lecture14-learning-ranking.ppt

Introduction to Information Retrieval
Introduction to
Information Retrieval
CS276: Information Retrieval and Web Search
Christopher Manning, Pandu Nayak, and
Prabhakar Raghavan
Lecture 14: Learning to Rank

Machine learning for IR ranking?
 We’ve looked at methods for ranking documents in IR
 Cosine similarity, inverse document frequency, proximity,
pivoted document length normalization, Pagerank, …
 We’ve looked at methods for classifying documents
using supervised machine learning classifiers
 Naïve Bayes, Rocchio, kNN, SVMs
 Surely we can also use machine learning to rank the
documents displayed in search results?
 Sounds like a good idea
 A.k.a. “machine-learned relevance” or “learning to rank”
Sec. 15.4

Machine learning for IR ranking
 This “good idea” has been actively researched – and
actively deployed by major web search engines – in
the last 7 or so years
 Why didn’t it happen earlier?
 Modern supervised ML has been around for about 20
years…
 Naïve Bayes has been around for about 50 years…

Machine learning for IR ranking
 There’s some truth to the fact that the IR community
wasn’t very connected to the ML community
 But there were a whole bunch of precursors:
 Wong, S.K. et al. 1988. Linear structure in information
retrieval. SIGIR 1988.
 Fuhr, N. 1992. Probabilistic methods in information
retrieval. Computer Journal.
 Gey, F. C. 1994. Inferring probability of relevance using the
method of logistic regression. SIGIR 1994.
 Herbrich, R. et al. 2000. Large Margin Rank Boundaries for
Ordinal Regression. Advances in Large Margin Classifiers.

Why weren’t early attempts very
successful/influential?
 Sometimes an idea just takes time to be appreciated…
 Limited training data
 Especially for real world use (as opposed to writing
academic papers), it was very hard to gather test collection
queries and relevance judgments that are representative of
real user needs and judgments on documents returned
 This has changed, both in academia and industry
 Poor machine learning techniques
 Insufficient customization to IR problem
 Not enough features for ML to show value

Why wasn’t ML much needed?
 Traditional ranking functions in IR used a very small
number of features, e.g.,
 Term frequency
 Inverse document frequency
 Document length
 It was easy to tune weighting coefficients by hand
 And people did
 You guys did in PA3
 Some of you even grid searched a bit

Why is ML needed now?
 Modern systems – especially on the Web – use a great
number of features:
 Arbitrary useful features – not a single unified model
 Log frequency of query word in anchor text?
 Query word in color on page?
 # of images on page?
 # of (out) links on page?
 PageRank of page?
 URL length?
 URL contains “~”?
 Page edit recency?
 Page length?
 The New York Times (2008-06-03) quoted Amit Singhal as
saying Google was using over 200 such features.

Simple example:
Using classification for ad hoc IR
 Collect a training corpus of (q, d, r) triples
 Relevance r is here binary (but may be multiclass, with 3–7 values)
 Document is represented by a feature vector
 x = (α, ω) α is cosine similarity, ω is minimum query window size
 ω is the the shortest text span that includes all query words
 Query term proximity is a very important new weighting factor
 Train a machine learning model to predict the class r of a document-
query pair
Sec. 15.4.1

Simple example:
 A linear score function is then
Score(d, q) = Score(α, ω) = aα + bω + c
 And the linear classifier is
Decide relevant if Score(d, q) > θ
 … just like when we were doing text classification
Sec. 15.4.1

Simple example:
0
2 3 4 5
0.05
0.025
cosine
score

Term proximity 
R
R
R
R
R R
R
R
R
R
R
N
N
N
N
N
N
N
N
N
N
Sec. 15.4.1
Decision surface

More complex example of using classification for
search ranking [Nallapati 2004]
 We can generalize this to classifier functions over
more features
 We can use methods we have seen previously for
learning the linear classifier weights

