Structured Prediction and Active Learning for Information Retrieval Presented at Microsoft Research...

Structured Prediction and Active Learning for Information Retrieval

Presented at Microsoft Research AsiaAugust 21st, 2008

Yisong YueCornell University

Joint work with:Thorsten Joachims (advisor), Filip Radlinski,

Thomas Finley, Robert Kleinberg, Josef Broder

Outline

• Structured Prediction– Complex Retrieval Goals– Structural SVMs (Supervised Learning)

• Active Learning– Learning From Real Users– Multi-armed Bandit Problems

Supervised Learning

• Find function from input space X to output space Y

such that the prediction error is low.

Microsoft announced today that they acquired Apple for the amount equal to the gross national product of Switzerland. Microsoft officials stated that they first wanted to buy Switzerland, but eventually were turned off by the mountains and the snowy winters…

x

y1

GATACAACCTATCCCCGTATATATATTCTATGGGTATAGTATTAAATCAATACAACCTATCCCCGTATATATATTCTATGGGTATAGTATTAAATCAATACAACCTATCCCCGTATATATATTCTATGGGTATAGTATTAAATCAGATACAACCTATCCCCGTATATATATTCTATGGGTATAGTATTAAATCACATTTA

x

y-1x

y7.3

Examples of Complex Output Spaces

• Natural Language Parsing– Given a sequence of words x, predict the parse tree y.– Dependencies from structural constraints, since y has to be a

tree.

The dog chased the catx

S

VPNP

Det NV

NP

Det N

y

• Part-of-Speech Tagging– Given a sequence of words x, predict sequence of tags y.

– Dependencies from tag-tag transitions in Markov model.

Similarly for other sequence labeling problems, e.g., RNA

Intron/Exon Tagging.

The rain wet the catx

Det NVDet Ny



• Multi-class Labeling

• Sequence Alignment

• Grammar Trees & POS Tagging

• Markov Random Fields• Clustering• Information Retrieval (Rankings)

– Average Precision & NDCG

– Listwise Approaches

– Diversity

– More Complex Goals

Information Retrieval

• Input: x (feature representation of a document/query pair)

• Conventional Approach– Real valued retrieval functions f(x) – Sort by f(xi) to obtain ranking

• Training Method– Human-labeled data (documents labeled by relevance)– Learn f(x) using relatively simple criterion– Computationally convenient– Works pretty well (but we can do better)

Conventional SVMs

• Input: x (high dimensional point)

• Target: y (either +1 or -1)

• Prediction: sign(wTx)

• Training:

subject to:

• The sum of slacks upper bounds the accuracy loss

N

ii

w N

Cw

1

2

, 2

1minarg

iiT xwi 1)(y : i

i

i

Pairwise Preferences SVM

ji

jiw N

Cw

,,

2

, 2

1minarg

ji

yyjixwxw

ji

jijijT

iT

, ,0

:, ,1

,

,

Such that:

Large Margin Ordinal Regression [Herbrich et al., 1999]

Can be reduced to time [Joachims, 2005] )log( nnO

Pairs can be reweighted to more closely model IR goals [Cao et al., 2006]

Mean Average Precision

• Consider rank position of each relevance doc– K1, K2, … KR

• Compute Precision@K for each K1, K2, … KR

• Average precision = average of P@K

• Ex: has AvgPrec of

• MAP is Average Precision across multiple queries/rankings

76.05

3

3

2

1

1

3

1

Optimization Challenges

• Rank-based measures are multivariate– Cannot decompose (additively) into document pairs– Need to exploit other structure

• Defined over rankings– Rankings do not vary smoothly– Discontinuous w.r.t model parameters– Need some kind of relaxation/approximation

[Y & Burges; 2007]

Optimization Approach

• Approximations / Smoothing– Directly define gradient

• LambdaRank [Burges et al., 2006]

– Gaussian smoothing • SoftRank GP [Guiver & Snelson, 2008]

• Upper bound relaxations– Exponential Loss w/ Boosting

• AdaRank [Xu et al., 2007]

– Hinge Loss w/ Structural SVMs • [Chapelle et al., 2007]• SVM-map [Yue et al., 2007]

Structured Prediction

• Let x be a structured input (candidate documents)• Let y be a structured output (ranking)

• Use a joint feature map to encode

the compatibility of predicting y for given x.– Captures all the structure of the prediction problem

• Consider linear models: after learning w, we can make predictions via

FR ),( xy

),(maxargˆ xyyy

Tw

Linear Discriminant for Ranking

• Let x = (x1,…xn) denote candidate documents (features)

• Let yjk = {+1, -1} encode pairwise rank orders

• Feature map is linear combination of documents.

