Machine Learning Fairness Correction via User-Specified Attributes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning model correction techniques fail to adapt to varying fairness evaluation standards, leading users to abandon the system as they cannot achieve the desired fairness metrics.
Innovation Solution
A machine learning program that allows users to specify attributes and labels based on desired fairness standards, enabling the model to be trained and corrected to meet specific fairness criteria through interaction with the user, including classification results and combined metrics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If conventional automatic model correction based on conventional fairness indices is used, then the model correction process is automated and efficient, but the model cannot achieve fairness according to varying user-defined standards
Solution Approach 1:
The system dynamically adjusts the fairness correction process by allowing users to specify their own fairness indices and attributes. The model correction transitions from a static, pre-defined approach to a dynamic, user-adaptive process where the fairness criteria can change based on user input and different evaluation scenarios.
Solution Approach 2:
The system enables parameter changes by allowing users to modify fairness indices, select different attributes for evaluation, and adjust the weighting of various fairness metrics. This flexibility in parameter specification allows the same correction framework to adapt to different fairness evaluation standards without requiring complete re-automation of the process.
2Device complexity
If only existing fairness metrics are presented to users, then the system structure remains simple, but users cannot achieve their desired fairness standards and may abandon the system
Solution Approach 1:
The system introduces an intermediary layer between the user and the model correction process. This intermediary allows users to specify their fairness criteria through a structured interface that translates user requirements into appropriate fairness indices and attributes, bridging the gap between simple system structure and user needs without requiring users to directly implement complex correction algorithms.
Solution Approach 2:
The system performs preliminary actions by pre-defining a framework of fairness indices and attributes that users can select from and modify. This preliminary structure provides users with ready-to-use fairness evaluation options while still allowing customization, reducing the operational complexity users would otherwise face in defining fairness criteria from scratch.
3Ease of manufacture
If the model is trained with fixed fairness indices, then the training process is straightforward, but the model lacks adaptability to various fairness evaluation standards
Solution Approach 1:
The training framework achieves universality by designing a multi-functional system that can handle multiple fairness indices and evaluation standards through a single unified training process. The model is trained to be adaptable to various fairness criteria by incorporating user-specified attributes and indices, allowing the same training mechanism to serve multiple fairness evaluation purposes.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A machine learning program including instructions for obtaining multiple classification results of classification of a plurality of data pieces outputted from a machine learning model into which the plurality of data pieces have been inputted; specifying in accordance with classification results, a first plurality of attributes among a plurality of attributes included in a first plurality of data pieces classified into a first group and a second plurality of data pieces classified into a second group, each difference between each value of the first plurality of attributes of the first plurality of data pieces and each value of the first plurality of attributes of the second plurality of data pieces satisfying a condition; determining labels of the data pieces based on a first index representing a combination of the first multiple attributes; and training the machine learning model, using the labels and the data pieces.