change their decision (fixing everyone else’s actions). • If Player 2 chooses to Stay Silent Player 1 will want to Confess (0 vs -1). • If Player 2 chooses to Confess Player 1 will want to Confess(-3 vs -5). • The Nash Equilibrium of this game is (Confess, Confess). -1 -5 0 -3 Silent Confess Silent Confess -1 0 -3 -5
(Stag, Stag) and (Hare, Hare). • This is an important problem of evolutionary game theory ◦ Stag Hunt shows the evolution of trust and the value of communication. ◦ (Stag, Stag) is a better equilibrium than (Hare, Hare), but this requires me to trust the other player. 5 0 2 2 Stag Hare Stag Hare 5 2 2 0
is a coalition, a group of players. • The goal of cooperative game theory is to build a team of players to achieve a common result. ◦ Ensemble Learning has a lot of theoretical basis in cooperative game theory. • With the recent interest in model interpretability, researchers have applied cooperative game theory to problem of evaluating feature importance in a model. ◦ The features of the model are seen as the “players” of the game, and the model’s output is “result” of the game.
contribution to the team. • Heuristic: If we remove a player from the team and the outcome doesn’t change, then the player wasn’t useful. • To compute the Shapley Value for a player, we compute each outcome where the player was present, and compare it to the counterfactual outcome where the player wasn’t present. Player i present Player i not present
SHAP integrated with XGBoost, a popular gradient boosting tree algorithm. • Brute-forcing all 2N feature combinations is inefficient, so TreeSHAP uses the structure of decision trees to quickly approximate the Shapley Values. • Python Library: https://github.com/slundberg/shap (Age, Education, Marital Status) (Age, Education, Marital Status, Occupation) (Age, Education, Marital Status, Workclass) (Age, Education, Marital Status, Occupation, Workclass)
in algorithmic fairness is how to compensate individuals for their data*. • Game Theoretic Set-up of the Problem: ◦ We view the training data points D as the “players” of the game, and given some learning algorithm A and performance metric V, we want to optimize the result V(S, A) where S is a subset of the data. ◦ This is not a universal framework: The result is dependent on the choice of the algorithm A and performance metric V. • Goal: We want to compute the marginal contribution of each data point Φ i (V) to the performance metric. • Gradient Shapley Algorithm ◦ Generate a random permutation of the training data. ◦ Update model using gradient descent, and compute the marginal contribution as the change in performance. Source: Ghorbani and Zou (2019), http://proceedings.mlr.press/v97/ghorba ni19c/ghorbani19c.pdf
structure, Find the optimal* behavior for each agent. Mechanism Design: Starting with the desired behavior for each agent, Find an incentive structure that induces this behavior. Economic Incentives Optimal Behavior Game Theory Mech. Design
and they must decide how to split the cake between themselves. • Assume each of them prefer larger portions of cake (monotonically increasing utility), how can we cut the cake to appease both of them? • Formally, we want the cake allocation to be envy-free (Alice doesn’t want Bob’s cake more than her own).
and Bob chooses which portion he wants. • Assuming Alice is rational*, she will realize if she cuts the cake unevenly (for example, 70/30 split), Bob can always take the larger portion. • In order to maximize her portion, Alice should cut the cake 50/50. • This frameworks in an optimal allocation that is both proportional (fair) and envy-free (neither strictly prefers the others’ allocation). *rational as defined by rational choice theory
is designing markets for exchange economies. ◦ Alvin E. Roth won the 2012 Nobel Prize in Economics for this and other related work. • Given a set of Donors D and recipients R, we can construct a bipartite graph from a donor d to a recipient r if they have blood-type compatibility. • With blood types, there are universal donors (blood-type O negative) and universal receivers (blood-type AB positive). ◦ Heuristic: We want to efficiently match the most number of donor/recipient pairs, so we shouldn’t waste universal donors on universal receivers. • Roth’s work allows for maximal number matchings in this Kidney exchange graph.
mathematical economics to improve access to opportunity for historically underserved and marginalized communities. • Many existing Data Science for Social Good initiatives often apply existing algorithms to problems of societal impact. • The goal of this research is to CREATE NEW mathematical and computational frameworks for optimizing resource allocation and improving information access.
to income shocks (unexpected medical bill, car accident). • Goal: We want to allocate income subsidies such that we reduce the risk that any household experiences financial ruin (bankruptcy) • Using ideas from ruin probabilities, the authors create a framework to minimize the likelihood that any household becomes financially ruined. Ruin Probability for Individual i: We model the probability an individual experiences financial ruin as a function of their initial income, net pay, and financial shock size and frequency.
reserve u i . • At each timestep t, the household receives some net income c i . • HOWEVER, household also randomly experience income shocks S ij during the time step • We want to keep the household’s total income R i (t) to never drop below 0. ◦ Connection to Random Walks Initial Income Net Income received Sum of Income Shocks Income at time t
for income subsidies. We want to allocate some subsidy x i for each household such that • Min-Sum Objective: ◦ Each household has some weight w i representing social cost of ruin. In this formulation we want to minimize the weighted expected number of agents who experience ruin. ◦ A solution to this objective function can be approximated using the Knapsack algorithm (Value of household is societal weight and weight of household is income subsidy). • Min-Max Objective: ◦ In this social objective, we want to keep the worst ruin probability experienced by any of the households as low as possible (Save everyone) ◦ A solution to this objective function can be approximated using modified Binary Search. Guess some min-max solution p, if it is infeasible, adjust down, else adjust up. Worst ruin probability
Scott Lundberg, Su-In Lee. https://arxiv.org/abs/1705.07874 • Data Shapley: Equitable Valuation of Data for Machine Learning; Amirata Ghorbani, James Zou. https://arxiv.org/abs/1904.02868 • Kidney Exchange; Alvin Roth, Tayfun Sonmez, M. Utku Unver. https://web.stanford.edu/~alroth/papers/kidney.qje.pdf • Subsidy Allocation in the Presence of Income Shocks; Rediet Abebe, Jon Kleinberg, S. Matthew Weinberg. https://www.cs.cornell.edu/home/kleinber/aaai20-shocks.pdf
◦ AI Methods for Social Good (CMU) ◦ Applied Mechanism Design for Social Good (UMD) • Books ◦ The Model Thinker by Scott Page ◦ Handbook of Computational Social Choice by Ariel Procaccia ◦ Algorithmic Game Theory by Nisan, Roughgarden, Tardos, and Vazirani.