An SVM classifier for information retrieval
[Nallapati 2004]
 Let g(r|d,q) = wf(d,q) + b
 SVM training: want g(r|d,q) ≤ −1 for nonrelevant
documents and g(r|d,q) ≥ 1 for relevant documents
 SVM testing: decide relevant iff g(r|d,q) ≥ 0
 Features are not word presence features (how would you
deal with query words not in your training data?) but
scores like the summed (log) tf of all query terms
 Unbalanced data (which can result in trivial always-say-
nonrelevant classifiers) is dealt with by undersampling
nonrelevant documents during training (just take some at
random) [there are other ways of doing this – cf. Cao et al. later]

[Nallapati 2004]
 Experiments:
 4 TREC data sets
 Comparisons with Lemur, a state-of-the-art open source IR
engine (Language Model (LM)-based – see IIR ch. 12)
 Linear kernel normally best or almost as good as quadratic
kernel, and so used in reported results
 6 features, all variants of tf, idf, and tf.idf scores

[Nallapati 2004]
Train Test Disk 3 Disk 4-5 WT10G (web)
Disk 3 LM 0.1785 0.2503 0.2666
SVM 0.1728 0.2432 0.2750
Disk 4-5 LM 0.1773 0.2516 0.2656
SVM 0.1646 0.2355 0.2675
 At best the results are about equal to LM
 Actually a little bit below
 Paper’s advertisement: Easy to add more features
 This is illustrated on a homepage finding task on
WT10G:
 Baseline LM 52% success@10, baseline SVM 58%
 SVM with URL-depth, and in-link features: 78% S@10

“Learning to rank”
 Classification probably isn’t the right way to think
about approaching ad hoc IR:
 Classification problems: Map to a unordered set of classes
 Regression problems: Map to a real value [Start of PA4]
 Ordinal regression problems: Map to an ordered set of
classes
 A fairly obscure sub-branch of statistics, but what we want here
 This formulation gives extra power:
 Relations between relevance levels are modeled
 Documents are good versus other documents for query
given collection; not an absolute scale of goodness
Sec. 15.4.2

“Learning to rank”
 Assume a number of categories C of relevance exist
 These are totally ordered: c1 < c2 < … < cJ
 This is the ordinal regression setup
 Assume training data is available consisting of document-
query pairs represented as feature vectors ψi and
relevance ranking ci
 We could do point-wise learning, where we try to map
items of a certain relevance rank to a subinterval (e.g,
Crammer et al. 2002 PRank)
 But most work does pair-wise learning, where the input
is a pair of results for a query, and the class is the
relevance ordering relationship between them

Point-wise learning
 Goal is to learn a threshold to separate each rank

Pairwise learning: The Ranking SVM
[Herbrich et al. 1999, 2000; Joachims et al. 2002]
 Aim is to classify instance pairs as correctly ranked or
incorrectly ranked
 This turns an ordinal regression problem back into a binary
classification problem
 We want a ranking function f such that
ci > ck iff f(ψi) > f(ψk)
 … or at least one that tries to do this with minimal
error
 Suppose that f is a linear function
f(ψi) = wψi
Sec. 15.4.2

The Ranking SVM
 Ranking Model: f(ψi)
f (yi)
Sec. 15.4.2

The Ranking SVM
 Then (combining the two equations on the last slide):
ci > ck iff w(ψi − ψk) > 0
 Let us then create a new instance space from such
pairs:
Φu = Φ(di, dj, q) = ψi − ψk
zu = +1, 0, −1 as ci >,=,< ck
 We can build model over just cases for which zu = −1
 From training data S = {Φu}, we train an SVM
Sec. 15.4.2

Two queries in the original space

Two queries in the pairwise space

The Ranking SVM
 The SVM learning task is then like other examples
that we saw before
 Find w and ξu ≥ 0 such that
 ½wTw + C Σ ξu is minimized, and
 for all Φu such that zu < 0, wΦu ≥ 1 − ξu
 We can just do the negative zu, as ordering is
antisymmetric
 You can again use SVMlight (or other good SVM
libraries) to train your model (SVMrank specialization)
Sec. 15.4.2

