Dynamics of Web: Analysis and Implications from Search Perspective

Web Dynamics
Analysis and Implications from
Search Perspective
1

Lecturers
Dr. Ismail Sengör Altingövde
– Senior researcher
– L3S Research Center
– Hannover, Germany
Dr. Nattiya Kanhabua
– Postdoctoral researcher
– L3S Research Center
– Hannover, Germany
2

L3S Research Center
 Computer science and interdisciplinary research
on all aspects related to the Web
3

People
 80 researchers: 53% post-docs & PhD students,
38% scientific research assistants & int’l interns
 15 countries from US and Latin-America, Asia,
Europe and Africa
 Equal opportunity: 27% females
4

All project inventory!
www.l3s.de
5

L3S: Some Projects
Cubrik: Human-enhanced
time-aware multimedia search
ARCOMEM: From collect-all
archives to community memories
Terence: Adaptive learning
for poor comprehenders
M-Eco: The medical ecosystem
detecting disease outbreaks
6

Outline
Day 1: Introduction
• Evolution of the Web
• Overview of research topics
• Content and query analysis
Day 3: Indexing the Past
• Indexing and searching
versioned documents
Day 2: Evolution of Web
Search Results
• Short-term impacts
• Longitudinal analysis
Day 4: Retrieval and
Ranking
• Searching the past
• Searching the future
8

Evolution of the Web
• Web is changing over time in many
aspects:
– Size: web pages are added/deleted all the time
– Content: web pages are edited/modified
– Query: users’ information needs changes,
entity-relationship changes over time
– Usage: users’ behaviors change over time
[Ke et al., CN 2006; Risvik et al., CN 2002]
[Dumais, SIAM-SDM 2012; WebDyn 2010]9

Size dynamics
• Challenge
– Crawling, indexing, and caching
10

2000
First billion-URL index
The world’s largest!
≈5000 PCs in clusters!
1995 2012
Web and index sizes
11

2000
≈5000 PCs in clusters!2004
Index grows to
4.2 billion pages
1995 2012
Web and index sizes
12

2000
≈5000 PCs in clusters!2004
Index grows to
4.2 billion pages
1995 2012
2008
Google counts
1 trillion
unique URLs
Web and index sizes
13

Web and index sizes
http://www.worldwidewebsize.com/
15

Content dynamics
• Challenge
– Document representation and retrieval
WayBack Machine: a web archive search tool by the Internet Archive
16

Content dynamics
• Challenge
– Document representation and retrieval
17

Query dynamics
• Challenge
– Time-sensitive queries
– Query understanding and processing
Google Insights for Search: http://www.google.com/insights/search/
Query: Halloween
18

Usage dynamics
• Challenge
– Browsing and search behavior
– User preference (e.g., “likes”)
19

• Searching documents created/edited over time
– E.g., web archives, news archives, blogs, or emails
Example application
20

Example application
Web
archives
news
archives
blogs emails
“temporal document
collections”
21

Example application
Web
archives
news
archives
blogs emails
“temporal document
collections”
Retrieve documents
about Pope Benedict
XVI written before 2005
Term-based IR approaches
may give unsatisfied results
22

• A web archive search tool by the Internet Archive
– Query by a URL, e.g., http://www.ntnu.no
Wayback Machine1
1
Retrieved on 15 January 2011
23

Wayback Machine1
No keyword query
1
24

Wayback Machine1
No keyword query
No relevance ranking
1
25

• A news archive search tool by Google
– Query by keywords
– Rank results by relevance or date
Google News Archive Search
26

• A news archive search tool by Google
– Query by keywords
– Rank results by relevance or date
Google News Archive Search
Not consider
terminology
changes over time
27

Research topics
• Content Analysis
– Determining timestamps of documents
– Temporal information extraction
• Query Analysis
– Determining time of queries
– Named entity evolution
– Query performance prediction
• Evolution of Search Results
– Short-term impacts on result caches
– Longitudinal analysis of search results
28

Research topics (cont’)
• Indexing
– Indexing and query processing techniques for the
versioned document collections
• Retrieval and Ranking
– Searching the past
– Searching the future
29

Content Analysis
(1) Determining timestamps of documents
(2) Temporal information extraction

