MLRG theme: Responsible ML

Introduction presentation

MLRG theme: Responsible ML

• Bias to address Introduce ranking bias
• What is unfair? Social welfare
• What is fair? Unequal opportunity

Today's paper

Classic ethics question

* picture from AFP

* picture from socialprotection.org

Taxation / social insurance

Notation

Setting

Notation

Process

Notation

Recommender system policy

Notation

Per user

Per provider

1-d examples

1-d examples

• Three content providers each with viability = 2
• Six users with utility based on distance to provider

1-d examples

• Myopic RS matching

1-d examples

• Result

• users 1, 2 and 3 --> provider 1
• users 4 and 5 --> provider 2
• user 6 --> provider 3

1-d examples

• Result of subsequent rounds

• users 1, 2 and 3 --> provider 1
• users 4, 5 and 6--> provider 2

• Per round total utility: 8 + 2

• user 6's utility goes from 2 to 0 without provider 3

1-d examples

• Non-myopic policy could be

• users 1 and 2 --> provider 1
• users 3 and 4 --> provider 2
• users 5 and 6 --> provider 3

• Per round total utility: 10 - 2

• Users 3 and 5 gives up 2     utility per round

Fairness? 

• subsidize provider 3
• slightly lower rewards for users 3 and 5 at every timestep for perpetuity
• much higher reward for user 6 for perpetuity

Social welfare

• group every T timesteps into an epoch
• at epoch k under policy

• long-run social welfare

Equilibrium

• subset of providers that are viable in the long run

• many equilibria - find the one that maximizes long-run social welfare

• achieve with a single rule to apply across all epochs

Objective (for any epoch)

(expectation over user query distribution, timesteps within epoch and policy)

Objective (for any epoch)

• maximize social welfare
• ensure that all providers remain viable
• how does this translate to optimality over infinite epochs?

Solving the problem

• linear programming 
• polynomial time for 1/     approximation

If utility function is additive

For non-linear utility functions

• linearize

Experiments

• linear programming solution to the welfare objective (LP-RS)
• three datasets

• synthetic data
• Movielens movie ratings
• SNAP 2010 Twitter follower

• myopic policies underperform LP-RS

Model refinements

• fairness to individual users
• incomplete information
• robustness to changes in query distributions
• full RL approach

Fairness to individual users

• long-run utility for       under 

• denote      as the best possible policy for 

• define fairness in terms of regret

Fairness to individual users

• regret of      under

• maximum regret under 

• add a penalty term to objective based on MR