• Prediction made by sorting on document scores wTxi

),(maxargˆ xyyy

Tw

relj relk

kjjk xxy: :!

)(),( xy

Linear Discriminant for Ranking

• Using pairwise preferences is common in IR

• So far, just reformulated using structured prediction notation.

• But we won’t decompose into independent pairs– Treat the entire ranking as a structured object– Allows for optimizing average precision

relj relk

kjT

jkT xxwyw

: :!

)(),( xy

Structural SVM

• Let x denote a structured input (candidate documents)• Let y denote a structured output (ranking)

• Standard objective function:

• Constraints are defined for each incorrect labeling y’ over the set of documents x.

i

iN

Cw 2

2

1

iiiTiiTi ww )'(),'(),( :' )()()()( yxyxyyy

[Y, Finley, Radlinski, Joachims; SIGIR 2007]

Structural SVM for MAP

• Minimize

subject to

where ( y jk = {-1, +1} )

and

• Sum of slacks is smooth upper bound on MAP loss.

relj relk

ik

ij

ii xxyjk

: :!

)()()()( )(),( xy

i

iN

Cw 2

2

1


)'(1)'( yy Avgprec

i

[Y, Finley, Radlinski, Joachims; SIGIR 2007]

Too Many Constraints!

• For Average Precision, the true labeling is a ranking where the relevant documents are all ranked in the front, e.g.,

• An incorrect labeling would be any other ranking, e.g.,

• This ranking has Average Precision of about 0.8 with (y’) ¼ 0.2

• Intractable number of rankings, thus an intractable number of constraints!

Structural SVM Training

Original SVM Problem• Intractable number of constraints

• Most are dominated by a small set of “important” constraints

Structural SVM Approach• Repeatedly finds the next most

violated constraint…

• …until set of constraints is a good approximation.

[Tsochantaridis et al., 2005]

Finding Most Violated Constraint

• A constraint is violated when

• Finding most violated constraint reduces to

• Highly related to inference/prediction:

0)'(),(),'( iTw yxyxy

)'(),'(maxargˆ'

yxyyy

Tw

),(maxarg xyy

Tw


Observations• MAP is invariant on the order of documents within a relevance

class– Swapping two relevant or non-relevant documents does not change MAP.

• Joint SVM score is optimized by sorting by document score, wTxj

• Reduces to finding an interleaving

between two sorted lists of documents

relj relk

kT

jT

jk xwxwy: :!'

)(')'(maxarg yy


►

• Start with perfect ranking• Consider swapping adjacent

relevant/non-relevant documents

relj relk

kT

jT

jk xwxwy: :!'

)(')'(maxarg yy


►


relevant/non-relevant documents• Find the best feasible ranking of

the non-relevant document

relj relk

kT

jT

jk xwxwy: :!'

)(')'(maxarg yy


►


relevant/non-relevant documents• Find the best feasible ranking of the

non-relevant document• Repeat for next non-relevant

document

relj relk

kT

jT

jk xwxwy: :!'

)(')'(maxarg yy


►




document• Never want to swap past previous

non-relevant document

relj relk

kT

jT

jk xwxwy: :!'

)(')'(maxarg yy


►




document• Never want to swap past previous

non-relevant document• Repeat until all non-relevant

documents have been considered

relj relk

kT

jT

jk xwxwy: :!'