Motivation
• Incorporating the time dimension into
search can increase retrieval effectiveness
– Only if temporal information is available
• Research question
– How to determine the temporal information of
documents?
[Kanhabua, SIGIR Forum 2012]
31

Two time aspects
1. Publication or modified time
• Problem: determining timestamps of documents
• Method: rule-based technique, or temporal language
models
2. Content or event time
• Problem: temporal information extraction
• Method: natural language processing, or time and
event recognition algorithms
32

Two time aspects
• Problem: determining timestamps of documents
• Method: rule-based technique, or temporal language
models
• Problem: temporal information extraction
• Method: natural language processing, or time and
event recognition algorithms
content time
publication time
33

Problem Statements
• Difficult to find the trustworthy time for web documents
– Time gap between crawling and indexing
– Decentralization and relocation of web documents
– No standard metadata for time/date
Determining time of documents
34

Problem Statements
I found a bible-like
document. But I have
no idea when it was
created?
““ For a given document with uncertain
timestamp, can the contents be used to
determine the timestamp with a sufficiently
high confidence? ””
35

Problem Statements
Let’s me see…
This document is
probably
written in 850 A.C.
with 95% confidence.
I found a bible-like
document. But I have
no idea when it was
created?
““ For a given document with uncertain
timestamp, can the contents be used to
determine the timestamp with a sufficiently
high confidence? ””
36

Current approaches
1. Content-based
2. Link-based
3. Hybrid
37

Content-based approach
Partition Word
1999 tsunami
1999 Japan
1999 tidal wave
2004 tsunami
2004 Thailand
2004 earthquake
Temporal Language Models
Temporal Language
Models
• Based on the statistic usage
of words over time
• Compare each word of a
non-timestamped document
with a reference corpus
• Tentative timestamp -- a
time partition mostly
overlaps in word usage
[de Jong et al., AHC 2005]
Freq
1
1
1
1
1
1
38

Partition Word
1999 tsunami
1999 Japan
1999 tidal wave
2004 tsunami
2004 Thailand
2004 earthquake
tsunami
Thailand
A non-timestamped
document
Temporal Language
Models
of words over time
Freq
1
1
1
1
1
1
39

Partition Word
1999 tsunami
1999 Japan
1999 tidal wave
2004 tsunami
2004 Thailand
2004 earthquake
tsunami
Thailand
A non-timestamped
document
Temporal Language
Models
of words over time
Freq
1
1
1
1
1
1
40

Partition Word
1999 tsunami
1999 Japan
1999 tidal wave
2004 tsunami
2004 Thailand
2004 earthquake
tsunami
Thailand
A non-timestamped
document
Temporal Language
Models
of words over time
Freq
1
1
1
1
1
1
41

Partition Word
1999 tsunami
1999 Japan
1999 tidal wave
2004 tsunami
2004 Thailand
2004 earthquake
tsunami
Thailand
A non-timestamped
document
Similarity ScoresSimilarity Scores
Score(1999) = 1
Score(2004) = 1 + 1 = 2 Most likely timestamp is 2004
Temporal Language
Models
of words over time
Freq
1
1
1
1
1
1
42

Normalized log-likelihood ratio
Partition Word
1999 tsunami
1999 Japan
1999 tidal wave
2004 tsunami
2004 Thailand
2004 earthquake
tsunami
Thailand
A non-timestamped
document
Similarity ScoresSimilarity Scores
Score(1999) = 1
Score(2004) = 1 + 1 = 2 Most likely timestamp is 2004
Normalized log-likelihood
ratio
• Variant of Kullback-Leibler
divergence
• Similarity of a document and
time partitions
• C is the background model
estimated on the corpus
• Linear interpolation
smoothing to avoid the zero
probability of unseen words
[Kraaij, SIGIR Forum 2005]
43

Improving temporal LMs
• Enhancement techniques
1. Semantic-based data preprocessing
2. Search statistics to enhance similarity scores
3. Temporal entropy as term weights
Intuition: Direct comparison between extracted words
and corpus partitions has limited accuracy
Approach: Integrate semantic-based techniques into
document preprocessing
[Kanhabua et al., ECDL 2008]
44

Intuition: Search statistics Google Zeitgeist (GZ) can
increase the probability of a tentative time partition
Approach: Linearly combine a GZ score with the
normalized log-likelihood ratio
45

