morrissharp
commented
2 years ago Suggest change to: Data Balance Analysis is relevant for the overall understanding of datasets, but is essential to building Machine Learning models in a responsible way, especially in term of fairness. It is all too easy to build an ML Model that produces biased results for subsets of the population by training or testing the model on biased ground truth data. There are multiple case studies of biased models assisting in granting loans healthcare, recruitment opportunities and many other decision-making tasks. In most of these examples, the data on which these models are trained was the common issue. These findings emphasize how important it is for model creators and auditors to analyze data balance:
- to measure training data across various sub-populations
- to ensure the data has good coverage, and a balanced representation of labels across sensitive categories and category combinations
- and to check that the test data is representative of the target population
In summary, Data Balance Analysis has the following benefits when used for building ML models.
- Reduces the risk of unbalanced models by:
- ensuring service fairness and reducing the costs of ML building by identifying data representation gaps early on
- prompting data scientists to seek mitigation steps before proceeding on the training portion of Machine Learning model development
- Enables easy end-to-end debugging of ML systems in combination with Fairlearn by providing a clear view if an issue in a model is tied to the data or the model itself.
I think this is missing a blank line before the bulleted list. The bulleted list is not displaying properly in the docs.
https://github.com/microsoft/responsible-ai-toolbox-mitigations/blob/83c26a3fc61f15c609734126bd4e76e0d922fdbc/raimitigations/databalanceanalysis/aggregate_measures.py#L36