Currently indirect bias detection works for independent features. For ex: race may have a high correlation with zip code, income etc.
But need to do the same for composite features like black female, or young black female. Sometimes these protected features by themselves may not have a significant effect but does when taken together
For a customer who is interested to identify which all features have a high correlation leading to indirect bias, it needs to be done automatically, since manual approach may not capture everything
Also, how would bias be detected for a composite feature like black male or older white male, where some feature are monitored group and some are protected groups
Why is it useful?
|Who would benefit from this IDEA?||Customer - Walmart|
How should it work?
1) Should automatically look for indirect bias not just for the protected features selected but also among the composite features that can be formed from them
2) Then identify all correlated features for such composite features and then identify bias via perturbation
3) Workaround : Manually create these composite features by doing cartesian product, but cannot consider all possible scenarios, and even if we can, cannot expect a customer to do this manually
|Priority Justification||Walmart likes to use this in their HR hiring usecase aswell as in other scenarios|