Elicitation for Preferences Single
Peaked on Trees
Palash Dey and Neeldhara Misra
*IJCAI 2016
and the query model for eliciting preferences.
The standard Voting Setup
and better query complexities for elicitation.
Single-peaked Preferences
and the query model for eliciting preferences.
The standard Voting Setup
and better query complexities for elicitation.
so we consider broader domains but still have nice properties.
GENERALIZING “SINGLE PEAKEDNESS”
Single-peaked Preferences
and the query model for eliciting preferences.
The standard Voting Setup
and better query complexities for elicitation.
so we consider broader domains but still have nice properties.
for preferences single-peaked on trees.
GENERALIZING “SINGLE PEAKEDNESS”
Back to QUERY COMPLEXITY
Single-peaked Preferences
and the query model for eliciting preferences.
The standard Voting Setup
The standard Voting Setup
and the query model for eliciting preferences.
Candidates/Alternatives
Voters express their preferences over alternatives
(here, as rankings).
When the number of candidates is large,
soliciting a full ranking can be a little unmanageable.
When the number of candidates is large,
soliciting a full ranking can be a little unmanageable.
Voters typically find it easier to answer
“comparison queries”:
Would you rather hang out over coffee
or join me for a concert?
How many such queries to do we need to make
to be able to reconstruct the full preference?
How many such queries to do we need to make
to be able to reconstruct the full preference?
Just like the weighing scale puzzles, except you can
only compare two single options at a time.
Using a “merge sort” like idea,
O(m log m) comparisons are enough.
Using a “merge sort” like idea,
O(m log m) comparisons are enough.
Recursively order half the alternatives.
Using a “merge sort” like idea,
O(m log m) comparisons are enough.
Recursively order half the alternatives.
Using a “merge sort” like idea,
O(m log m) comparisons are enough.
Merge the two lists with a linear number of queries.
21 25 32 41 55 63 72
10 22 45 53 62 64 90
21 25 32 41 55 63 72
10 22 45 53 62 64 90
21 25 32 41 55 63 72
10
22 45 53 62 64 90
21
25 32 41 55 63 72
10
22 45 53 62 64 90
21
25 32 41 55 63 72
10 22
45 53 62 64 90
21 25
32 41 55 63 72
10 22
45 53 62 64 90
21 25 32
41 55 63 72
10 22
45 53 62 64 90
21 25 32 41
55 63 72
10 22
45 53 62 64 90
21 25 32 41
55 63 72
10 22 45
53 62 64 90
21 25 32 41
55 63 72
10 22 45
53
62 64 90
21 25 32 41
55 63 72
10 22 45
53 62 64 90
Using a “merge sort” like idea,
O(m log m) comparisons are enough.
Merge the two lists linear number of queries.
Single Peaked Preferences
and better query complexities for elicitation. [1]
Left RightCenter
A B C D E F G
Left RightCenter
A B C D E F G
Left RightCenter
A B C D E F G
E D C F G B A
Left RightCenter
A B C D E F G
E D C F G B A
E D CF GB A
Left RightCenter
A B C D E F G
Left RightCenter
A B C D E F G
Left RightCenter
A B C D E F G
If an agent with single-peaked preferences prefers x to y,
one of the following must be true:
- x is the agent’s peak,
- x and y are on opposite sides of the agent’s peak, or
- x is closer to the peak than y.
Left RightCenter
A B C D E F G
The notion is popular for several reasons:
- No Condorcet Cycles.
- No incentive for an agent to misreport its preferences.
- Identifiable in polynomial time.
- Reasonable (?) model of actual elections.
Left RightCenter
A B C D E F G
Left RightCenter
A B C D E F G
…and better query complexities for elicitation.
Left RightCenter
A B C D E F G
…and better query complexities for elicitation.
(i) Identify the peak: use binary search.
Left RightCenter
A B C D E F G
(i) Identify the peak: use binary search.
[Better query complexities for elicitation.]
Left RightCenter
A B C D
E
F G
(i) Identify the peak: use binary search.
[Better query complexities for elicitation.]