)(')'(maxarg yy

Comparison with other SVM methods

0.1

0.12

0.14

0.16

0.18

0.2

0.22

0.24

0.26

0.28

0.3

TREC 9 Indri TREC 10 Indri TREC 9Submissions

TREC 10Submissions

TREC 9Submissions(without best)


Dataset

Mea

n A

vera

ge

Pre

cisi

on

SVM-MAP

SVM-ROC

SVM-ACC

SVM-ACC2

SVM-ACC3

SVM-ACC4

Structural SVM for MAP

• Treats rankings as structured objects

• Optimizes hinge-loss relaxation of MAP– Provably minimizes the empirical risk

– Performance improvement over conventional SVMs

• Relies on subroutine to find most violated constraint

– Computationally compatible with linear discriminant

)'(),'(maxargˆ'

yxyyy

Tw

Need for Diversity (in IR)

• Ambiguous Queries– Users with different information needs issuing the

same textual query– “Jaguar”– At least one relevant result for each information need

• Learning Queries– User interested in “a specific detail or entire breadth

of knowledge available” • [Swaminathan et al., 2008]

– Results with high information diversity

Result #18

Top of First Page

Bottom of First Page

Results From 11/27/2007

Query: “Jaguar”

Learning to Rank

• Current methods– Real valued retrieval functions f(q,d) – Sort by f(q,di) to obtain ranking

• Benefits:– Know how to perform learning– Can optimize for rank-based performance measures– Outperforms traditional IR models

• Drawbacks:– Cannot account for diversity– During prediction, considers each document independently

Example

•Choose K documents with maximal information coverage.•For K = 3, optimal set is {D1, D2, D10}

Diversity via Set Cover

• Documents cover information– Assume information is partitioned into discrete units.

• Documents overlap in the information covered.

• Selecting K documents with maximal coverage is a set cover problem– NP-complete in general

– Greedy has (1-1/e) approximation [Khuller et al., 1997]

Diversity via Subtopics

• Current datasets use manually determined subtopic labels– E.g., “Use of robots in the world today”

• Nanorobots• Space mission robots• Underwater robots

– Manual partitioning of the total information– Relatively reliable– Use as training data

Weighted Word Coverage

• Use words to represent units of information

• More distinct words = more information

• Weight word importance

• Does not depend on human labeling

• Goal: select K documents which collectively cover as many distinct (weighted) words as possible– Greedy selection yields (1-1/e) bound.– Need to find good weighting function (learning problem).

Example

D1 D2 D3 Best

Iter 1 12 11 10 D1

Iter 2

Marginal Benefit

V1 V2 V3 V4 V5

D1 X X X

D2 X X X

D3 X X X X

Word Benefit

V1 1

V2 2

V3 3

V4 4

V5 5

Document Word Counts

Example

D1 D2 D3 Best

Iter 1 12 11 10 D1

Iter 2 -- 2 3 D3

Marginal Benefit

V1 V2 V3 V4 V5

D1 X X X

D2 X X X

D3 X X X X

Word Benefit

V1 1

V2 2

V3 3

V4 4

V5 5

Document Word Counts

Related Work Comparison

• Essential Pages [Swaminathan et al., 2008]

– Uses fixed function of word benefit– Depends on word frequency in candidate set

• Our goals– Automatically learn a word benefit function

• Learn to predict set covers • Use training data• Minimize subtopic loss

– No prior ML approach • (to our knowledge)

Linear Discriminant

• x = (x1,x2,…,xn) - candidate documents

• y – subset of x • V(y) – union of words from documents in y.

• Discriminant Function:

• (v,x) – frequency features (e.g., ¸10%, ¸20%, etc).

• Benefit of covering word v is then wT(v,x)

)(

),(),(y

xxyVv

TT vww

),(maxargˆ xyyy

Tw

[Y, Joachims; ICML 2008]

Linear Discriminant

• Does NOT reward redundancy – Benefit of each word only counted once

• Greedy has (1-1/e)-approximation bound

• Linear (joint feature space) – Allows for SVM optimization

)(

),(),(y

xxyVv

TT vww


More Sophisticated Discriminant

• Documents “cover” words to different degrees– A document with 5 copies of “Microsoft” might cover it

better than another document with only 2 copies.

• Use multiple word sets, V1(y), V2(y), … , VL(y)

• Each Vi(y) contains only words satisfying certain importance criteria.


More Sophisticated Discriminant

)(

)( 1

),(

),(

),(1

y

y

x

x

xy

LVv L

Vv

v

v

•Separate i for each importance level i. •Joint feature map is vector composition of all i

),(maxargˆ xyyy

Tw

•Greedy has (1-1/e)-approximation bound. •Still uses linear feature space.


