RAISP02-BP03 Mitigate unwanted bias directly in the core AI system design
Consider incorporating fairness mitigations such as sampling and optimization methods during training, alignment and calibration techniques that actively mitigate biased system responses, and post-processing strategies that review and adjust outputs before they reach users. The specific fairness strategies you use should directly support the fairness goals in your release criteria.
Level of risk exposed if this best practice is not established: High
Implementation considerations
-
Use sampling-based methods during training to improve model performance on underrepresented groups. Apply techniques like weighted sampling to give more importance to underrepresented examples, oversampling to create more training instances from minority groups, or stratified sampling to achieve balanced representation. Consider error-based weighted sampling where you identify groups that experience higher error rates on a validation set and sample datapoints from those groups at higher rates during training. These methods assist your model learn better patterns for each group instead of just the majority.
-
Consider using fairness metrics within the model loss function to guide model training to penalize unfair outputs.
-
Consider if demographic features or proxy features factor significantly into the model predictions by analyzing feature attributions for indications of a biased model. Consider using Amazon SageMaker AI Clarify for feature attributions and bias detection.
Resources
Related documents:
-
Fairness, model explainability and bias detection with SageMaker AI Clarify
-
Automate model retraining with Amazon SageMaker AI Pipelines when drift is detected
-
Accenture Enterprise AI – Scaling Machine Learning and Deep Learning Models
-
Amazon SageMaker AI AI: Prepare ML Data with Amazon SageMaker AI Data Wrangler
-
ISO/IEC 42001:2023 Information technology — Artificial intelligence — Management system
Related tools:
