Fairness: Mitigating bias

Once a source of bias has been identified in the training data, we can take proactive steps to mitigate its effects. There are two main strategies that machine learning (ML) engineers typically employ to remediate bias:

• Augmenting the training data.
• Adjusting the model's loss function.

Augmenting the training data

If an audit of the training data has uncovered issues with missing, incorrect, or skewed data, the most straightforward way to address the problem is often to collect additional data.

However, while augmenting the training data can be ideal, the downside of this approach is that it can also be infeasible, either due to a lack of available data or resource constraints that impede data collection. For example, gathering more data might be too costly or time-consuming, or not viable due to legal/privacy restrictions.

Adjusting the model's optimization function

In cases where collecting additional training data is not viable, another approach for mitigating bias is to adjust how loss is calculated during model training. We typically use an optimization function like log loss to penalize incorrect model predictions. However, log loss does not take subgroup membership into consideration. So instead of using log loss, we can choose an optimization function designed to penalize errors in a fairness-aware fashion that counteracts the imbalances we've identified in our training data.

The TensorFlow Model Remediation Library provides utilities for applying two different bias-mitigation techniques during model training:

• MinDiff: MinDiff aims to balance the errors for two different slices of data (male/female students versus nonbinary students) by adding a penalty for differences in the prediction distributions for the two groups.

• Counterfactual Logit Pairing: Counterfactual Logit Pairing (CLP) aims to ensure that changing a sensitive attribute of a given example doesn't alter the model's prediction for that example. For example, if a training dataset contains two examples whose feature values are identical, except one has a gender value of male and the other has a gender value of nonbinary, CLP will add a penalty if the predictions for these two examples are different.

The techniques you choose for adjusting the optimization function are dependent on the use cases for the model. In the next section, we'll take a closer look at how to approach the task of evaluating a model for fairness by considering these use cases.

Exercise: Check your understanding

Which of the following statements regarding bias-mitigation techniques are true?

Both MinDiff and CLP penalize discrepancies in model performance tied to sensitive attributes

Both techniques aim to mitigate bias by penalizing prediction errors resulting from imbalances in how sensitive attributes are represented in training data.

MinDiff penalizes differences in the overall distribution of predictions for different slices of data, whereas CLP penalizes discrepancies in predictions for individual pairs of examples.

MinDiff addresses bias by aligning score distributions for two subgroups. CLP tackles bias by ensuring that individual examples are not treated differently solely because of their subgroup membership.

Adding more examples to the training dataset will always help counteract bias in a model's predictions.

Adding more training examples is an effective strategy for mitigating bias, but the composition of the new training data matters. If the additional training examples exhibit similar imbalances to the original data, they probably won't help mitigate the existing bias.

If you are mitigating bias by adding more training data, you shouldn't also apply MinDiff or CLP during training.

Augmenting training data and applying techniques like MinDiff or CLP can be complementary. For example, an ML engineer might be able to collect enough additional training data to reduce a discrepancy in performance by 30%, and then use MinDiff to further reduce the discrepancy by another 50%.

Key terms:

• Bias (ethics/fairness)
• Log Loss