Weighted Subtopic Loss

• Example:– x1 covers t1

– x2 covers t1,t2,t3

– x3 covers t1,t3

• Motivation– Higher penalty for not covering popular subtopics– Mitigates effects of label noise in tail subtopics

# Docs Loss

t1 3 1/2

t2 1 1/6

t3 2 1/3

Structural SVM

• Input: x (candidate set of documents)• Target: y (subset of x of size K)

• Same objective function:

• Constraints for each incorrect labeling y’.

• Scoreof best y at least as large as incorrect y’ plus loss

• Finding most violated constraint also set cover problem

i

iN

Cw 2

2

1


TREC Experiments

• TREC 6-8 Interactive Track Queries• Documents labeled into subtopics.

• 17 queries used, – considered only relevant docs– decouples relevance problem from diversity problem

• 45 docs/query, 20 subtopics/query, 300 words/doc

TREC Experiments

• 12/4/1 train/valid/test split– Approx 500 documents in training set

• Permuted until all 17 queries were tested once

• Set K=5 (some queries have very few documents)

• SVM-div – uses term frequency thresholds to define importance levels

• SVM-div2 – in addition uses TFIDF thresholds

TREC Results

Method Loss

Random 0.469

Okapi 0.472

Unweighted Model 0.471

Essential Pages 0.434

SVM-div 0.349

SVM-div2 0.382

Methods W / T / L

SVM-div vs Ess. Pages

14 / 0 / 3 **

SVM-div2 vs Ess. Pages

13 / 0 / 4

SVM-div vs SVM-div2

9 / 6 / 2

Can expect further benefit from having more training data.

IR as Structured Prediction

• As a general approach:– Structured prediction encapsulates goals of IR– Recent work have demonstrated the benefit of

using structured prediction.

• Future directions:– Apply to more general retrieval paradigms– XML retrieval

XML Retrieval

• Can retrieve information at different scopes– Individual documents– Smaller components of documents– Larger clusters of documents

• Issues of objective function & diversity still apply– Complex performance measures– Inter-component dependencies

[Image src: http://www.dcs.bbk.ac.uk/~ptw/teaching/html/]

http://www.dcs.bbk.ac.uk/~ptw/teaching/html/

Blog Retrieval

• Special case of XML retrieval

• Query: “High Energy Physics”– Return a blog feed?– Return blog front page?– Return individual blog posts?

– Optimizing for MAP?– Diversity?

Active Learning

• Batch Learning– Learns a model using pre-collected training data– Assumes training data representative of unseen data– Most studied machine learning paradigm

• Very successful in wide range of applications

– Includes most work on structured prediction

• Active Learning:– Can be applied directly to live users

• Representative of real users

– Removes cost of human labeled training data• Time / Money / Reliability

Implicit Feedback

• Users provide feedback while searching– What results they click on– How they reformulate queries– The length of time from issuing query to

clicking on a result– Geographical– User-specific data

• Personal search history• Age / gender / profession / etc.

Presentation Bias in Click Results

[Granka et al., 2004]

Biased Implicit Feedback

• Users biased towards top of rankings

– Passive collection results in very biased training data– No feedback for relevant documents outside top 10– Most prior work focus on passive collection

• Our goals– Use active learning methods to gather unbiased

implicit feedback– Still present good results to users while learning

• Learn “on the fly”

Preferences Between Rankings

• Interleave two rankings into one ranking– Users click more on documents from better ranking.

[Radlinski et al., 2008]

Active Learning Approach

• Leverage ranking preference test– Compare relative quality of competing retrieval functions

– Not biased towards any specific rank position

• Avoid showing users poor results– Quickly determine bad rankings

– Algorithm must learn “online”

• Formulate as a multi-armed bandit problem

Dueling Bandits Problem

• Given bandits (retrieval functions) r1, …, rN

• Each time step compares two bandits

• Comparison is noisy– Some probability of saying worse bandit is better– Each comparison independent

• Choose pair (rt,rt’) to minimize regret:

• (% users who prefer best bandit over chosen ones)

[Broder, Kleinberg, Y; work in progress]

T

tttT rrPrrPR

1

1)'*()*(

Regret Minimization

• If regret is sublinear in T (e.g., log T)– Then average regret (over time) tends to 0

• Want average regret to approach 0 quickly– RT should be as small as possible

T

tttT rrPrrPR

1

1)'*()*(


Results

• Let ε be the differentiability of top two (r*,r**)– P(r* > r**) = ½ + ε

• Known lower bound:

