Learn to Rank search results

Recruiting SolutionsRecruiting SolutionsRecruiting Solutions1

Learning-to-Rank Search ResultsGanesh Venkataraman

http://www.linkedin.com/in/npcomplete@gvenkataraman

Audience Classification

Search background ML background Search + ML background

2

Search

ML

Search+ML

0% 10% 20% 30% 40% 50% 60% 70%

Outline

Search Overview Why Learning to Rank (LTR)? Biases with collecting training data from click logs

– Sampling bias– Presentation bias

Three basic approaches– Point wise– Pair wise– List wise

Key Takeaways/Summary

3

tl;dr

Ranking interacts heavily with retrieval and query understanding

Ground truth > features > model*

List wise > pair wise > point wise

4

* Airbnb engineering blog: http://nerds.airbnb.com/architecting-machine-learning-system-risk/

5

Primer on Search

Information need query select from results

rank using IR model

user:

system:

bird’s-eye view of how a search engine works

6

7

Pre Retrieval/Retrieval/Post Retrieval

Pre retrieval– Process input query, rewrite, check for spelling etc.– Hit search (potentially several) nodes with appropriate query

Retrieval– Given a query, retrieve all documents matching query along with

a score

Post retrieval– Merge sort results from different search nodes– Add relevant information to search results used by front end

8

9

Claim #1 Search is about understanding the query/user intent

10

Understanding intent

10

TITLE CO GEO

TITLE-237software engineersoftware developer

programmer…

CO-1441Google Inc.

Industry: Internet

GEO-7583Country: US

Lat: 42.3482 NLong: 75.1890 W

(RECOGNIZED TAGS: NAME, TITLE, COMPANY, SCHOOL, GEO, SKILL )

11

Fixing user errors

typos

Help users spell names

12

Claim #2 Search is about understanding systems

13

The Search Index

Inverted Index: Mapping from (search) terms to list of documents (they are present in)

Forward Index: Mapping from documents to metadata about them

14

Posting List

Term Posting List

DO = “it is what it is”

D1 = “what is it”

D2 = “it is a banana”

DocId

a

banana

is

it

what

2

2

0

0

0 Frequency

1

1

2

2

1

Bold

B

1 1 2 1

1 1 2 1

1 1

15

Candidate selection for “abraham lincoln”

Posting lists– “abraham” => {5, 7, 8, 23, 47, 101}– “lincoln” => {7, 23, 101, 151}

Query = “abraham AND lincoln”– Retrieved set => {7, 23, 101}

Some systems level issues– How to represent posting lists efficiently?– How does one traverse a very long posting list? (for words like

“the”, “an” etc.)?

16

Claim #3 Search is ranking problem

17

What is search ranking?

Ranking – Find a ordered list of documents according to relevance between

documents and query.

Traditional search– f(query, document) => score

Social networks context– f(query, document, user) => score– Find an ordered list of documents according to relevance

between documents, query and user

18

Why LTR?

Manual models become hard to tune with very large number of features and non-convex interactions

Leverages large volume of click through data in an automated way

Unique challenges involved in crowdsourcing personalized ranking

Key Issues– How do we collect training data?– How do we avoid biases?– How do we train the model?

TRAINING

Documents for training

Features

Human evaluation

Labels

Machine learning model

TRAINING

Documents for training

Features

Human evaluation

Labels

Machine learning model

21

training options – crowdsourcing judgment

Crowd source judgments• (query, user, document) -> label• {1, 2, 3, 4, 5}, higher label =>

better• Issues

• Personalized world• Difficult to scale

Mining click stream

Approach: Clicked = Relevant, Not-Clicked = Not Relevant

User eye scan direction

Unfairly penalized?

23

Position Bias

“Accurately interpreting clickthrough data as implicit feedback” – Joachims et. al, ACM SIGIR, 2005.– Experiment #1

Present users with normal Google search results 55.56% users clicked first result 5.56% clicked second result

– Experiment #2 Same result page as first experiment, but 1st and 2nd result were

flipped 57.14% users clicked first result 7.14% clicked second result

FAIR PAIRS

[Radlinski and Joachims, AAAI’06]

• Fair Pairs: • Randomize, Clicked= R,

Skipped= NR

FAIR PAIRS

Flipped

[Radlinski and Joachims, AAAI’06]

• Fair Pairs: • Randomize, Clicked= R,

Skipped= NR

FAIR PAIRS

Flipped

[Radlinski and Joachims, AAAI’06]

• Fair Pairs: • Randomize, Clicked= R,

Skipped= NR• Great at dealing with position bias• Does not invert models

27

Issue #2 – Sampling Bias

Sample bias– User clicks or skips only what is shown.– What about low scoring results from existing model?– Add low-scoring results as ‘easy negatives’ so model

learns bad results not presented to user.