Intuition: A term weight depends on how good the
term is for separating time partitions (discriminative)
Approach: Propose temporal entropy, based on a
term selection presented in Lochbaum and Streeter
46

Semantic-based preprocessing
Intuition: Direct comparison between extracted
words and corpus partitions has limited accuracy
Approach: Integrate semantic-based techniques
into document preprocessing
Semantic-based
Preprocessing
Description
Part-of-speech tagging Select only interesting classes of words, e.g. nouns, verbs, and adjectives
Collocation extraction Co-occurrence of different words can alter the meaning, e.g. “United States”
Word sense disambiguation Identify the correct sense of a word from context, e.g. “bank”
Concept extraction Compare concepts instead of original words, e.g. “tsunami” and “tidal wave”
have the common concept of “disaster”
Word filtering Select the top-ranked words according to TF-IDF scores for a comparison
47

Leveraging search statistics
48

49

P(wi) is the probability that wi occurs:
P(wi) = 1.0 if a gaining query
P(wi) = 0.5 if a declining query
f(R) converts a ranked
number into weight. The
higher ranked query is
more important.
An inverse partition
frequency, ipf = log N/n
50

Temporal entropy
Temporal Entropy
A measure of temporal information which a word conveys.
Captures the importance of a term in a document collection
whereas TF-IDF weights a term in a particular document.
Tells how good a term is in separating a partition from others.
A term occurring in few partitions has higher temporal
entropy compared to one appearing in many partitions.
The higher temporal entropy a term has, the better
representative of a partition.
tern is for separating time partitions (discriminative)
51

Temporal entropy
52

Temporal entropy
Np is the total number of
partitions in a corpus
53

Temporal entropy
A probability of a partition
p containing a term wi
Np is the total number of
partitions in a corpus
54

Link-based approach
• Dating a document using its neighbors
1. Web pages linking to the document
• Incoming links
1. Web pages pointed by the document
• Outgoing links
1. Media assets associated with the document
• E.g., images
• Averaging the last-modified dates of its
neighbors as timestamps
[Nunes et al., WIDM 2007]
55

Hybrid approach
• Inferring timestamps using machine
learning
– Exploit links, contents of a web pages and its neighbors
– Features: linguistic, position, page formats, and tags
[Chen et al., SIGIR 2010]
56

Temporal information extraction
• Three types of temporal expressions
1. Explicit: time mentions being mapped directly to
a time point or interval, e.g., “July 4, 2012”
2. Implicit: imprecise time point or interval, e.g.,
“Independence Day 2012”
3. Relative: resolved to a time point or interval
using other types or the publication date, e.g.,
“next month”
[Alonso et al., SIGIR Forum 2007]
57

Current approaches
• Extract temporal expressions from
unstructured text
– Time and event recognition algorithms
• Harvest temporal knowledge from semi-
structured contents like Wikipedia
infoboxes
– Rule-base methods
[Verhagen et al., ACL 2005; Strötgen et al., SemEval 2010]
[Hoffart et al., AIJ 2012]
58

References
• [Alonso et al., SIGIR Forum 2007] Omar Alonso, Michael Gertz, Ricardo A. Baeza-Yates: On the
value of temporal information in information retrieval. SIGIR Forum 41(2): 35-41 (2007)
• [Chen et al., SIGIR 2010] Zhumin Chen, Jun Ma, Chaoran Cui, Hongxing Rui, Shaomang Huang: Web
page publication time detection and its application for page rank. SIGIR 2010: 859-860
• [Dumais, SIAM-SDM 2012] Susan T. Dumais: Temporal Dynamics and Information Retrieval. SIAM-
SDM 2012
• [de Jong et al., AHC 2005] Franciska de Jong, Henning Rode, Djoerd Hiemstra: Temporal language
models for the disclosure of historical text. AHC 2005: 161-168
• [Kanhabua et al., ECDL 2008] Nattiya Kanhabua, Kjetil Nørvåg: Improving Temporal Language
Models for Determining Time of Non-timestamped Documents. ECDL 2008: 358-370
• [Kanhabua, SIGIR Forum 2012] Nattiya Kanhabua: Time-aware approaches to information retrieval.
SIGIR Forum 46(1): 85 (2012)
• [Ke et al., CN 2006] Yiping Ke, Lin Deng, Wilfred Ng, Dik Lun Lee: Web dynamics and their
ramifications for the development of Web search engines. Computer Networks 50(10): 1430-1447
(2006)
• [Kraaij, SIGIR Forum 2005] Wessel Kraaij: Variations on language modeling for information retrieval.
SIGIR Forum 39(1): 61 (2005)
• [Nunes et al., WIDM 2007] Sérgio Nunes, Cristina Ribeiro, Gabriel David: Using neighbors to date
web documents. WIDM 2007: 129-136
• [Risvik et al., CN 2002] Knut Magne Risvik, Rolf Michelsen: Search engines and Web dynamics.
Computer Networks 39(3): 289-302 (2002) 59