Aside: The SVM loss function
 The minimization
minw ½wTw + C Σ ξu
and for all Φu such that zu < 0, wΦu ≥ 1 − ξu
 can be rewritten as
minw (1/2C)wTw + Σ ξu
and for all Φu such that zu < 0, ξu ≥ 1 − (wΦu)
 Now, taking λ = 1/2C, we can reformulate this as
minw Σ [1 − (wΦu)]+ + λwTw
 Where []+ is the positive part (0 if a term is negative)

Aside: The SVM loss function
 The reformulation
minw Σ [1 − (wΦu)]+ + λwTw
 shows that an SVM can be thought of as having an
empirical “hinge” loss combined with a weight
regularizer
Loss
1 wΦu
Hinge loss Regularizer of‖w‖

Adapting the Ranking SVM for (successful)
Information Retrieval
[Yunbo Cao, Jun Xu, Tie-Yan Liu, Hang Li, Yalou Huang, Hsiao-
Wuen Hon SIGIR 2006]
 A Ranking SVM model already works well
 Using things like vector space model scores as features
 As we shall see, it outperforms them in evaluations
 But it does not model important aspects of practical
IR well
 This paper addresses two customizations of the
Ranking SVM to fit an IR utility model

The ranking SVM fails to model the IR problem
well …
1. Correctly ordering the most relevant documents is
crucial to the success of an IR system, while
misordering less relevant results matters little
 The ranking SVM considers all ordering violations as the
same
2. Some queries have many (somewhat) relevant
documents, and other queries few. If we treat all
pairs of results for a query equally, queries with
many results will dominate the learning
 But actually queries with few relevant results are at least
as important to do well on

Based on the LETOR test collection
 From Microsoft Research Asia
 An openly available standard test collection with
pregenerated features, baselines, and research results for
learning to rank
 It’s availability has really driven research in this area
 OHSUMED, MEDLINE subcollection for IR
 350,000 articles
 106 queries
 16,140 query-document pairs
 3 class judgments: Definitely relevant (DR), Partially Relevant
(PR), Non-Relevant (NR)
 TREC GOV collection (predecessor of GOV2, cf. IIR p. 142)
 1 million web pages
 125 queries

Principal components projection of 2 queries
[solid = q12, open = q50; circle = DR, square = PR, triangle = NR]

Ranking scale importance discrepancy
[r3 = Definitely Relevant, r2 = Partially Relevant, r1 = Nonrelevant]

Number of training documents per query
discrepancy [solid = q12, open = q50]

Recap: Two Problems with Direct Application of
the Ranking SVM
 Cost sensitiveness: negative effects of making errors on top
ranked documents
d: definitely relevant, p: partially relevant, n: not relevant
ranking 1: p d p n n n n
ranking 2: d p n p n n n
 Query normalization: number of instance pairs varies according to
query
q1: d p p n n n n
q2: d d p p p n n n n n
q1 pairs: 2*(d, p) + 4*(d, n) + 8*(p, n) = 14
q2 pairs: 6*(d, p) + 10*(d, n) + 15*(p, n) = 31

These problems are solved with a new Loss
function
 τ weights for type of rank difference
 Estimated empirically from effect on NDCG
 μ weights for size of ranked result set
 Linearly scaled versus biggest result set
min
w
L(w) = tk(i)
mq(i)
1- zi
w,xi
(1)
- xi
(2)
é
ë
ù
û
i=1
l
å
+
+l w
2

Experiments
 OHSUMED (from LETOR)
 Features:
 6 that represent versions of tf, idf, and tf.idf factors
 BM25 score (IIR sec. 11.4.3)
 A scoring function derived from a probabilistic approach to IR,
which has traditionally done well in TREC evaluations, etc.

