What does fairness mean (ml modeling vs production ml)

In the slide that compares academic ml vs production ml, fairness is chosen as a parameter.

What is does fairness mean in this context ?

Fairness is important for both academic and industrial settings.
The primary driver for an academic work is the overall model performance (say, accuracy). A paper that doesn’t pay heavy attention to fairness can still be published.
In a production ML setting, when fairness is not met, the product receives negative feedback from the audience and is rejected quickly. Here’s an example of how amazon recruitment tool went away from the market:

1 Like

Hi there

and welcome to the community!

Thanks for your post, @kshashankrao.
Bias or unfairness usually can emerge from two dimensions:

  • data (like measurement bias)
  • algorithms (like algorithmic bias)

Every dataset is the result of several design decisions made by the data curator. […]
For instance, authors try to satisfy subgroup fairness in classification, equality of opportunity and equalized odds.

see also: https://arxiv.org/pdf/1908.09635.pdf

As mentioned already, if you have a rather limited and specific experiment, you can still publish a paper , describing the limits and potential of your findings, e.g. classification of animals or so.

In a prod setting, not only intended use but at least also unintended use (and with certain limits also misuse) of the AI system needs to be considered as @balaji.ambresh mentioned already! In this paper this interaction with the user is also discussed.

Best regards
Christian

1 Like