Left RightCenter
A B C D
E
F G
(i) Identify the peak: use binary search.
this signals that the
peak is to the right.
[Better query complexities for elicitation.]
Left RightCenter
A B C D E F G
(i) Identify the peak: use binary search.
[Better query complexities for elicitation.]
Left RightCenter
A B C
D
E F G
(i) Identify the peak: use binary search.
[Better query complexities for elicitation.]
Left RightCenter
A B C
D
E F G
(i) Identify the peak: use binary search.
this signals that the
peak is to the left.
[Better query complexities for elicitation.]
Left RightCenter
A B C D E F G
(ii) Once we know the peak, the rest is O(m) queries.
[Better query complexities for elicitation.]
Left Right
A B C D E F G
(ii) Once we know the peak, the rest is O(m) queries.
[Better query complexities for elicitation.]
Left Right
A B C D E F G
(ii) Once we know the peak, the rest is O(m) queries.
C
[Better query complexities for elicitation.]
Left Right
A B C D E F G
(ii) Once we know the peak, the rest is O(m) queries.
C
[Better query complexities for elicitation.]
Left Right
A B C D E F G
(ii) Once we know the peak, the rest is O(m) queries.
C
[Better query complexities for elicitation.]
Left Right
A
B
C D E F G
(ii) Once we know the peak, the rest is O(m) queries.
C
[Better query complexities for elicitation.]
Left Right
A
B
C D E F G
(ii) Once we know the peak, the rest is O(m) queries.
C
[Better query complexities for elicitation.]
Left Right
A B C D E F G
(ii) Once we know the peak, the rest is O(m) queries.
C
[Better query complexities for elicitation.]
Left Right
A B C
D
E F G
(ii) Once we know the peak, the rest is O(m) queries.
C
[Better query complexities for elicitation.]
Left Right
A B C
D
E F G
(ii) Once we know the peak, the rest is O(m) queries.
C
[Better query complexities for elicitation.]
Left Right
A B C D E F G
Center
[Better query complexities for elicitation.]
Left Right
A B C D E F G
Center
(ii) Once we know the peak, the rest is O(m) queries.
(i) Identify the peak: use binary search (O(log m) queries).
[Better query complexities for elicitation.]
GeneraliZing “single-Peakedness”
so we consider broader domains but still have nice properties. [2]
Profiles single-peaked on trees
A preference profile is single-peaked with respect to a
tree T (labeled with the alternatives)
if it is single peaked when restricted to the candidates
involved in any path on the tree.
Profiles single-peaked on trees
A preference profile is single-peaked with respect to a
tree T (labeled with the alternatives)
if it is single peaked when restricted to the candidates
involved in any path on the tree.
Profiles single-peaked on trees
A preference profile is single-peaked with respect to a
tree T (labeled with the alternatives)
if it is single peaked when restricted to the candidates
involved in any path on the tree.
Profiles single-peaked on trees
A preference profile is single-peaked with respect to a
tree T (labeled with the alternatives)
if it is single peaked when restricted to the candidates
involved in any path on the tree.
The special case when T is a path corresponds to the
usual notion of single-peakedness.
Profiles single-peaked on trees
A preference profile is single-peaked with respect to a
tree T (labeled with the alternatives)
if it is single peaked when restricted to the candidates
involved in any path on the tree.
Profiles single-peaked on trees
Are these structures still “nice enough”, like single-peaked profiles?
Profiles single-peaked on trees
Are these structures still “nice enough”, like single-peaked profiles?
- No Condorcet Cycles.
- No incentive for an agent to misreport its preferences.
- Identifiable in polynomial time.
- Reasonable (?) model of actual elections.
Back to elicitation
for preferences single-peaked on trees.
The main idea:
- Split off the tree into paths.
- Elicit along each path.
- Merge the information across paths.
The main idea:
- Split off the tree into paths.
- Elicit along each path.
- Merge the information across paths.
If the number of leaves is t, then the tree can be
partitioned into t disjoint paths.
The main idea:
- Split off the tree into paths.
- Elicit along each path.
- Merge the information across paths.
