near-optimal batch mode active learning and adaptive
TRANSCRIPT
Near-optimal Batch Mode Active Learning and Adaptive Submodular Optimization
Yuxin Chen Andreas KrauseDepartment of Computer Science, ETH Zurich
Batch Mode Active Learning
w
?
?
?
?
?
?
?
? ?
. . .
. . .
?
?
. . .
. . .
?
. . .
?
Multi-stage Influence Maximization in Social Networks
How should we construct the batches?
B
A
s
. . .
THE BatchGreedy ALGORITHM
B
A
s
. . .
Conditional marginal benefit of an item s:
Expectation over all realizations
within the batch
Conditioning on previous
observations
�f (s | A,yB) = EyV
⇥f(y{s}[A[B)� f(yA[B) | yB
⇤.
THE BatchGreedy ALGORITHM
B
A
s
. . .
si,j = argmax
s2V�f (s | {s1,j , . . . , si�1,j}| {z }
the jth batch A
,yB)
BatchGreedy will greedily select the i-th element in the j-th batch
BatchGreedy VS. OPTIMAL BATCH
Cost of BatchGreedy Cost of optimal batch policy
. . .
. . . . . .
�����
�����
�����
�����
BatchGreedy VS. OPTIMAL BATCH
Cost of BatchGreedy Cost of optimal batch policy
. . .
. . . . . .
�����
�����
�����
����� O (lnQ) ·
How many extra items will we select?
BatchGreedy VS. SEQUENTIAL
. . .
Cost of BatchGreedy Cost of optimal sequential policy
. . .
. . .
�����
�����
�����
�����
BatchGreedy VS. SEQUENTIAL
. . .
Cost of BatchGreedy Cost of optimal sequential policy
. . .
. . .
�����
�����
�����
�����
COMPETITIVE
EXPERIMENTAL RESULTS
# POSTER
w
?
. . .
. . .
EXPERIMENTAL RESULTS
# POSTER
w
?
. . .
. . .
10 20 30 40 500
2
4
6
8
10
12
14
Number of labels requested
% M
ista
kes
EXPERIMENTAL RESULTS
# POSTER
w
?
. . .
. . .
10 20 30 40 500
2
4
6
8
10
12
14
Number of labels requested
% M
ista
kes
10 20 30 40 500
2
4
6
8
10
12
14
random
EXPERIMENTAL RESULTS
# POSTER
w
?
. . .
. . .
10 20 30 40 500
2
4
6
8
10
12
14
Number of labels requested
% M
ista
kes
10 20 30 40 500
2
4
6
8
10
12
14
random
10 20 30 40 500
2
4
6
8
10
12
14
sequential
EXPERIMENTAL RESULTS
# POSTER
w
?
. . .
. . .
10 20 30 40 500
2
4
6
8
10
12
14
Number of labels requested
% M
ista
kes
10 20 30 40 500
2
4
6
8
10
12
14
random
10 20 30 40 500
2
4
6
8
10
12
14
KLR−BMAL
10 20 30 40 500
2
4
6
8
10
12
14
sequential
EXPERIMENTAL RESULTS
# POSTER
w
?
. . .
. . .
10 20 30 40 500
2
4
6
8
10
12
14
Number of labels requested
% M
ista
kes
10 20 30 40 500
2
4
6
8
10
12
14
10−batch greedy
10 20 30 40 500
2
4
6
8
10
12
14
random
10 20 30 40 500
2
4
6
8
10
12
14
KLR−BMAL
10 20 30 40 500
2
4
6
8
10
12
14
sequential
EXPERIMENTAL RESULTS
# POSTER
w
?
. . .
. . .
10 20 30 40 500
2
4
6
8
10
12
14
Number of labels requested
% M
ista
kes
10 20 30 40 500
2
4
6
8
10
12
14
10−batch greedy
10 20 30 40 500
2
4
6
8
10
12
14
random
10 20 30 40 500
2
4
6
8
10
12
14
KLR−BMAL
10 20 30 40 500
2
4
6
8
10
12
14
sequential
EXPERIMENTAL RESULTS
# POSTER
w
?
