Matching Markets with Ordinal Preferences

Matching Markets with Ordinal Preferences
TIFR, May 2013
Matching Markets
πœ‹1
πœ‹2
πœ‹3
β€’ N agents, N items, N complete preferences.
β€’ Outcome:
Agent-Item Matching
Outline of Talk
β€’ Mechanisms
– Random Serial Dictatorship (RSD)
– Rank Maximal Matching (RMM)
β€’ Welfare
– Ordinal Welfare Factor
– Rank Approximation
β€’ Truthfulness
– Dealing with randomness.
Random Serial Dictatorship
β€’ Agents arrive in a random permutation and
pick their best unallocated item.
πœ‹ = (2,1,3)
…
πœ‹ = (3,2,1)
Choice 1
Choice 2
Choice 3
Rank Maximal Matching
β€’ Maximize #(top choice), then Maximize #(top 2),...
β€’ Polytime computable.
Irving, 2003
Irving, Kavitha, Melhorn, Michail, Paluch, 2004.
Social Welfare
β€’ Pareto Optimality.
No other outcome makes everyone happier.
β€’ RMM leads to a Pareto Optimal outcome.
β€’ RSD leads to ex-post Pareto Optimal outcome.
Social Welfare
β€’ Cardinal Welfare
Each pair associated with cardinal number.
Social welfare = Sum of utilities.
β€’ What to do when no numbers are known?
Ordinal Welfare Factor (OWF)
β€’ Outcome 𝑀 is 𝛼-efficient, if for any 𝑀′,
#agents with 𝑀 β‰₯ 𝑀′ β‰₯ 𝛼𝑁
β€’ Problem: Everyone has same ordering.
(1, 1)
(2, 2)
M = (3, 3)
…
(N,N)
(1, N)
(2, 1)
M’ = (3, 2)
…
(N,N-1)
𝛼 < 1/𝑁
Ordinal Welfare Factor (OWF)
β€’ Randomization.
A distribution 𝑴 is 𝛼-efficient, if for any other
distribution 𝑴′,
𝐄𝐱𝐩 𝑀←𝑴,𝑀′←𝑴′ [#agents with (𝑀 β‰₯ 𝑀′ )] β‰₯ 𝛼𝑁
β€’ Mechanism has OWF 𝛼 if it returns an 𝛼efficient distribution.
Symmetric β€œBad” Example
β€’ Every agent has same preference order.
β€’ 𝑴 is uniform over all matchings.
β€’ Fix matching 𝑀′ = { 1,1 , 2,2 … , 𝑁, 𝑁 },
βˆ€π‘–,
β€’ 𝑴 is
𝐏𝐫 𝑀←𝑴
1
1
+
2
2𝑛
𝑖
𝑀 β‰₯𝑖 𝑀 β‰₯
𝑁
-efficient.
β€²
Performance of Mechanisms
β€’ Theorem. RSD has OWF β‰₯ 1/2
Bhalgat, C, Khanna 2011.
β€’ RMM is deterministic.
Many agents can be made better off at the
expense of one agent.
Strengths and Weaknesses
β€’ Comparative Measure.
β€’ Notion of β€œapproximation”.
Quantify mechanisms.
β€’ Not good for deterministic mechanisms.
β€’ No notion of β€œhow much better off”.
Rank Approximation
β€’ Let π‘€π‘–βˆ— maximize #(agents getting top i)
𝑛𝑖 ≔ π‘€π‘–βˆ—
β€’ 𝑀 is 𝛼-rank approximate if
#(agents getting top 𝑖 in 𝑀) β‰₯ 𝛼𝑛𝑖 .
β€’ Mechanism has 𝛼-rank approximation if it
returns an 𝛼-rank approximate matching.
Connection to Cardinal Welfare
β€’ Homogenous agents:
Each agent has same cardinal profile
𝑒1 > 𝑒2 > β‹― > 𝑒𝑁
β€’ 𝑀 is 𝛼-rank approximate implies
𝛼-approximation for homogenous agents.
Performance of Mechanisms
β€’ Theorem. RMM has ½-rank approximation.
1
- Maximal/Maximum β‰₯
2
- Optimal.
β€’ RSD is not 𝛼-approximate for any constant 𝛼.
β‰ˆ 𝑁
Choice 1
Strengths and Weaknesses
β€’ Deterministic mechanisms can have good rank
approximation.
β€’ Cardinal welfare for homogenous agents.
β€’ Could improve many while hurting only a few.
β€’ No good rank appx known in non-matching
setting.
Truthfulness
β€’ If an agent lies, he gets a worse item.
If an agent lies, he doesn’t get a better item.
β€’ Issues with randomized mechanisms.
What are worse and better distributions?
β€’ Hierarchy of truthfulness.
Randomization vs Truthfulness
Universally Truthful.
Distribution over deterministic mechanisms
Strongly Truthful. (Gibbard, 77)
Lying gives a stochastically dominated allocation.
Lex Truthful. (?)
Lying gives a lexicographically dominated allocation.
Weakly Truthful. (Bogomolnaia-Moulin, 01)
Lying can’t give stochastically dominating allocation.
Lex Truthful Implementation
β€’ A deterministic algorithm A can be πœ–-lextruthful implemented if there is a randomized
mechanism M such that
– M is Lex Truthful.
– With probability > (1-πœ–), outcome of M is same as
that of A
Theorem. Any pseudomonotone algorithm A
is πœ–-lex-implementable, for any πœ– > 0.
C, Swamy 2013
Pseudomonotonicity
A πœ‹π‘– , πœ‹βˆ’π‘– = 𝑀
πœ‹π‘–
M(i)
M’(i)
M’(i) is below M(i) in πœ‹π‘–
A πœ‹β€²π‘– , πœ‹βˆ’π‘– = 𝑀′
πœ‹π‘–
πœ‹β€²π‘–
b
M’(i)
M(i)
b
or there’s b above M’(i) in πœ‹π‘–
which has been demoted.
Performance of Mechanisms
β€’ RSD is Universally Truthful.
Under certain conditions, it is the only
strongly truthful mechanism. (Larsson, 94)
β€’ RMM satisfies pseudomonotonicity.
Therefore, it can be πœ–-LT implemeneted.
Summary
β€’ Welfare definitions unclear in ordinal settings.
Saw two notions.
Generalizes to Social choice settings.
β€’ Truthfulness of randomized mechanisms also
tricky. Hierarchy of truthfulness.
β€’ Can results be extended to general settings?