1 differences between observed and latent confidence in rank ordering brent j. miller mark steyvers...
TRANSCRIPT
1
Differences Between Observed and Latent Confidence in Rank Ordering
Brent J. MillerMark Steyvers
University of California, Irvine
2
Getting Expert Information
3
Getting Expert Information
Delphi Method Iterative Collaboration
4
George Washington
John Adams
James Monroe
James Madison
Thomas Jefferson
John Adams
James Madison
James Monroe
Independent Rank-Ordering Tasks
Order these U.S. Presidents, by term of office:
George Washington
1st Participant 2nd Participant 3rd Participant
John Adams
George Washington
Thomas Jefferson
James Madison
James Monroe
Thomas Jefferson
5
James Madison
Thomas Jefferson
Thomas Jefferson
John Adams
James Madison
James Monroe
James Monroe
Iterative Rank-Ordering Tasks
Order these U.S. Presidents, by term of office:
George Washington
1st ParticipantGeorge Washington
John Adams
George Washington
James Madison
Thomas Jefferson
James Monroe
2nd Participant 3rd Participant
John Adams
6
• 100 Subjects• 17 Ranking Questions
– Chronological Events (Holidays, U.S. Presidents)– Physical Measures (Landmass, City Pop., State Locations)– Ordinal Information (Ten Amendments, Ten Commandments)
• 2 Conditions
Rank-Ordering Experiment
7
Rank Ordering – Independent & Iteratives
5 10 15 20 25 30 35 40 45 504
6
8
10
12
14
16
18
Group Size
Ave
rage
Tau
17 Questions via Mechanical Turk
5 10 15 20 25 30 35 40 45 506
7
8
9
10
11
12
13
14
Group Size
Ave
rage
Tau
17 Questions from University Students
x 4
88
Participant1 Y1: A < B < CX11 X12 X13
A B C
Thurstonian Model for Rank Ordering
σ1
99
Participant1 Y1: A < B < CX11 X12 X13
Y2 : A < C < BParticipant2 X22 X23X21
A B C
σ2
Thurstonian Model for Rank Ordering
10
Iterative Collaboration
George Washington
John Adams
Thomas Jefferson
James Monroe
Andrew Jackson
Theodore Roosevelt
Woodrow Wilson
Franklin D. Roosevelt
Harry S. Truman
Dwight D. Eisenhower
76 77 78 79 80 81 correct
1
2
3
4
5
6
7
8
9
10
Participants
Pos
ition
U.S. Presidents
33 34 35 36 37 38 39
Participants
Posi
tion
11
Participant1 Y1: A < B < CX11 X12 X13
A B C
Thurstonian Model for Item Representation
σ11 σ12 σ13
12
• 50 Subjects• 8 Ranking Questions
– Chronological Events (Holidays, U.S. Presidents)– Physical Measures (Landmass, City Pop., State Locations)– Ordinal Information (Ten Amendments, Ten Commandments)
Rank-Ordering Experiment
7 5 6 4 1 8 3 2 7 5 6 4 1 8 3 2
Independent (8) Iterative (8)
13
Independent vs. Iterative Aggregate Performance
5 10 15 20 25 30 35 40 457
8
9
10
11
12
13
14
15
16
Group Size
Ave
rage
Tau
Independent
Iterative
14
Independent Subject Generative Model
μi
xij σij
yj
i Items
j Subjects
µi ≡ Ground truthσ2
ij ~ InvGamma(.2,1)xi ~ Norm(µi,σ2
ij)Yj ~ Rank(xj)
15
Iterative Subject Generative Model
μij
xi,j-1
σi,j-1
yj-1
i Items
j Subjects
xij
σij
cutyj
µi ≡ Ground truthσ2
ij ~ InvGamma(.2,1)Xij-1 ~ Norm(µi,σ2
ij-1)Yj-1 ~ Rank(xj-1)
16
Simulated Aggregate Performance
1 10 20 30 40 50 652
4
6
8
10
12
14
16
Group Size
Ave
rag
et
Independent Subjects
Iterative Subjects
17
Reduction in Iterative Error by Item Confidence
−0.4 −0.05 0.3 0.65 1
1
2
3
4
5
DError
Sub
ject
Con
fide
nce
Confidence for Items with Error > 4 Tau
No Change
18
Conclusions
• Item-level knowledge may allow partial expertise between participants, allowing for improvement in group aggregation.
• Subject item-level confidence appears to support the validity of our proposed model.
• We should be able to use subjects’ confidence as an estimate of underlying item-level expertise.
19
Danke!
Differences Between Observed and Latent Confidence in Rank Ordering