When should I use MinDiff?
We recommend applying MinDiff in instances where your model performs well generally, but produces harmful errors more frequently on examples belonging to a sensitive group, and you wish to close the performance gap. The sensitive groups of interest may vary depending on your use case, but often include protected classes, such as race, religion, gender, sexual orientation, and more. Throughout this document, we will use “sensitive group” to refer to any set of examples belonging to a protected class.
There are two primary conditions for using MinDiff to address underperforming slices of data:
- You have already tuned and evaluated your model, identifying metrics that show underperforming slices of data. This must be done before applying model remediation.
- You have, or can obtain, a sufficient number of relevant labeled examples belonging to the underperforming group (more details below).
MinDiff is one of many techniques for remediating unequal behavior. In particular, it may be a good choice when you’re trying to directly equalize performance between groups. MinDiff can be used in conjunction with other approaches, such as data augmentation and others, which may lead to better results. However, if you need to prioritize which technique to invest in, you should do so according to your product needs.
When applying MinDiff, you may see performance degrade or shift slightly for your best performing groups, as your underperforming groups improve. This tradeoff is expected, and should be evaluated in the context of your product requirements. In practice, we have often seen that MinDiff does not cause top performing slices to drop below acceptable levels, but this is application- specific and a decision that needs to be made by the product owner.
On what model types can I apply MinDiff?
MinDiff has been shown to be consistently effective when applied to binary classifiers. Adapting the method for other applications is possible, but has not been fully tested. Some work has been done to show success in multi- classification and ranking tasks1 but any use of MinDiff on these or other types of models should be considered experimental.
On what metrics can I apply MinDiff?
MinDiff may be a good solution when the metric you’re trying to equalize across groups is false positive rate (FPR), or false negative rate (FNR), but it may work for other metrics. As a general rule, MinDiff may work when the metric you’re targeting is a result of differences in the score distributions between examples belonging to a sensitive group and examples not belonging to a sensitive group.
Building your MinDiff dataset
When preparing to train with MinDiff, you’ll need to prepare three separate datasets. As with regular training, your MinDiff datasets should be representative of the users your model serves. MinDiff may work without this but you should use extra caution in such cases.
Assuming you’re trying to improve your model’s FPR for examples belonging to a sensitive class, you’ll need:
- The original training set - The original dataset that was used for training your baseline model
- The MinDiff sensitive set - A dataset of examples belonging to the sensitive class with only negative ground truth labels. These examples will be used only for calculating the MinDiff loss.
- The MinDiff non-sensitive set - A dataset of examples not belonging to the sensitive class with only negative ground truth labels. These examples will be used only for calculating the MinDiff loss.
When using the library, you will combine all three of these datasets into a single dataset, which will serve as your new training set.
Picking examples for MinDiff
It may have seemed counterintuitive in the example above to carve out sets of negatively labeled examples if you are primarily concerned with disparities in false positive rate. However, remember that a false positive prediction comes from a negatively labeled example incorrectly classified as positive.
When collecting your data for MinDiff, you should pick examples where the disparity in performance is evident. In our example above, this meant choosing negatively labeled examples to address FPR. Had we been interested in targeting FNR, we would have needed to choose positively labeled examples.
How much data do I need?
Good question--it depends on your use case! Based on your model architecture, data distribution, and MinDiff configuration, the amount of data needed can vary significantly. In past applications, we have seen MinDiff work well with 5,000 examples in each MinDiff training set (sets 2 and 3 in the previous section). With less data, there is increased risk of lowered performance, but this may be minimal or acceptable within the bounds of your production constraints. After applying MinDiff, you will need to evaluate your results thoroughly to ensure acceptable performance. If they are unreliable, or do not meet performance expectations, you may still want to consider gathering more data.
When is MinDiff not right for me?
MinDiff is a powerful technique that can provide impressive results, but this does not mean that it is the right method for all situations. Applying it haphazardly does not guarantee that you will achieve an adequate solution.
Beyond the requirements discussed above, there are cases where MinDiff may be technically feasible, but not suitable. You should always design your ML workflow according to known recommended practices. For instance, if your model task is ill-defined, the product needs unclear, or your example labels overly skewed, you should prioritize addressing these issues. Similarly, if you do not have a clear definition of the sensitive group, or are unable to reliably determine whether examples belong to the sensitive group, you will not be able to apply MinDiff effectively.
At a higher level, you should always consider whether your product is an appropriate use for ML at all. If it is, consider the potential vectors for user harm it creates. The pursuit of responsible ML is a multi-faceted effort which aims to anticipate a broad range of potential harms; MinDiff can help mitigate some of these, but all outcomes deserve careful consideration.
1Beutel A., Chen, J., Doshi, T., Qian, H., Wei, L., Wu, Y., Heldt, L., Zhao, Z., Hong, L., Chi, E., Goodrow, C. (2019). Fairness in Recommendation Ranking through Pairwise Comparisons.