References (cont’)
• [Risvik et al., CN 2002] Knut Magne Risvik, Rolf Michelsen: Search engines and Web dynamics.
Computer Networks 39(3): 289-302 (2002)
• [Strötgen et al., SemEval 2010] Jannik Strötgen, Michael Gertz: Heideltime: High quality rule-based
extraction and normalization of temporal expressions. SemEval 2010: 321-324
• [WebDyn 2010] Web Dynamics course: http://www.mpi-
inf.mpg.de/departments/d5/teaching/ss10/dyn/, Max-Planck Institute for Informatics, Saarbrücken,
Germany, 2010
• [Verhagen et al., ACL 2005] Marc Verhagen, Inderjeet Mani, Roser Sauri, Jessica Littman, Robert
Knippen, Seok Bae Jang, Anna Rumshisky, John Phillips, James Pustejovsky: Automating Temporal
Annotation with TARSQI. ACL 2005
60

Query Analysis
(1) Determining time of queries
(2) Named entity evolution
(3) Query performance prediction

Temporal queries
• Temporal information needs
– Searching temporal document collections
• E.g., digital libraries, web/news archives
– Historians, librarians, journalists or students
• Temporal queries exist in both standard
collections and the Web
– Relevancy is dependent on time
– Documents are about events at particular time
[Berberich et al., ECIR 2010]
63

Distribution over time of Qrel
Recency query Time-sensitive query
Time-insensitive query
[Li et al., CIKM 2003]
64

Types of temporal queries
• Two types of temporal queries
1. Explicit: time is provided, "Presidential election 2012“
2. Implicit: time is not provided, "Germany World Cup"
• Temporal intent can be implicitly inferred
• I.e., refer to the World Cup event in 2006
• Studies of web search query logs show a
significant fraction of temporal queries
– 1.5% of web queries are explicit
– ~7% of web queries are implicit
– 13.8% of queries contain explicit time and 17.1% of
queries have temporal intent implicitly provided
[Nunes et al., ECIR 2008; Metzler et al., SIGIR 2009; Zhang et al., EMNLP 2010]
65

• Semantic gaps: lacking knowledge about
1. possibly relevant time of queries
2. terminology changes over time
Challenges with temporal queries
66

query
time1
time2
…
timek
suggest
67

query
time1
time2
…
timek
suggest
68

1. Possibly relevant time of queries
2. Named entity changes over time
69

Determining time of queries
• Problem statements
– Implicit temporal queries: users have no knowledge
about the relevant time of a query
– Difficult to achieve high accuracy using only keywords
– Relevant results associated to particular time not given
• Research question
– How to determine the time of an implicit temporal query
and use the determined time for improving search
results?
70

Current approaches
1. Query log analysis
2. Search result analysis
71

Query log analysis
• Mining query logs
– Analyze query frequencies over time for identifying
the relevant time of queries
– Re-rank search results of implicit temporal queries
using the determined time
[Metzler et al., SIGIR 2009; Zhang et al., EMNLP 2010]
72

Search result analysis
• Using temporal language models
– Determine time of queries when no time is given explicitly
– Re-rank search results using the determined time
• Exploiting time from search snippets
– Extract temporal expressions (i.e., years) from the
contents of top-k retrieved web snippets for a given query
– Content-based language-independent approach
[Kanhabua et al., ECDL 2010; Campos et al., TempWeb
73

Approach I. Dating using keywords*
Approach II. Dating using top-k documents*
– Queries are short keywords
– Inspired by pseudo-relevance feedback
Approach III. Using timestamp of top-k documents
– No temporal language models are used
*Using Temporal Language Models proposed by de Jong et al.
Determining time of queries
74

