Active Learning for Matching Problems
Laurent Charlin, Richard Zemel and Craig Boutilier
Prediction of
missing scores
(M)
232
012
112
232
012
112
Final objectives
and constraints
Elicited & predicted
scores
Match Matrix
Our contributions:
- Matching-aware active learning methods
- Probabilistic matching procedure
(P) Score Prediction
Matching Max-score query
3
2
Difference in matching obj.
250
5000
200
4800
4600
150
4400
100
4200
50
4000
0
3800
−50
0
50
100
150
Queries
Queries
200
250
100
80
60
40
20
0
3600
−100
−50
2
3
Num. rand. queries
Absolute Matching Objective
Elicited Scores
(P)
Matching Performance
- Run the LP (M):
- Value of a match:
0
3400
300
100
200
300
50
100
150
Queries
4
Any score prediction model can be used
-Assume model keeps a distr. over
Matching Max-entropy query
700
232
012
112
232
012
112
4.8
600
:
(M) Matching Problem
x 10
4.9
800
Dating
Items
3
S
1
1
2
Query the scores that will be most helpful for
matching:
- Active learning methods should take the
matching objective ( ) into account
- Computing EVOI is expensive
- We propose methods that operate
in matching space
4.6
400
300
4.5
200
4.4
4.2
−100
- Simulate elicitation procedure
- Every user starts with a few obs. scores
- At each round each user is queried (batch)
- Bayesian PMF is used to predict scores
−200
−20
3000
2000
1000
4.3
0
Experimental Setup
4000
4.7
500
100
0
20
40
60
Queries
80
100
120
200
Conference
Elicitation
Users
(E)
Experimental Results
Jokes
- Learning user preferences
eases user elicitation burden
- Active learning provides a further
improvement
(E) Elicitation: Active Learning
Methods
Difference in Matching Objective
Problem: Recommend items to users under
matching constraints
0
0
4.1
140
3000
2800
150
2600
1000
800
100
2400
600
400
- Data
Suitabilities
200
- Jokes (Jester):
0
0
50
100
150
Taking predicted score uncertainty into account:
- 300 users, 10 jokes
- Match 10 to 30 users per joke
- Matching-aware active learning is a win
- Dating (LibimSeTi.cz):
- Intractable but we use sampling:
- Further results: our approach is robust to
- 250 users, 250 items (users)
matching constraints, batch sizes.
- Match 15 to 25 users per item
- Conference (NIPS'10):
Conclusion: Effective for determining high-quality
- 1250 papers, 48 revs
matches
with significantly less elicitation
- Match 20 to 30 papers per rev.
Current work: EVOI extensions, different types of
Constraints:
Number of scores
200
50
2200
0
2000
150
−50
−20
100
0
20
40
60
Queries
80
100
1800
120
50
0
−10
−5
0
5
6000
5000
4000
3000
2000
1000
0
1
2
3
4
5
6
7
8
9 10
3000
Number of Scores
2500
2000
1500
1000
500
- 2 items per user
- 1 user per item
0
0
1
2
3
queries (side-information, higher level)