Timeline for Avoiding social discrimination in model building
Current License: CC BY-SA 4.0
8 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Jan 17 at 23:18 | comment | added | Ben | I think one of the big issues here is that it is a choice to use complex "black box" models from machine learning. There are many well-established statistical models where it is easy to understand and control the way that input variables affect predictions/decisions. It seems strange to me that people would choose black-box models and then fret about the complexity of controlling predictions. | |
| Jan 17 at 23:17 | history | edited | Ben | CC BY-SA 4.0 |
added 1777 characters in body
|
| Aug 26, 2021 at 18:37 | comment | added | Ben | @Alexis: That is of course the opening to the view that everything is caused by gender (due e.g., to "institutional discrimination") and therefore I will discriminate to "correct" this. Once that decision is made, the model will be discriminatory because the builder of the model is using discrimination to impose their God's eye view of how society "should have been". | |
| Apr 13, 2019 at 8:43 | history | edited | Ben | CC BY-SA 4.0 |
added 536 characters in body
|
| Jan 22, 2019 at 2:17 | comment | added | Alexis | +1 for thoughtful. "All you need to do is to formulate your model in such a way that it does not use gender (and inherent gender characteristics) as predictors." That's easy to write, but beginning to create algorithms for social decisions like hiring when society is in medias res means that things like income history, educational attainment, and previous position are causally downstream of gender. | |
| Jan 20, 2019 at 21:02 | comment | added | Carl | Tedious. Whereas there is merit in forethought, for example, assumptions, e.g., "How is gender bias problematic?" no-one is all knowing and there is no substitute for checking results post hoc. | |
| Jan 16, 2019 at 9:34 | comment | added | kjetil b halvorsen♦ | An alternative to removing correlated variables would be to train separate models for men and women. The question then is how to use those separate models? | |
| Jan 16, 2019 at 2:59 | history | answered | Ben | CC BY-SA 4.0 |