This requires O(|Pi|) queries for the path Pi. Recall that the
paths constitute a partition of m nodes - so the overall
time here is O(m).
The main idea:
- Split off the tree into paths.
- Elicit along each path.
- Merge the information across paths.
Merging t lists of m elements in total can be done in time
O(m log t), using divide and conquer.
The main idea:
- Split off the tree into paths.
- Elicit along each path.
- Merge the information across paths.
Overall queries: O(m log t)
Does your profile have “d outliers” to single-peakedness?
Does your profile have “d outliers” to single-peakedness?
O(d log d) + O(m — d) + O(m)
Does your profile have “d outliers” to single-peakedness?
O(d log d) + O(m — d) + O(m)
the usual merge sort
Does your profile have “d outliers” to single-peakedness?
O(d log d) + O(m — d) + O(m)
the usual merge sort
the single-peaked algorithm
Does your profile have “d outliers” to single-peakedness?
O(d log d) + O(m — d) + O(m)
the usual merge sort
the single-peaked algorithm
(putting the outputs together)
A quick Lower Bound
for preferences single-peaked on trees.
Any path that doesn’t involve the root goes along one of the spines.
The number of orderings that are single peaked over
the subdivided star of depth d on t leaves: (t!)d
The number of orderings that are single peaked over
the subdivided star of depth d on t leaves: (t!)d
A standard adversary argument gives the following
lower bound: log((t!)d) = dt log t = m log t
1. Vincent Conitzer. Eliciting single-peaked preferences
using comparison queries.
J. Artif. Intell. Res., 35:161–191, 2009.
2. Gabrielle Demange. Single-peaked orders on a tree.
Math. Soc. Sci, 3(4):389–396, 1982.
ReferenceS
Thank you!

Elicitation for Preferences Single Peaked on Trees

  • 1.
    Elicitation for PreferencesSingle Peaked on Trees Palash Dey and Neeldhara Misra *IJCAI 2016
  • 2.
    and the querymodel for eliciting preferences. The standard Voting Setup
  • 3.
    and better querycomplexities for elicitation. Single-peaked Preferences and the query model for eliciting preferences. The standard Voting Setup
  • 4.
    and better querycomplexities for elicitation. so we consider broader domains but still have nice properties. GENERALIZING “SINGLE PEAKEDNESS” Single-peaked Preferences and the query model for eliciting preferences. The standard Voting Setup
  • 5.
    and better querycomplexities for elicitation. so we consider broader domains but still have nice properties. for preferences single-peaked on trees. GENERALIZING “SINGLE PEAKEDNESS” Back to QUERY COMPLEXITY Single-peaked Preferences and the query model for eliciting preferences. The standard Voting Setup
  • 6.
    The standard VotingSetup and the query model for eliciting preferences.
  • 7.
  • 8.
    Voters express theirpreferences over alternatives (here, as rankings).
  • 9.
    When the numberof candidates is large, soliciting a full ranking can be a little unmanageable.
  • 10.
    When the numberof candidates is large, soliciting a full ranking can be a little unmanageable. Voters typically find it easier to answer “comparison queries”: Would you rather hang out over coffee or join me for a concert?
  • 11.
    How many suchqueries to do we need to make to be able to reconstruct the full preference?
  • 12.
    How many suchqueries to do we need to make to be able to reconstruct the full preference? Just like the weighing scale puzzles, except you can only compare two single options at a time.
  • 13.
    Using a “mergesort” like idea, O(m log m) comparisons are enough.
  • 14.
    Using a “mergesort” like idea, O(m log m) comparisons are enough. Recursively order half the alternatives.
  • 15.
    Using a “mergesort” like idea, O(m log m) comparisons are enough. Recursively order half the alternatives.
  • 16.
    Using a “mergesort” like idea, O(m log m) comparisons are enough. Merge the two lists with a linear number of queries.
  • 17.
    21 25 3241 55 63 72 10 22 45 53 62 64 90
  • 18.
    21 25 3241 55 63 72 10 22 45 53 62 64 90
  • 19.
    21 25 3241 55 63 72 10 22 45 53 62 64 90
  • 20.