• Interleaved Filter achieves regret

• Information theoretically optimal – (up to constant factors)

TN

RT log

TN

ORT log


Assumptions

• Strong Stochastic Transitivity– For three bandits ri > rj > rk :

• Stochastic Triangle Inequality– For three bandits ri > rj > rk :

• Satisfied by many standard generative models– E.g., Logistic/Bradley-Terry (K=2)

jkijik , max

jkijik K

Interleaved Filter

• Choose candidate bandit at random ►

Interleaved Filter

• Choose candidate bandit at random

• Make noisy comparisons (Bernoulli trial)

against all other bandits in turn– Maintain mean and confidence interval

►

Interleaved Filter



against all other bandits in turn…– Maintain mean and confidence interval

• …until another bandit is better – With confidence 1 – δ

►

Interleaved Filter





• Repeat process with new candidate– Remove all empirically worse bandits

►

Interleaved Filter





• Repeat process with new candidate– Remove all empirically worse bandits

• Continue until 1 candidate left ►

Regret Analysis

• Stops comparing at 1 – δ confidence– Concludes one bandit is better

• Appropriate choice of δ ** leads to

1 - 1/T probability of finding r*

• Regret is 0 whenever we choose r*– Only accumulate regret when finding r*

T

tttT rrPrrPR

1

1)'*()*(

** δ = N-2T-1

Naïve Approach

• In deterministic case, O(N) comparisons to find max

• Extend to noisy case:– Maintain current candidate

– Run comparisons against 1 other bandit until 1 – δ confidence

– Take better bandit as candidate

– Repeat until all bandits considered

• Problem:– If current candidate awful

– Many comparisons to determine which awful bandit is better

– Incur high regret for each comparison

TN

ORT log2

Naïve vs Interleaved Filter

• Naïve performs poorly due to matches

between two awful bandits – Too many comparisons– Accumulates high regret

• Interleaved Filter bounds matches using

bounds on current candidate vs best– Stops when better bandit found– Regret bounded

T

tttT rrPrrPR

1

1)'*()*(

Naïve vs Interleaved Filter

• But Naïve concentrates on 2 bandits at

any point in time

• Interleaved Filter compares 1 bandit vs

rest simultaneously

– Should experience N2 blowup in regret– … or at least N log N

TN

ORT log2

TN

ORT log

Regret Analysis

• Define a round to be all the time steps for

a particular candidate bandit– O(log N) rounds total w.h.p.

• Define a match to be all the comparisons

between two bandits in a round– O(N) matches in each round– At most O(N log N) total matches

• End of each round – Remove empirically inferior bandits– “Constant fraction” of bandits removed

after each round

Regret Analysis

• O(log N) rounds played

• “Constant fraction” of bandits removed

at end of each round– O(N) total matches w.h.p.

• Each match incurs regret

• Expected regret:

TO log1

TN

ORE

TOT

TN

OT

RE

T

T

log

)(1

log1

1

Dueling Bandits Problem

• Uses a natural (and simple) regret formulation

– Captures preference for the best possible retrieval function– Consistent with unbiased ranking preference feedback

• [Radlinski et al., 2008]

– Online/Bandit formulation of finding the max w/ noisy compares

• Interleaved Filter achieves best possible regret bound

- Logarithmic in T - Linear in N

TN

ORT log

T

tttT rrPrrPR

1

1)'*()*(

Related Work

• Other forms of implicit feedback– Preferences between documents within a ranking

• Other active learning techniques– Bandit algorithm for minimizing abandonment

• [Radlinski et al., 2008]

– Active exploration of pairwise document preferences• [Radlinski et al., 2007]

– These approaches cannot generalize across queries

• Most learning approaches use passive collection– Susceptible to presentation bias

Moving Forward

• Limitations– Assumes users preferences are static

• Interleaved filter first explores, then commits

– Assumes a finite set of ranking functions• Should assume a continuous parameter space

• Future directions– Use Interleaved Filter as an optimization engine

• Collect finite sample from continuous parameter space

– Look at completely new problem formulations– Progress towards live user studies

Summary

• Structured prediction for complex retrieval problems– Rank-based performance measures– Diversity– Potentially much more!

