View
219
Download
0
Tags:
Embed Size (px)
Citation preview
A Comparative Study of Methods for
Transductive Transfer Learning
Andrew Arnold, Ramesh Nallapati, William W. CohenMachine Learning Department
Carnegie Mellon University
IEEE ICDMWorkshop on Mining and Management of Biological Data
October 28, 2007
What we are able to do:
The neuronal cyclin-dependent kinase p35/cdk5 comprises a catalytic subunit (cdk5) and an activator subunit (p35)
Reversible histone acetylation changes the chromatin structure and can modulate gene transcription. Mammalian histone deacetylase 1 (HDAC1)
Training data: Test:
Train:Test:
• Supervised learning – Train on large, labeled data sets drawn from same
distribution as testing data– Well studied problem
What we’re getting better at doing:• Semi-supervised learning
– Same as before, but now• Add large unlabelled or weakly labeled data sets from same domain
– [Zhu ’05, Grandvalet ’05]
The neuronal cyclin-dependent kinase p35/cdk5 comprises a catalytic subunit (cdk5) and an activator subunit (p35)
Reversible histone acetylation changes the chromatin structure and can modulate gene transcription. Mammalian histone deacetylase 1 (HDAC1)
Train: Auxiliary (available for training):
Train:Auxiliary:
• Transductive learning – Unlabeled test data is available during training– Easier than inductive learning:
• Learning specific predictions rahter than general function• [Joachims ’99, ’03, Sindhwani ’05, Vapnik ‘98]
The neuronal cyclin-dependent kinase p35/cdk5 comprises a catalytic subunit (cdk5) and an activator subunit (p35)
Reversible histone acetylation changes the chromatin structure and can modulate gene transcription. Mammalian histone deacetylase 1 (HDAC1)
Train: Both Auxiliary & Eventual Test:
Train:Auxiliary & Test:
What we’re getting better at doing:
• Transfer learning (domain adaptation):– Leverage large, previously labeled data from a related domain
• Related domain we’ll be training on (with lots of data): Source• Domain we’re interested in and will be tested on (data scarce): Target
– [Ng ’06, Daumé ’06, Jiang ’06, Blitzer ’06, Ben-David ’07, Thrun ’96]
The neuronal cyclin-dependent kinase p35/cdk5 comprises a catalytic subunit (cdk5) and an activator subunit (p35)
Neuronal cyclin-dependent kinase p35/cdk5 (Fig 1, a) comprises a catalytic subunit (cdk5, left panel) and an activator subunit (p35, fmi #4)
Train (source domain: E-mail): Test (target domain: IM):
Train (source domain: Abstract):Test (target domain: Caption):
What we’d like to be able to do:
The neuronal cyclin-dependent kinase p35/cdk5 comprises a catalytic subunit (cdk5) and an activator subunit (p35)
Reversible histone acetylation changes the chromatin structure and can modulate gene transcription. Mammalian histone deacetylase 1 (HDAC1)
What we’d like to be able to do: • Transfer learning (multi-task):
• Same domain, but slightly different task• Related task we’ll be training on (with lots of data): Source• Task we’re interested in and will be tested on (data scarce): Target
– [Ando ’05, Sutton ’05]
Train (source task: Names): Test (target task: Pronouns):
Train (source task: Proteins):Test (target task: Action Verbs):
7
Motivation• Why is transfer important?
– Often we violate non-transfer assumption without realizing. How much data is truly identically distributed (i.i.d.)?
• E.g. Different authors, annotators, time periods, sources
– Large amounts of labeled data/trained classifiers already exist• Why waste data & computation?• Can learning be made easier by leveraging related domains/problems?
– Life-long learning
• Why is transduction important?– Why solve a harder problem than we need to?– Unlabeled data is vast and cheap
• Are transduction and transfer so different?– Can we learn more about one by studying the other?
8
Outline• Motivating Problems
– Supervised learning– Semi-supervised learning– Transductive learning– Transfer learning: domain adaptation– Transfer learning: multi-task
• Methods– Maximum entropy (MaxEnt)– Source regularized maximum entropy– Feature space expansion– Feature selection– Feature space transformation– Iterative Pseudo Labeling (IPL)– Biased thresholding– Support Vector Machines (SVMs)
• Inductive SVM• Transductive SVM
• Experiment:– Domain & Data
• Results• Conclusions & Contributions• Limitations & future work
9
Maximum Entropy (MaxEnt)
• Discriminative model– Matches feature expectations of model to data
Conditional likelihood:
Regularized optimization:
10
Summary of Learning Settings
11
Source-regularized MaxEnt
• Instead of regularizing towards zero– Learn model Λ’s on source data– During target training
• Regularize towards source-trained Λ’s
[Chelba’04]
12
Feature Space Expansion
• Add extra degrees of freedom– Allow classifier to discern general/specific features
[Daumé ’06, ’07]
13
Feature selection
• Emphasize features shared by source and target data
• Minimize different features
• How to measure?– Fisher exact test:
Is P(feature | source) == P(feature | target) ?