    21 25 32 4155 63 72 10 22 45 53 62 64 90
  • 21.
    21 25 32 4155 63 72 10 22 45 53 62 64 90
  • 22.
    21 25 32 4155 63 72 10 22 45 53 62 64 90
  • 23.
    21 25 32 4155 63 72 10 22 45 53 62 64 90
  • 24.
    21 25 3241 55 63 72 10 22 45 53 62 64 90
  • 25.
    21 25 3241 55 63 72 10 22 45 53 62 64 90
  • 26.
    21 25 3241 55 63 72 10 22 45 53 62 64 90
  • 27.
    21 25 3241 55 63 72 10 22 45 53 62 64 90
  • 28.
    Using a “mergesort” like idea, O(m log m) comparisons are enough. Merge the two lists linear number of queries.
  • 29.
    Single Peaked Preferences andbetter query complexities for elicitation. [1]
  • 30.
  • 31.
  • 32.
    Left RightCenter A BC D E F G E D C F G B A
  • 33.
    Left RightCenter A BC D E F G E D C F G B A E D CF GB A
  • 34.
  • 35.
  • 36.
    Left RightCenter A BC D E F G If an agent with single-peaked preferences prefers x to y, one of the following must be true: - x is the agent’s peak, - x and y are on opposite sides of the agent’s peak, or - x is closer to the peak than y.
  • 37.
    Left RightCenter A BC D E F G The notion is popular for several reasons: - No Condorcet Cycles. - No incentive for an agent to misreport its preferences. - Identifiable in polynomial time. - Reasonable (?) model of actual elections.
  • 38.
  • 39.
    Left RightCenter A BC D E F G …and better query complexities for elicitation.
  • 40.
    Left RightCenter A BC D E F G …and better query complexities for elicitation. (i) Identify the peak: use binary search.
  • 41.
    Left RightCenter A BC D E F G (i) Identify the peak: use binary search. [Better query complexities for elicitation.]
  • 42.
    Left RightCenter A BC D E F G (i) Identify the peak: use binary search. [Better query complexities for elicitation.]
  • 43.
    Left RightCenter A BC D E F G (i) Identify the peak: use binary search. this signals that the peak is to the right. [Better query complexities for elicitation.]
  • 44.
    Left RightCenter A BC D E F G (i) Identify the peak: use binary search. [Better query complexities for elicitation.]
  • 45.
    Left RightCenter A BC D E F G (i) Identify the peak: use binary search. [Better query complexities for elicitation.]
  • 46.
    Left RightCenter A BC D E F G (i) Identify the peak: use binary search. this signals that the peak is to the left. [Better query complexities for elicitation.]
  • 47.
    Left RightCenter A BC D E F G (ii) Once we know the peak, the rest is O(m) queries. [Better query complexities for elicitation.]
  • 48.
    Left Right A BC D E F G (ii) Once we know the peak, the rest is O(m) queries. [Better query complexities for elicitation.]
  • 49.
    Left Right A BC D E F G (ii) Once we know the peak, the rest is O(m) queries. C [Better query complexities for elicitation.]
  • 50.
    Left Right A BC D E F G (ii) Once we know the peak, the rest is O(m) queries. C [Better query complexities for elicitation.]
  • 51.
    Left Right A BC D E F G (ii) Once we know the peak, the rest is O(m) queries. C [Better query complexities for elicitation.]
  • 52.
    Left Right A B C DE F G (ii) Once we know the peak, the rest is O(m) queries. C [Better query complexities for elicitation.]
  • 53.
    Left Right A B C DE F G (ii) Once we know the peak, the rest is O(m) queries. C [Better query complexities for elicitation.]
  • 54.
    Left Right A BC D E F G (ii) Once we know the peak, the rest is O(m) queries. C [Better query complexities for elicitation.]
  • 55.
    Left Right A BC D E F G (ii) Once we know the peak, the rest is O(m) queries. C [Better query complexities for elicitation.]
  • 56.
    Left Right A BC D E F G (ii) Once we know the peak, the rest is O(m) queries. C [Better query complexities for elicitation.]
  • 57.