. . .
. . .
10 20 30 40 500
2
4
6
8
10
12
14
Number of labels requested
% M
ista
kes
10 20 30 40 500
2
4
6
8
10
12
14
10−batch greedy
10 20 30 40 500
2
4
6
8
10
12
14
random
10 20 30 40 500
2
4
6
8
10
12
14
KLR−BMAL
10 20 30 40 500
2
4
6
8
10
12
14
sequential
0 200 400 600 8000
5
10
15
20
25
30
35
% it
em n
ot c
over
ed
Number of items selected
EXPERIMENTAL RESULTS
# POSTER
w
?
. . .
. . .
10 20 30 40 500
2
4
6
8
10
12
14
Number of labels requested
% M
ista
kes
10 20 30 40 500
2
4
6
8
10
12
14
10−batch greedy
10 20 30 40 500
2
4
6
8
10
12
14
random
10 20 30 40 500
2
4
6
8
10
12
14
KLR−BMAL
10 20 30 40 500
2
4
6
8
10
12
14
sequential
0 200 400 600 8000
5
10
15
20
25
30
35
% it
em n
ot c
over
ed
Number of items selected
0 200 400 600 8000
5
10
15
20
25
30
35
Non−adaptive
EXPERIMENTAL RESULTS
# POSTER
w
?
. . .
. . .
10 20 30 40 500
2
4
6
8
10
12
14
Number of labels requested
% M
ista
kes
10 20 30 40 500
2
4
6
8
10
12
14
10−batch greedy
10 20 30 40 500
2
4
6
8
10
12
14
random
10 20 30 40 500
2
4
6
8
10
12
14
KLR−BMAL
10 20 30 40 500
2
4
6
8
10
12
14
sequential
0 200 400 600 8000
5
10
15
20
25
30
35
% it
em n
ot c
over
ed
Number of items selected
0 200 400 600 8000
5
10
15
20
25
30
35
sequential
0 200 400 600 8000
5
10
15
20
25
30
35
Non−adaptive
EXPERIMENTAL RESULTS
# POSTER
w
?
. . .
. . .
10 20 30 40 500
2
4
6
8
10
12
14
Number of labels requested
% M
ista
kes
10 20 30 40 500
2
4
6
8
10
12
14
10−batch greedy
10 20 30 40 500
2
4
6
8
10
12
14
random
10 20 30 40 500
2
4
6
8
10
12
14
KLR−BMAL
10 20 30 40 500
2
4
6
8
10
12
14
sequential
0 200 400 600 8000
5
10
15
20
25
30
35
% it
em n
ot c
over
ed
Number of items selected
0 200 400 600 8000
5
10
15
20
25
30
35
sequential
0 200 400 600 8000
5
10
15
20
25
30
35
10−batch
0 200 400 600 8000
5
10
15
20
25
30
35
Non−adaptive
EXPERIMENTAL RESULTS
# POSTER
w
?
. . .
. . .
10 20 30 40 500
2
4
6
8
10
12
14
Number of labels requested
% M
ista
kes
10 20 30 40 500
2
4
6
8
10
12
14
10−batch greedy
10 20 30 40 500
2
4
6
8
10
12
14
random
10 20 30 40 500
2
4
6
8
10
12
14
KLR−BMAL
10 20 30 40 500
2
4
6
8
10
12
14
sequential
0 200 400 600 8000
5
10
15
20
25
30
35
% it
em n
ot c
over
ed
Number of items selected
0 200 400 600 8000
5
10
15
20
25
30
35
sequential
0 200 400 600 8000
5
10
15
20
25
30
35
10−batch
0 200 400 600 8000
5
10
15
20
25
30
35
100−batch
0 200 400 600 8000
5
10
15
20
25
30
35
Non−adaptive
Code Available
www.inf.ethz.ch/~chenyux/icml13/bmal-src.zip
Thanks for your attention
Come to see our poster!