37.3 Consumer Fairness
Also see (Diao, Harutyunyan, and Jiang 2019)
37.3.1 (Fu et al. 2021) Unfair machine learning algorithms
Compares the consequences of Equal treatment (ET) and equal opportunity (EO)
Regular and protected groups are better off if we change from ET to EO?
Optimal learning efforts under ET and EO?
Is the firm better off if ET to EO?
Strategic behavior of a firm: decide the amount of leaning effort. The firm invest for better ML algorithm to predict the outcomes.
Settings
Consumers
Good consumer (non-defaulters) and bad consumers (defaulters)
Consumers are in either regular or protected group
Firm
Risk-neutral, and want to select good consumers
Invest in ML by paying the learning costs to separate the good and bad consumers
Set approval thresholds for each group
A fairness constraint: either ET or EO
Main Results:
Firm less motivated to invest in learning under EO than under ET
Profit is lower under EO because it’s hard to separate the good and bad consumers in the protected group
EO makes everyone (regular + protected) worse off
Regular group has higher threshold under EO (than under ET)
Firm invests less on the algorithm under EO
Model
Firm
accepts or rejects an application
A consumer can either be good or bad
the profit/loss (\(\alpha, \beta\)) where it comes from good/bad consumer
firm exerts leaning efforts (s) for better separability between the two groups
Consumer
Either regular or protected (size = 1)
% of bad consumer in each group is \(d\)
Expected gains:
\(\alpha_p = \alpha(1 - d_p), \beta_p = \beta d_p\)
\(\alpha_r = \alpha(1 - d_r), \beta_r = \beta d_r\)
The bank uses ML and assigns scores to candidates to represent their goodness where \(\gamma_r, \gamma_p\) are learning efficiency
Summary
- The optimal threshold and the profit depend on \(\beta_p\)