SVM Training Debiasing via Uniform Distribution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Inference models, particularly those based on support vector machines, often exhibit latent bias due to biased training data, leading to undesirable impacts on computer-implemented services by generating inferences that discriminate based on unintended features.
Innovation Solution
A training procedure is implemented that uses a debiasing term to incentivize a uniform distribution of records across the decision boundary in support vector machine-based models, reducing the likelihood of latent bias by incorporating an objective function that balances classification accuracy with non-discrimination on bias features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional support vector machine models are trained on biased training data, then classification accuracy is improved, but latent bias is introduced leading to discrimination based on unintended features
Solution Approach 1:
The patent converts the harmful latent bias introduced by biased training data into a beneficial effect by using the bias as a signal to adjust the decision boundary. The debiasing term in the objective function utilizes the presence of bias to identify and correct discriminatory patterns, transforming the harmful bias signal into a mechanism for achieving more equitable classifications while maintaining accuracy.
Solution Approach 2:
The patent changes the parameters of the support vector machine by introducing a debiasing term to the objective function. This modifies the traditional SVM optimization to include additional constraints that penalize discrimination based on protected features. The parameter change allows the model to maintain classification accuracy while adjusting the decision boundary to reduce latent bias and promote fairness.
2Object-generated harmful factors
If a debiasing term is added to the objective function to reduce latent bias, then fairness is improved, but model complexity increases
Solution Approach 1:
The patent introduces a debiasing term as an intermediary component in the objective function. This term acts as a mediator between the classification accuracy requirement and the fairness constraint. The debiasing term processes the bias information and translates it into appropriate penalty signals, simplifying the overall model structure while achieving the dual goal of accuracy and fairness without requiring complex additional architectures.
Data Source
AI summary
Methods, systems, and devices for providing computer implemented services are disclosed. To provide the computer implemented services, inference models may generate and provide inferences used in the computer implemented services. The inference models may be obtained through training using training data. Training processes used to train the inference models may proactively to attempt to reduce the likelihood of the trained inference models exhibiting latent bias. The training process may disincentivize predictive power with respect to bias features and incentivize predictive power for labels through use of debiasing terms.