from real-world sources, then the data reflects actual distribution patterns, but minority attribute data remains scarce causing bias
Solution Approach 1:
The patent applies preliminary action by generating adversarial examples in advance to augment minority attribute data before model training. The data generation unit creates synthetic minority samples using GANs or other generation methods, then incorporates them into the training dataset beforehand, allowing the model to learn fair determination without waiting for sufficient real minority data to be collected naturally.
Solution Approach 2:
The patent applies parameter changes by modifying data distribution parameters through adversarial example generation. The system changes the quantity and distribution parameters of minority attribute data by generating synthetic samples with controlled attributes, transforming the imbalanced data distribution into a more balanced one that enables fair model determination.
2Reliability
If adversarial examples are generated to balance minority attributes, then fairness improves, but data processing complexity increases
Solution Approach 1:
The patent applies self-service by implementing automated adversarial example generation and integration systems. The data generation unit automatically detects minority attributes, generates appropriate adversarial examples, and incorporates them into the training dataset without manual intervention. This self-serving automation reduces the operational complexity burden despite the increased processing requirements.
Solution Approach 2:
The patent introduces an intermediary data generation unit that mediates between the original imbalanced dataset and the model training process. This intermediary component handles the complex adversarial example generation and integration tasks, isolating the complexity from the core model training pipeline and making the overall system more manageable.
3Reliability
If manual data adjustment is performed to balance attributes, then fairness can be improved, but time consumption and consent hurdles increase
Solution Approach 1:
The patent applies mechanics substitution by replacing manual mechanical data adjustment processes with automated computational methods. Instead of manually collecting, labeling, and balancing data (mechanical human operations), the system uses algorithmic adversarial example generation and automated data augmentation techniques to achieve the same fairness objective much faster and at scale.
Solution Approach 2:
The patent applies parameter changes by transforming the time-consuming manual data balancing process into rapid automated parameter manipulation. The system changes data distribution parameters through computational generation rather than manual collection, dramatically reducing the time parameter from months of manual work to minutes or hours of automated processing.
Data Source
AI summary
Provided is an information processing apparatus for artificially increasing data of a minority attribute to generate learning data for making fair determination on each piece of input data.An information processing apparatus includes: a data holding unit configured to hold first learning data to be used for learning a machine learning model; an acquisition unit configured to acquire information regarding bias of the learning data; a data generation unit configured to generate second learning data by using data included in the learning data on the basis of the information regarding bias; and a learning unit configured to learn the machine learning model by using the first learning data and the second learning data.