Rank Aggregation - PowerPoint PPT Presentation

1 / 25
About This Presentation
Title:

Rank Aggregation

Description:

Combine many different rank orderings on the same set of candidates/alternatives ... Direct Aggregation of Cardinal Utility Methods. summation, multiplication ... – PowerPoint PPT presentation

Number of Views:81
Avg rating:3.0/5.0
Slides: 26
Provided by: cheng3
Category:

less

Transcript and Presenter's Notes

Title: Rank Aggregation


1
Rank Aggregation
  • Chengkai Li

2
Rank Aggregation
  • Rank Aggregation
  • Combine many different rank orderings on the
    same set of candidates/alternatives to obtain a
    better ordering.
  • Example abcd, acdb, cabd ? acbd

3
Overview
  • Desired Properties
  • Rank Aggregation Methods
  • Condorcet Criterion vs. Borda
  • Kemeny Optimal Aggregation
  • Rank Aggregation in fields other than social
    choice
  • Web Search
  • Machine Learning
  • Information Retrieval
  • Database Middleware
  • Conclusion

4
Desired Properties
  • Condorcet Criterion (1785)
  • Condorcet alternative
  • Defeats every other alternative in pairwise
    simple majority voting
  • ab 21, ac 21

5
Desired Properties
  • Condorcet Criterion (contd.)
  • Condorcet Criterion
  • Condorcet alternative wins (if one exists)
  • cycle possible
  • ab 21, bc 21, ca 21

6
Desired Properties
  • Consistency
  • If the voters are divided into 2 parts and
    separate voting in the two parts get the same
    winner, then the overall voting result should be
    the same
  • Independence from Irrelevant Alternatives
  • X gt Y (overall), XZY ? ZXY (one voter), still
    XgtY (overall)
  • Monotonicity
  • X wins (overall), YZX ? YXZ (one voter), still X
    wins (overall)
  • If only Xs rank is raised.

7
Desired Properties
  • Pareto Optimality
  • X gt Y (all individual orderings) ? XgtY (overall)
  • Neutrality
  • Should not favor any alternative
  • Manipulability
  • insincere voting

8
Manipulability ?
  • Gibbard (1973) and Satterthwaite (1975)
  • Any non-dictatorial voting systems with at least
    three possible outcomes is subject to individual
    manipulation.
  • Although impossible
  • Some systems tend to encourage manipulation more
    than other.
  • It is easier to identify manipulation strategies
    in some systems than others

9
Arrows Impossibility Theorem
  • KENNETH J. ARROW Social Choice and Individual
    Values (1951). Won Nobel Prize in 1972
  • No voting scheme over three or more alternatives
    can satisfies all of several reasonable and
    logical conditions.
  • Universality
  • no restriction on individual ordering. All
    orderings are achievable.
  • Non-dictatorship
  • Monotonicity
  • Independence of irrelevant alternatives
  • Pareto Optimality

10
Rank Aggregation Methods
  • Majority Rule and Majoritarian Methods
  • Positional Methods
  • Utilitarian Methods
  • Multi-Stage Methods

11
Majority Rule and Majoritarian Methods
  • Use only information from pariwise comparisons
    between candidates
  • Majority Rule (2 alternatives)
  • The one preferred by most voters is selected.
  • Majoritarian Methods (gt2 alternatives)
  • select the Condorcet winner if one exists
  • Various Majoritarian Methods (no Condorcet
    winner)
  • The Amendment Procedure
  • The Successive Procedure

12
Positional Methods
  • Not pairwise comparisons, but orderings.
  • Plurality Voting (only top-1 alternative)
  • Approval Voting (only top-k alternatives)
  • Borda Count (entire ordering)

13
Borda Count
14
Borda Count
Condorcet Winner a (ab 32) Borda Winner b
(a55441, b55443)
15
Borda Count
Condorcet Winner a (ab 32) Borda Winner a
(a55444, b55443) Violate Independence
from Irrelevant Alternatives
16
Borda Count
  • Borda (1770)
  • Member of French Academy of Sciences
  • Noted for work in hydraulics, optics, navigation
    instrument
  • Purpose Reforming the election procedure of
    French Academy.
  • Criticize plurality method
  • Independency violated
  • Condorcet (1785)
  • Viewd Borda as enemy
  • Finding best ordering by hypothesis testing
  • Switch to propose Condorcet winner
  • Cycles ? eliminate candidates ? monotonicity
    violated
  • (Systems that eliminate before selecting violate
    monotonicity)

17
Utilitarian Methods
  • Require voters to assign utilities to
    alternatives
  • Majoritarian pairwise xgty
  • Positional position xyz ? 321
  • Utilitarian scale x0.5, y0.2, z0.1,
    x0.5, y0.49, z0.1
  • Direct Aggregation of Cardinal Utility Methods
  • summation, multiplication
  • Demand-Revealing Methods

18
Demand-Revealing Methods
  • Tideman and Tullock (1976)
  • Each voter offers to pay money m to obtain a
    preferred alternative.
  • The alternative with the largest sum of offered
    money wins.
  • Voters need to pay tax proportional to their
    contribution if their choices win.
  • This tax is wasted(!) or given to people not
    involved.

19
Multi-Stage Methods
  • Black's Method (1940)
  • Condorcet Winner, if one exists, otherwise Borda
    count.
  • The Runoff Election Procedure
  • Nanson's Borda-Elimination Procedure
  • Single Transferable Voting
  • Coomb's Procedure

20
Condorcets intention to order
  • 3 candidates, 13 voters
  • ordering abc has support 26
  • agtb8, agtc7, bgtc11
  • ordering bca has support 22
  • altb5, cgta6, bgtc11

21
Kemeny Optimal Aggregation
  • John Kemeny 1959 (developed BASIC language)
  • Satisfies Condorcet criterion
  • Equal to Condorcets ordering

22
Kemeny Optimal Aggregation
  • Kemeny distance
  • Number of different pairwise orders.
  • Example
  • 2 ranked list r1abc, r2bac
  • K(r1,r2) 1 (a gtr1 b, a ltr2 b)
  • Kemeny Optimal Aggregation
  • ranked list r1, r2, , rn
  • aggreate ranked list r argmin ? K(r,ri)
  • NP-hard

23
Application
  • Web Search
  • Machine Learning
  • Information Retrieval
  • Database Middleware

24
Web Search
  • Rank Aggregation Methods for the Web. Cynthia
    Dwork, Ravi Kumar, Moni Naor, D. Sivakumar.
    WWW10.
  • Spam-fighting in Meta-Search
  • Local Kemenization
  • efficient, satisfy Condorcet criterion.

25
Conclusion
  • Desired Properties
  • Rank Aggregation Methods
  • Condorcet Criterion vs. Borda
  • Kemeny Optimal Aggregation
  • Rank Aggregation in fields other than social
    choice
Write a Comment
User Comments (0)
About PowerShow.com