I. Dating using keywords
75

I. Dating using keywords
Query’s temporal
profiles
76

II. Dating using top-k documents
77

II. Dating using top-k documents
Query’s temporal
profiles
78

III. Using timestamp of documents
79

III. Using timestamp of documents
Query’s tempora
profiles
80

• Intuition: documents published closely to the
time of queries are more relevant
– Assign document priors based on publication dates
Re-ranking search results
query
News archive
Determine time 2005, 2004, 2006, ...
D2009
Initial retrieved results
81

• Intuition: documents published closely to the
time of queries are more relevant
– Assign document priors based on publication dates
Re-ranking search results
query
News archive
Determine time 2005, 2004, 2006, ...
D2009
Initial retrieved results
D2005
Re-ranked results
82

Challenges of temporal search
1. Possibly relevant time of queries
2. Named entity changes over time
query
synonym@2001
synonym@2002
…
synonym@2011
suggest
83

Problem Statements
• Queries of named entities (people, company, place)
– Highly dynamic in appearance, i.e., relationships between
terms changes over time
– E.g. changes of roles, name alterations, or semantic shift
Named entity evolution
84

Problem Statements
Scenario 1
Query: “Pope Benedict XVI” and written before 2005
Documents about “Joseph Alois Ratzinger” are relevant
85

Problem Statements
Scenario 1
Query: “Pope Benedict XVI” and written before 2005
Documents about “Joseph Alois Ratzinger” are relevant
Scenario 2
Query: “Hillary R. Clinton” and written from 1997 to 2002
Documents about “New York Senator” and “First Lady of
the United States” are relevant
86

Research question
• How to detect named entity changes in web
documents?
87

QUEST Demo: http://research.idi.ntnu.no/wislab/quest/
88

Current approaches
1. Temporal co-occurrence
2. Temporal association rule mining
3. Temporal knowledge extraction
– Ontology
– Wikipedia history
89

Temporal co-occurrence
• Temporal co-occurrence
– Measure the degree of relatedness of two entities at
different times by comparing term contexts
– Require a recurrent computation at querying time,
which reduce efficiency and scalability
[Berberich et al., WebDB 2009]
90

Association rule mining
• Temporal association rule mining
– Discover semantically identical concepts (or named
entities) that are used in different time
– Two entities are semantically related if their
associated events occur multiple times in a collection
– Events are represented as sentences containing a
subject, a verb, objects, and nouns
[Kaluarachchi et al., CIKM 2010]
91

Temporal knowledge extraction
• YAGO ontology
– Extract named entities from the YAGO ontology
– Track named entity evolution using the New York
Times Annotated Corpus
• Wikipedia history
– Define a time-based synonym as a term semantically
related to a named entity at a particular time period
– Extract synonyms of named entities from anchor texts
in article links using the whole history of Wikipedia
[Mazeika et al., CIKM 2011; Kanhabua et al., JCDL 2010]
92

Searching with name changes
• Extract time-based synonyms from Wikipedia
– Synonyms are words with similar meanings
– In this context, synonyms refer name variants (name
changes, titles, or roles) of a named entity
• E.g., "Cardinal Joseph Ratzinger" is a synonym of
"Pope Benedict XVI" before 2005
• Two types of time-based synonyms
1.Time-independent
2.Time-dependent
[Kanhabua et al., JCDL 2010]
93

Recognize named entities
94

95

96

Find synonyms
• Find a set of entity-synonym relationships at time tk
97

Find synonyms
• For each ei ϵ Etk , extract anchor texts from article
links:
– Entity: President_of_the_United_States
– Synonym: George W. Bush
– Time: 11/2004
98

Find synonyms
links:
– Time: 11/2004
99

Find synonyms
links:
– Time: 11/2004
100

Initial results
• Time periods are not accurate
Note: the time of synonyms are timestamps of Wikipedia articles (8 years)
101

• Analyze NYT Corpus to discover accurate time
– 20-year time span (1987-2007)
• Use the burst detection algorithm
– Time periods of synonyms = burst intervals
Enhancement using NYT
102

– 20-year time span (1987-2007)
103

– 20-year time span (1987-2007)
Initial results
104

Query expansion
1. A user enters an entity as a query
[Kanhabua et al., ECML PKDD 2010]
105