    Left Right A BC D E F G Center [Better query complexities for elicitation.]
  • 58.
    Left Right A BC D E F G Center (ii) Once we know the peak, the rest is O(m) queries. (i) Identify the peak: use binary search (O(log m) queries). [Better query complexities for elicitation.]
  • 59.
    GeneraliZing “single-Peakedness” so weconsider broader domains but still have nice properties. [2]
  • 60.
  • 61.
    A preference profileis single-peaked with respect to a tree T (labeled with the alternatives) if it is single peaked when restricted to the candidates involved in any path on the tree. Profiles single-peaked on trees
  • 65.
    A preference profileis single-peaked with respect to a tree T (labeled with the alternatives) if it is single peaked when restricted to the candidates involved in any path on the tree. Profiles single-peaked on trees
  • 66.
    A preference profileis single-peaked with respect to a tree T (labeled with the alternatives) if it is single peaked when restricted to the candidates involved in any path on the tree. Profiles single-peaked on trees
  • 67.
    A preference profileis single-peaked with respect to a tree T (labeled with the alternatives) if it is single peaked when restricted to the candidates involved in any path on the tree. The special case when T is a path corresponds to the usual notion of single-peakedness. Profiles single-peaked on trees
  • 68.
    A preference profileis single-peaked with respect to a tree T (labeled with the alternatives) if it is single peaked when restricted to the candidates involved in any path on the tree. Profiles single-peaked on trees Are these structures still “nice enough”, like single-peaked profiles?
  • 69.
    Profiles single-peaked ontrees Are these structures still “nice enough”, like single-peaked profiles? - No Condorcet Cycles. - No incentive for an agent to misreport its preferences. - Identifiable in polynomial time. - Reasonable (?) model of actual elections.
  • 70.
    Back to elicitation forpreferences single-peaked on trees.
  • 71.
    The main idea: -Split off the tree into paths. - Elicit along each path. - Merge the information across paths.
  • 72.
    The main idea: -Split off the tree into paths. - Elicit along each path. - Merge the information across paths. If the number of leaves is t, then the tree can be partitioned into t disjoint paths.
  • 73.
    The main idea: -Split off the tree into paths. - Elicit along each path. - Merge the information across paths. This requires O(|Pi|) queries for the path Pi. Recall that the paths constitute a partition of m nodes - so the overall time here is O(m).
  • 74.
    The main idea: -Split off the tree into paths. - Elicit along each path. - Merge the information across paths. Merging t lists of m elements in total can be done in time O(m log t), using divide and conquer.
  • 75.
    The main idea: -Split off the tree into paths. - Elicit along each path. - Merge the information across paths. Overall queries: O(m log t)
  • 76.
    Does your profilehave “d outliers” to single-peakedness?
  • 77.
    Does your profilehave “d outliers” to single-peakedness? O(d log d) + O(m — d) + O(m)
  • 78.
    Does your profilehave “d outliers” to single-peakedness? O(d log d) + O(m — d) + O(m) the usual merge sort
  • 79.
    Does your profilehave “d outliers” to single-peakedness? O(d log d) + O(m — d) + O(m) the usual merge sort the single-peaked algorithm
  • 80.
    Does your profilehave “d outliers” to single-peakedness? O(d log d) + O(m — d) + O(m) the usual merge sort the single-peaked algorithm (putting the outputs together)
  • 81.
    A quick LowerBound for preferences single-peaked on trees.
  • 83.
    Any path thatdoesn’t involve the root goes along one of the spines.
  • 89.
    The number oforderings that are single peaked over the subdivided star of depth d on t leaves: (t!)d
  • 90.
    The number oforderings that are single peaked over the subdivided star of depth d on t leaves: (t!)d A standard adversary argument gives the following lower bound: log((t!)d) = dt log t = m log t
  • 92.
    1. Vincent Conitzer.Eliciting single-peaked preferences using comparison queries. J. Artif. Intell. Res., 35:161–191, 2009. 2. Gabrielle Demange. Single-peaked orders on a tree. Math. Soc. Sci, 3(4):389–396, 1982. ReferenceS
  • 93.