Experimental Results (OHSUMED)

MSN Search [now Bing]
 Second experiment with MSN search
 Collection of 2198 queries
 6 relevance levels rated:
 Definitive 8990
 Excellent 4403
 Good 3735
 Fair 20463
 Bad 36375
 Detrimental 310

Experimental Results (MSN search)

Alternative: Optimizing Rank-Based Measures
[Yue et al. SIGIR 2007]
 If we think that NDCG is a good approximation of the
user’s utility function from a result ranking
 Then, let’s directly optimize this measure
 As opposed to some proxy (weighted pairwise prefs)
 But, there are problems …
 Objective function no longer decomposes
 Pairwise prefs decomposed into each pair
 Objective function is flat or discontinuous

Discontinuity Example
 NDCG computed using rank positions
 Ranking via retrieval scores
 Slight changes to model parameters
 Slight changes to retrieval scores
 No change to ranking
 No change to NDCG
d1 d2 d3
Retrieval Score 0.9 0.6 0.3
Rank 1 2 3
Relevance 0 1 0
NDCG = 0.63
NDCG discontinuous w.r.t
model parameters!

Structural SVMs [Tsochantaridis et al., 2007]
 Structural SVMs are a generalization of SVMs where the output
classification space is not binary or one of a set of classes, but some
complex object (such as a sequence or a parse tree)
 Here, it is a complete (weak) ranking of documents for a query
 The Structural SVM attempts to predict the complete ranking for
the input query and document set
 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.,
 There are an intractable number of rankings, thus an intractable
number of constraints!

Structural SVM training
[Tsochantaridis et al., 2007]
Original SVM Problem
 Exponential constraints
 Most are dominated by a small
set of “important” constraints
Structural SVM Approach
 Repeatedly finds the next most violated
constraint…
 …until a set of constraints which is a
good approximation is found
Structural SVM training proceeds incrementally by starting with a working set of
constraints, and adding in the most violated constraint at each iteration

Other machine learning methods for learning to
rank
 Of course!
 I’ve only presented the use of SVMs for machine
learned relevance, but other machine learning
methods have also been used successfully
 Boosting: RankBoost
 Ordinal Regression loglinear models
 Neural Nets: RankNet
 (Gradient-boosted) Decisision Trees

The Limitations of Machine Learning
 Everything that we have looked at (and most work in
this area) produces linear models of features by
weighting different base features
 This contrasts with most of the clever ideas of
traditional IR, which are nonlinear scalings and
combinations of basic measurements
 log term frequency, idf, pivoted length normalization
 At present, ML is good at weighting features, but not
at coming up with nonlinear scalings
 Designing the basic features that give good signals for
ranking remains the domain of human creativity

Summary
 The idea of learning ranking functions has been around
for about 20 years
 But only recently have ML knowledge, availability of
training datasets, a rich space of features, and massive
computation come together to make this a hot research
area
 It’s too early to give a definitive statement on what
methods are best in this area … it’s still advancing rapidly
 But machine learned ranking over many features now
easily beats traditional hand-designed ranking functions
in comparative evaluations [in part by using the hand-designed functions as features!]
 There is every reason to think that the importance of
machine learning in IR will grow in the future.

Resources
 IIR secs 6.1.2–3 and 15.4
 LETOR benchmark datasets
 Website with data, links to papers, benchmarks, etc.
 http://research.microsoft.com/users/LETOR/
 Everything you need to start research in this area!
 Nallapati, R. Discriminative models for information
retrieval. SIGIR 2004.
 Cao, Y., Xu, J. Liu, T.-Y., Li, H., Huang, Y. and Hon, H.-W.
Adapting Ranking SVM to Document Retrieval, SIGIR
2006.
 Y. Yue, T. Finley, F. Radlinski, T. Joachims. A Support
Vector Method for Optimizing Average Precision. SIGIR
2007.

lecture14-learning-ranking.ppt

More Related Content

What's hot (20)

Similar to lecture14-learning-ranking.ppt (20)

Recently uploaded (20)

lecture14-learning-ranking.ppt