Query expansion
106

Query expansion
2. The system retrieves synonyms wrt. the query
107

Query expansion
108

Query expansion
3. The user select synonyms to expand the query
109

Query expansion
3. The user select synonyms to expand the query
110

1. Performance prediction
– Predict the retrieval effectiveness wrt. a ranking model
Query prediction problems
111

query
precision = ?
recall = ?
MAP = ?
predict
112

2. Retrieval model prediction
– Predict the retrieval model that is most suitable
113

2. Retrieval model prediction
– Predict the retrieval model that is most suitable
query ranking = ?
predict max(precision)
max(recall)
max(MAP)
114

Problem Statement
• Predict the effectiveness (e.g., MAP) that a query will
achieve in advance of, or during retrieval
– high MAP  “good”
– low MAP  “poor”
Objective
• Apply query enhancement techniques to improve the
overall performance
– Query suggestion is applied for “poor” queries
Query performance prediction
[Hauff et al., CIKM 2008 ;Hauff et al., ECIR 2010; Carmel et al., 2010]
115

• First study of performance prediction for temporal
queries
– Propose 10 time-based pre-retrieval predictors
• Both text and time are considered
• Experiment
– Collection: NYT Corpus and 40 temporal queries
• Results
– Time-based predictors outperform keyword-based predictors
– Combined predictors outperform single predictors in most
cases
• Open issue
– Consider time uncertainty
Temporal query performance prediction
[Kanhabua et al., SIGIR 2011]
116

• Problem statement
– Two time aspects: publication time and content time
• Content time = temporal expressions mentioned in documents
– Difference in effectiveness for temporal queries when
ranking using publication time or content time
Time-aware ranking prediction
117

• First study of the impact on retrieval
effectiveness of ranking models using two time
aspects
• Three features from analyzing top-k results
– Temporal KL-divergence [Diaz et al., SIGIR 2004]
– Content Clarity [Cronen-Townsend et al., SIGIR 2002]
– Divergence of retrieval scores [Peng et al., ECIR 2010]
Learning to select time-aware ranking
118

• Measure the difference between the distribution
over time of top-k retrieved documents of q and the
collection
– Consider both time dimensions
Temporal KL-divergence
[Diaz et al., SIGIR 2004]
119

• The content clarity is measured by the Kullback-
Leibler (KL) divergence between the distribution of
terms of retrieved documents and the background
collection
Content Clarity
[Cronen-Townsend et al., SIGIR 2002]
120

• Measure the divergence of scores from the base
ranking, e.g., a non time-aware ranking model
– To determine the extent that a ranking model alters the
scores of the initial ranking
• Features
1. averaged scores of the base ranking
2. averaged scores of PT-Rank
3. averaged scores of CT-Rank
4. divergence from the base ranking model
Divergence of ranking scores
[Peng et al., ECIR 2010]
121

Discussion
• Results
– A small number of top-k documents achieves
better performance
– The larger number k, the more irrelevant
documents are introduced into the analysis
• Open issue
– When comparing with the optimal case there
is still room for further improvements
122

References
• [Berberich et al., WebDB 2009] Klaus Berberich, Srikanta J. Bedathur, Mauro Sozio, Gerhard
Weikum: Bridging the Terminology Gap in Web Archive Search. WebDB 2009
• [Berberich et al., ECIR 2010] Klaus Berberich, Srikanta J. Bedathur, Omar Alonso, Gerhard
Weikum: A Language Modeling Approach for Temporal Information Needs. ECIR 2010: 13-25
• [Campos et al., TempWeb 2012] Ricardo Campos, Gaël Dias, Alípio Jorge, Celia Nunes:
Enriching temporal query understanding through date identification: How to tag implicit temporal
queries? TWAW 2012: 41-48
• [Carmel et al., 2010] David Carmel, Elad Yom-Tov: Estimating the Query Difficulty for
Information Retrieval Morgan & Claypool Publishers 2010
• [Cronen-Townsend et al., SIGIR 2002] Stephen Cronen-Townsend, Yun Zhou, W. Bruce Croft:
Predicting query performance. SIGIR 2002: 299-306
• [Diaz et al., SIGIR 2004] Fernando Diaz, Rosie Jones: Using temporal profiles of queries for
precision prediction. SIGIR 2004: 18-24
• [Hauff et al., CIKM 2008] Claudia Hauff, Vanessa Murdock, Ricardo A. Baeza-Yates: Improved
query difficulty prediction for the web. CIKM 2008: 439-448
• [Hauff et al., ECIR 2010] Claudia Hauff, Leif Azzopardi, Djoerd Hiemstra, Franciska de Jong:
Query Performance Prediction: Evaluation Contrasted with Effectiveness. ECIR 2010: 204-216
• [Kaluarachchi et al., CIKM 2010] Amal Chaminda Kaluarachchi, Aparna S. Varde, Srikanta J.
Bedathur, Gerhard Weikum, Jing Peng, Anna Feldman: Incorporating terminology evolution for
query translation in text retrieval with association rules. CIKM 2010: 1789-1792
123