…

label 0

label 0

label 0

label 0

…

page 1 page 2 page 3 page n

28

Issue #2 – Sampling Bias

29

Avoiding Sampling Bias – Easy negatives

Invasive way– For a small sample of users add bad results in the SERP page to

test that the results were indeed bad– Not really recommended since it affects UX

Non-Invasive way– Assume we have a decent model– Take tail results and add them to model as an “easy negative”– Similar approach can be done for “easy positives” depending on

applications

30

How to collect training data?

Implicit relevance judgments from click logs – including clicked and unclicked results from SERP (avoids position bias)

Add easy negatives (avoids sampling bias)

Mining click stream

Approach: Relevance labels

Label = 5 (Most relevant)

Label = 0 (least relevant)

Label = 2

Model

Learning to Rank

Pointwise: Reduce ranking to binary classification

33

+

+

+

-

Q1

+

-

-

-

Q2

+

+

-

-

Q3

Learning to Rank

Pointwise: Reduce ranking to binary classification

34

+

+

+

-

Q1

+

-

-

-

Q2

+

+

-

-

Q3

Learning to Rank

Pointwise: Reduce ranking to binary classification

35

+

+

+

-

Q1

+

-

-

-

Q2

+

+

-

-

Q3

Limitations Assume relevance is absolute Relevant documents associated with different queries are put into the

same class

Learning to Rank

Pairwise: Reduce ranking to classification of document pairs w.r.t. the same query

– {(Q1, A>B), (Q2, C>D), (Q3, E>F)}

36

Learning to Rank

Pairwise: Reduce ranking to classification of document pairs w.r.t the same query

– {(Q1, A>B), (Q2, C>D), (Q3, E>F)}

37

Learning to Rank Pairwise

– No longer assume absolute relevance– Limitation: Does not differentiate inversions at top vs. bottom positions

38

Listwise approach - DCG

Objective – Come up with a function to convert entire set of ranked search results, each with relevance labels into a score

Characteristics of such a function– Higher relevance in ranked set => higher score– Higher relevance in ranked set on higher positions => higher

score

p documents in the search results, each document ‘i’ has a relevance reli.

39

DCG

40

Rank Discounted Gain

1 3

2 4.4

3 0.5

7.9

NDCG based optimization

NDCG@k = Normalized(DCG@k) Ensures value is between 0.0 and 1.0 Since NDCG directly represents the “value” of particular

ranking given the relevance labels, one can directly formulate ranking as maximizing NDCG@k (say k = 5)

Directly pluggable into a variety of algorithms including coordinate ascent

41

Learning to Rank

42

Point wise

Simple to understand and debugStraight forward to use ✕Query independent ✕Assumes relevance is absolute

Pair wise

Assumes relevance is relativeDepends on query✕Loss function agnostic to position

List Wise

Directly operate on ranked listsLoss function aware of position✕More complicated, non-convex functions, higher

training time

43

Search Ranking

Click Logs Training Data ModelOffline

Evaluation

Online A/B test/debug

score = f(query, user, document)

tl;dr revisited

Ranking interacts heavily with retrieval and query understanding– Query understanding affects intent detection, fixing user errors etc.– Retrieval affects candidate selection, speed etc.

Ground truth > features > model*– Truth data is affected by biases

List wise > pair wise > point wise– Listwise while more complicated avoids some model level issues in

pairwise and point wise methods

44

* Airbnb engineering blog: http://nerds.airbnb.com/architecting-machine-learning-system-risk/

Useful references

“From RankNet to LambdaRank to LambdaMART: An overview” – Christopher Burges

“Learning to Rank for Information Retrieval” – Tie-Yan Liu

RankLib – has implementations of several LTR approaches

45

LinkedIn search is powered by …

46

We are hiring !!

careers.linkedin.com