Cost-sensitive Decision Tree and Test Strategies

Download Report

Transcript Cost-sensitive Decision Tree and Test Strategies

Get Another Label?
Improving Data Quality and Data
Mining Using Multiple, Noisy
Labelers
Victor Sheng, Foster Provost, Panos Ipeirotis
KDD 2008
New York University
Stern School
Outsourcing preprocessing
• Traditionally, data mining teams have invested substantial
internal resources in data formulation, information extraction,
cleaning, and other preprocessing
– Raghu from his Innovation Lecture
“the best you can expect are noisy labels”
• Now, we can outsource preprocessing tasks, such as labeling,
feature extraction, verifying information extraction, etc.
– using Mechanical Turk, Rent-a-Coder, etc.
– quality may be lower than expert labeling (much?)
– but low costs can allow massive scale
• The ideas may apply also to focusing user-generated tagging,
crowdsourcing, etc.
2
ESP Game (by Luis von Ahn)
3
Other “free” labeling schemes
• Open Mind initiative (http://www.openmind.org)
• Other GWAP games
– Tag a Tune
– Verbosity (tag words)
– Matchin (image ranking)
• Web 2.0 systems?
– Can/should tagging be directed?
4
Noisy labels can be problematic
Many tasks rely on high-quality labels for objects:
–
–
–
–
–
learning predictive models
searching for relevant information
finding duplicate database records
image recognition/labeling
song categorization
• Noisy labels can lead to degraded performance
5
Quality and Classification Performance
• Labeling quality (labeling accuracy P) increases
classification quality increases
P = 1.0
100
P = 0.8
80
P = 0.6
70
60
P = 0.5
50
300
280
260
240
220
200
180
160
140
120
100
80
60
40
20
40
1
Accuracy
90
Number of examples (Mushroom)
6
Majority Voting and Label Quality
• Ask multiple “noisy” labelers, keep majority label as “true”
label
• Given 2N+1 labelers with uniform accuracy P, integrated
quality is P(Bin(2N+1, P) <= N)
P is probability
of individual labeler
being correct
Integrated quality
1
0.9
P=1.0
0.8
P=0.9
0.7
P=0.8
0.6
P=0.7
0.5
P=0.6
0.4
P=0.5
0.3
P=0.4
0.2
1
3
5
7
9
11
13
Number of labelers
 (1) removing noisy labelers, (2) collect more labels as much as possible
Labeling Methods
• MV: majority voting
• Uncertainty preserving labeling (soft label)
– Multiplied Examples (ME): for each example xi, ME considers the
multiset of existing labels (Li,j), and for each Lij, it creates a replica with
weight 1/|Li,j| (??)
• These replicas with weights are fed into the classifier for training
• Another method is to use Naive Bayes (see WSDM’11):
Modeling Annotator Accuracies for Supervised Learning, WSDM 2011
Get Another Label?
• Single Labeling (SL)
– One label per each sample; get more samples
• Repeated labeling (w/ a fixed set of samples)
– Round-robin Repeated Labeling
• Fixed Round Robin (FRR)
– Keep labeling the same set of samples
• Generalized Round Robin (GRR)
– Keep labeling the same set of samples, yet giving highest preference to a
sample with the fewest labels
– Selective Repeated-Labeling
• Consider label uncertainty of a sample (LU)
• Considerer classification (model) uncertainty of a sample (MU)
• Consider both label uncertainty and model uncertainty (LMU)
Experiment Setting
• 70/30 division (70% for training, 30% for testing)
• Uniform accuracy of labeling as P; for each sample, a correct
label is given with probability P
• Classifier: C4.5 in WEKA
Single Labels vs. Majority Voting
• Sample size of training data set matters
• When sample size is large enough, MV is better than SL
• With low noise, more (single labeled) examples better
MV-FRR
(50 examples)
Tradeoffs for Modeling
• Get more labels  Improve label quality  Improve classification
• Get more examples  Improve classification
P = 1.0
100
P = 0.8
80
P = 0.6
70
60
P = 0.5
50
80
10
0
12
0
14
0
16
0
18
0
20
0
22
0
24
0
26
0
28
0
30
0
60
40
20
40
1
Accuracy
90
Number of examples (Mushroom)
Selective Repeated-Labeling
• We have seen:
– With enough examples and noisy labels, getting multiple labels is
better than single-labeling
– When we consider costly preprocessing, the benefit is magnified
• Can we do better than the basic strategies?
• Key observation: we have additional information to guide
selection of data for repeated labeling
– Multi-set labels; e.g., {+,-,+,+,-,+} vs. {+,+,+,+}
13
Natural Candidate: Entropy
• Entropy is a natural measure of label uncertainty:
– E({+,+,+,+,+,+})=0
– E({+,-, +,-, +,- })=1
| S |
| S | | S |
| S |
E(S )  
log2

