(generated image for illustration only)
Algorithmic bias mitigation techniques are categorized into three main stages relative to the model training process. Pre-processing methods modify the training data itself (e.g., reweighing, resampling). In-processing methods incorporate fairness constraints directly into the model’s learning algorithm. Post-processing methods adjust the model’s predictions after they have been made to improve fairness.
This three-part classification provides a structured framework for addressing bias. Pre-processing is data-centric; it aims to create a ‘fair’ dataset before the model sees it. Techniques like reweighing assign different importance to data points to counteract imbalances, while over/under-sampling adjusts the number of instances from different groups. This approach is model-agnostic but may alter the data’s underlying distribution.
In-processing is model-centric. It modifies the learning algorithm’s objective function to include a penalty term for unfairness. For example, a model might be optimized to maximize accuracy while simultaneously minimizing the difference in error rates between groups. This can lead to more integrated solutions but requires modifying the core algorithm, making it less flexible.
Post-processing is prediction-centric. It takes the outputs of a trained, potentially biased model and adjusts them to satisfy a fairness criterion. This can involve changing classification thresholds for different groups. It’s the least invasive method as it treats the model as a black box, but it may reduce overall utility and can seem ad-hoc. The choice of stage depends on factors like access to the training data, ability to modify the model, and the specific fairness goals.
Bias Mitigation Processing Stages
(if date is unknown or not relevant, e.g. "fluid mechanics", a rounded estimation of its notable emergence is provided)
Full size images and downloads are only available, 100% free, for registered members.
> Login <
