This is exploration vs exploitation dilemma. For example let's say that 10% of ads are thrown randomly, and from these random rolls these patterns are discovered:
> [Denver+<40-50> years+men]: mechanics +10%
> [Denver+<40-50> years+men]: nurse -5%
Then the system can apply these coefficients on 90% of the other traffic.
If you are making 100% exploration (so 100% random), then it means the people are going to miss their relevant job opportunity (having a net negative impact on the society).
Increasing exploration is a solution that would legally actually reduce biases of previouses patterns, but at the cost of less relevant content.
In all cases, if the bias is real, exploration discovers them and the coefficients already naturally adjust.
One exception, advertisers can artificially restrict saying "I want only men between 30-40" in their targeting filters.
https://www.tutorialspoint.com/machine_learning/machine_lear...
This is exploration vs exploitation dilemma. For example let's say that 10% of ads are thrown randomly, and from these random rolls these patterns are discovered:
> [Denver+<40-50> years+men]: mechanics +10%
> [Denver+<40-50> years+men]: nurse -5%
Then the system can apply these coefficients on 90% of the other traffic.
If you are making 100% exploration (so 100% random), then it means the people are going to miss their relevant job opportunity (having a net negative impact on the society).
Increasing exploration is a solution that would legally actually reduce biases of previouses patterns, but at the cost of less relevant content.
In all cases, if the bias is real, exploration discovers them and the coefficients already naturally adjust.
One exception, advertisers can artificially restrict saying "I want only men between 30-40" in their targeting filters.
Then what Meta can do ? Not much.