log2
|S|
|S| |S|
|S|
| S  |: positive | S  |: negative
Strategy: Get more labels for examples with high-entropy label
multisets
14
What Not to Do: Use Entropy
0.95
Improves at first,
hurts in long run
Labeling quality
0.9
0.85
0.8
0.75
0.7
ENT ROPY
GRR
0.65
0.6
0
400
800
1200
1600
Number of labels (waveform, p=0.6)
2000
Why not Entropy
• In the presence of noise, entropy will be high
even with many labels
• Entropy is scale invariant
(3+ , 2-) has same entropy as (600+ , 400-)
16
Binomial Dist. with Uniform Prior Dist.
Let Y ~ Bin(, n) where  ~ Uniform( 0, 1 ).
p( | Y )  fUnif (0,1) f Bin (Y |  )
n Y
 1 *   (1   )n Y
Y 
  Y (1   )n Y
You cannot just call the posterior a
binomial distribution because you are
conditioning on Y and  is a random
variable, not the other way around.
T hepdf for thebeta distribution which is known to be properis :
Γ(α  β) α 1
Beta(x|α,β) 
x (1  x ) β 1 ( 0  x  1 and α,β  0).
Γ(α)Γ(β)
Let x   , α  Y  1,β  n  Y  1
T hus,p(π | Y, n) ~ Beta(Y 1, n - Y  1) 
p(π | Y, n) 
Note: (k )  (k  1)!
[Gamma Function]
Γ(n  2 )
x (Y 1)1 (1  x )( n Y 1)1
Γ(Y  1 )Γ (n  Y  1 )
Γ(n  2 )
xY (1  x ) n Y
Γ(Y  1 )Γ ( n  Y  1 )
This is the normalization constant to
transform y(1-)n-y into a beta distribution.
Estimating Label Uncertainty (LU)
• Observe +’s and –’s and compute Pr{+|obs} and Pr{-|obs}
• Label uncertainty = tail of beta distribution
– P=0.5, alpha1 and alpha2
– For more accurate estimation, we can instead use 95% HDR,
Highest Density Region (or interval)
Beta(18,8)
95%
HDR
1
SLU
2
posterior
3
4
Beta probability
density function
0
[.51,.84]
0.0
0.5
0.0
1.0
0.2
0.4
0.6
p
0.8
1.0
Label Uncertainty
• p=0.7
• 5 labels
(3+, 2-)
• Entropy ~ 0.97
• CDFb=0.34
Label Uncertainty
• p=0.7
• 10 labels
(7+, 3-)
• Entropy ~ 0.88
• CDFb=0.11
Label Uncertainty
• p=0.7
• 20 labels
(14+, 6-)
• Entropy ~ 0.88
• CDFb=0.04
Labeling quality
Label Uncertainty vs. Round Robin
1
0.9
0.8
GRR
LU
0.7
0.6
0
400
800
1200
1600
2000
Number of labels (waveform, p=0.6)
similar results across a dozen data sets
Another strategy:
Model Uncertainty (MU)
• Learning a model of the data provides an
alternative source of information about label
certainty
• Model uncertainty: get more labels for
instances that cannot be modeled well
• Intuition?
– for data quality, low-certainty “regions” may be
due to incorrect labeling of corresponding
instances
– for modeling: why improve training data quality if
model already is certain there? (LMU)
?
- -- +
+
+
++ - - - - - + +
-+ ++
+ + ++ - - - -- - + +
- - ----++
-+
-+-
?
23
Yet another strategy:
Label & Model Uncertainty (LMU)
• Label and model uncertainty (LMU): avoid
examples where either strategy is certain
S LMU 
S LU  S MU
24
Comparison
Model Uncertainty
Label & Model
alone also improves
Label
Uncertainty
quality
Uncertainty
1
Labeling quality
0.95
0.9
0.85
0.8
GRR
0.75
GRR
MU
LU
LMU
0.7
0.65
0.6
0
400
800
1200
Number of labels (waveform, p=0.6)
1600
2000
Comparison: Model Quality
• Across 12 domains, LMU is always better than GRR.
• LMU is statistically significantly better than LU and MU
90
Label & Model uncertainty
Accuracy
85
80
75
GRR
MU
LU
LMU
70
65
60
0
400
800
1200
Number of labels (spambase, p=0.6)
1600
2000
Summary
Micro-task outsourcing (e.g., MTurk, RentaCoder
ESP game) has changed the landscape for data
formulation
• Repeated labeling can improve data quality and
model quality (but not always)
• When labels are noisy, repeated labeling can be
preferable to single labeling even when labels
aren’t particularly cheap
• When labels are relatively cheap, repeated
labeling can do much better
• Round-robin repeated labeling can do well
• Selective repeated labeling improves substantially
27