– If so, shared feature keep– If not, different feature discard
14
Feature Space Transformation
• Source and target originally independently separable• Learn transformation, G, to allow joint separation:
15
Iterative Pseudo Labeling (IPL)
• Novel algorithm for MaxEnt based transfer• Adjust feature values to match feature
expectation in source and target• θ trades off certainty vs adaptativity
16
IPL analysis
Given linear transform:
We can express conditional feature expectations of target data in terms of a transformation of source:
17
Biased Thresholding• Different proportions of positive examples
– Learning to predict rain in in humid and arid climates
• How to maximize F1 (and not accuracy)?– Score Cut (s-cut)
• Select score threshold over ranked train scores
• Apply to test data
– Percentage Cut (p-cut)
• Estimate proportion of positive examples expected in target data
• Set threshold so as to select this amount
18
Support Vector Machines (SVMs)
• Inductive (standard) SVM:– Learn separating hyperplane on labeled training data.
Then evaluate on held-out testing data.
• Transductive SVM:– Learn hyperplane in the presence of labeled training
data AND unlabeled testing data. Use distribution of testing points to assist you.
– Easier to learn particular labels than a whole function.– More expensive than inductive
19
Transductive vs. Inductive SVM
[Joachims ’99, ‘03]
20
Domain
21
DataUT
% positive 7%
positive 14,360.00 negative 202,435.00
total 216,795.00 abstracts 747
yapex% positive 15%positive 9,058.00 negative 51,472.00
total 60,530.00 abstracts 200
• Notice difference in:– Length and density of protein names– Number of training examples: ||UT|| ~ 4*||Yapex||– % positive examples: twice as many in Yapex
<prot> p38 stress-activated protein kinase </prot> inhibitor reverses <prot> bradykinin B(1) receptor </prot>-mediated component of inflammatory hyperalgesia.
<Protname>p35</Protname>/<Protname>cdk5 </Protname> binds and phosphorylates <Protname>beta-catenin</Protname> and regulates <Protname>beta-catenin </Protname> / <Protname>presenilin-1</Protname> interaction.
22
Experiment
• Examining three dimensions:– Labeled vs unlabeled vs prior auxiliary data
• eg. % target positive examples, few labeled target data
– Transduction vs induction– Transfer vs non-transfer
• Since few true positives, focused on: F1 := (2 * Precision * Recall) / (Precision + Recall)
• Source = UT, target = Yapex
• For IPL, θ = .95 (conservative)
23
Results: Transfer
• Transfer is much more difficult– Accuracy is not the problem
24
Results: Transduction
• Transduction helps in transfer setting– TSVM copes better than MaxEnt, ISVM
25
Results: IPL
• IPL can help boost performance– Makes transfer MaxEnt competitive with TSVM
– But bounded by quality of initial pseudo-labels
Results: Priors
• Priors improve unsupervised transfer– Threshold helps balance recall and precision better F1
– A little bit of knowledge can help a lot
27
Results: Supervision
• Supervised transfer beats supervised non-transfer– Significant at 99% binomial CI on precision and recall
• But not by as much as might be hoped for
• Even relatively simple transfer methods can help
28
Conclusions & Contributions• Introduced novel MaxEnt transfer method: IPL
– Can match transduction in unsupervised setting
– Gives probabilistic results
• Analyzed and compared various methods related to transfer learning and concluded:– Transfer is hard
• But made easier when explicitly addressed
– Transduction is a good start• TSVM excels even with scant prior knowledge
– A little prior target knowledge is even better• No need for fully labeled target data set
• Threshold is important:– Currently only using at test time– Why not incorporate earlier, get better pseudo labels
• Priors seem to help a lot:– Currently only using feature means, what about
variances?
• Can structuring feature
space lead to parsimonious
transferable priors?
Limitations & Future Work
token
rightleft
token.is.capitalized token.is.numeric
30
Limitations & Future Work: high-level• How to better make use of source data?
– Why doesn’t source data help so much?
• Is IPL convex?– Is this exactly what we want to optimize?– How does regularization affect convexity?
• What, exactly, is the relationship between transduction and transfer?– Can their theories be unified?
• When is it worth explicitly modeling transfer?– How different do the domains need to be?– How much source/target data do we need?– What kind of priors do we need?
31
☺ Thank you! ☺
¿ Questions ?
32
References