• Active learning using unbiased implicit feedback– Learn directly from users (cheaper & more accurate)– Active learning for structured prediction models?

http://www.yisongyue.com

• Thanks to Tie-Yan Liu and Hang Li

http://www.yisongyue.com/

Extra Slides

SVM-map

Experiments

• Used TREC 9 & 10 Web Track corpus.

• Features of document/query pairs computed from outputs of existing retrieval functions.

(Indri Retrieval Functions & TREC Submissions)

• Goal is to learn a recombination of outputs which improves mean average precision.

Comparison with Best Base methods

0.1

0.12

0.14

0.16

0.18

0.2

0.22

0.24

0.26

0.28

0.3


TREC 10Submissions



Dataset

Mea

n A

vera

ge

Pre

cisi

on

SVM-MAP

Base 1

Base 2

Base 3

Comparison with other SVM methods

0.1

0.12

0.14

0.16

0.18

0.2

0.22

0.24

0.26

0.28

0.3


TREC 10Submissions



Dataset

Mea

n A

vera

ge

Pre

cisi

on

SVM-MAP

SVM-ROC

SVM-ACC

SVM-ACC2

SVM-ACC3

SVM-ACC4

SVM-div

Essential Pages

),(1log),(),(

),,(),(),,(

2 xxx

xxx

vrvrv

xvTFvxvC ii

[Swaminathan et al., 2008]

Benefit of covering word v with document xi

Importance of covering word v

Intuition: - Frequent words cannot encode information diversity.- Infrequent words do not provide significant information

x = (x1,x2,…,xn) - set of candidate documents for a queryy – a subset of x of size K (our prediction).

v

vC ),,(maxargˆ xyyy


• Encode each subtopic as an additional “word” to be covered.

• Use greedy prediction to find approximate most violated constraint.

)(

)(

)( 1

),(

),(

),('

1

y

y

y

x

x

xy

T T

Vv L

Vv

Lv

v

),'('maxargˆ'

xyyy

Tw

Approximate Constraint Generation

• Theoretical guarantees no longer hold.– Might not find an epsilon-close approximation to the feasible

region boundary.

• Performs well in practice.

Approximate constraint generation seems to work perform well.

Synthetic Dataset

• Trec dataset very small• Synthetic dataset so we can vary retrieval size K

• 100 queries• 100 docs/query, 25 subtopics/query, 300 words/doc

• 15/10/75 train/valid/test split

Consistently outperforms Essential Pages

Interleaved Filter

Lower Bound Example

• All suboptimal bandits roughly equivalent– ε12 = ε13 = … = ε1N

• comparisons to differentiate best from suboptimal

• Pay Θ(ε) regret for each comparison

• Accumulated regret over all comparisons is at least

Tlog

12

TN

log

Per-Match Regret

• Number of comparisons in match ri vs rj :

– ε1i > εij : round ends before concluding ri > rj

– ε1i < εij : conclude ri > rj before round ends, remove rj

• Pay ε1i + ε1j regret for each comparison

– By triangle inequality ε1i + ε1j ≤ (2K+1)max{ε1i , εij}

– Thus by stochastic transitivity accumulated regret is

TO

iji

log, max

122

1

TOTO

iji

log1

log, max

1

1

Number of Rounds

• Assume all superior bandits have

equal prob of defeating candidate– Worst case scenario under transitivity

• Model this as a random walk– rj transitions to each ri (i < j) with equal probability

– Compute total number of steps before reaching r1 (i.e., r*)

• Can show O(log N) w.h.p. using Chernoff bound

Total Matches Played

• O(N) matches played in each round– Naïve analysis yields O(N log N) total

• However, all empirically worse bandits

are also removed at the end of each round– Will not participate in future rounds

• Assume worst case that inferior bandits have ½ chance of

being empirically worse– Can show w.h.p. that O(N) total matches are played

over O(log N) rounds

Removing Inferior Bandits

• At conclusion of each round– Remove any empirically worse bandits

• Intuition:– High confidence that winner is better

than incumbent candidate

– Empirically worse bandits cannot be “much better” than

incumbent candidate

– Can show that winner is also better than empirically worse

bandits with high confidence

– Preserves 1-1/T confidence overall that we’ll find the best bandit

Structured Prediction and Active Learning for Information Retrieval Presented at Microsoft Research...

Documents

Transcript of Structured Prediction and Active Learning for Information Retrieval Presented at Microsoft Research...