References (cont’)
• [Kanhabua et al., JCDL 2010] Nattiya Kanhabua, Kjetil Nørvåg: Exploiting time-based synonyms
in searching document archives. JCDL 2010: 79-88
• [Kanhabua et al., ECDL 2010] Nattiya Kanhabua, Kjetil Nørvåg: Determining Time of Queries for
Re-ranking Search Results. ECDL 2010: 261-272
• [Kanhabua et al., SIGIR 2011] Nattiya Kanhabua, Kjetil Nørvåg: Time-based query performance
predictors. SIGIR 2011: 1181-1182
• [Kanhabua et al., SIGIR 2012] Nattiya Kanhabua, Klaus Berberich, Kjetil Nørvåg: Learning to
select a time-aware retrieval model. (To appear) SIGIR 2012
• [Li et al., CIKM 2003] Xiaoyan Li, W. Bruce Croft: Time-based language models. CIKM 2003:
469-475
• [Metzler et al., SIGIR 2009] Donald Metzler, Rosie Jones, Fuchun Peng, Ruiqiang Zhang:
Improving search relevance for implicitly temporal queries. SIGIR 2009:
• [Mazeika et al., CIKM 2011] Arturas Mazeika, Tomasz Tylenda, Gerhard Weikum: Entity
timelines: visual analytics and named entity evolution. CIKM 2011: 2585-2588
• [Nunes et al., ECIR 2008] Sérgio Nunes, Cristina Ribeiro, Gabriel David: Use of Temporal
Expressions in Web Search. ECIR 2008: 580-584
• 700-701
• [Peng et al., ECIR 2010] Jie Peng, Craig Macdonald, Iadh Ounis: Learning to Select a Ranking
Function. ECIR 2010: 114-126
• [Zhang et al., EMNLP 2010] Ruiqiang Zhang, Yuki Konda, Anlei Dong, Pranam Kolari, Yi Chang,
Zhaohui Zheng: Learning Recurrent Event Queries for Web Search. EMNLP 2010: 1129-1139124

Evolution of Web Search
Results
(1) Short-term impacts: Caching Results
(2) Longitudinal analysis of search results
126RuSSIR 2012 126

Architecture of a Search Engine
• How search results change?
1. Crawl the Web
2. Index the content
3. Go to line 1
Crawler
Document
Collection
Indexer
Indexes
Query
processor
AnswersWWW
127

Architecture of a Search Engine
• How search results change?
1. Crawl the Web
2. Index the content
3. Go to line 1
Crawler
Document
Collection
Indexer
Indexes
Query
processor
AnswersWWW
on-line
128

Indexing the Dynamic Web
• How to refresh the index?
– Batch update (re-build)
– Re-merge
– In-place
• Batch update
– Shadowing
– Simplest, the old index can keep serving at
high rates
[Lester et al., IPM 2006]
130

• Re-merge
– A “buffer” B of new index entries
– Compute queries over index I and B and
merge results
– Merge B with I when size(B) > threshold
– Optimizations: logarithmic and geometric
merging
131

• In-place
– Over-allocation: Leave free space at the end
of each list
– Add new entries to the free space; o.w.,
relocate to a new position on disk
132

Query Processing
• Is updating the underlying index enough?
If a query is submitted for the first time, it is
processed over an up-to-date index
But... what about the cached results?
133

Why Result Caching?
• Around 50% of a query stream is
composed of repeating queries
[Fagni et al., TOIS 2006]
Query popularity: heavy tail distribution
134

Result Caches
• Result cache: top-k urls with snippets per query
• Results caching helps to reduce
– query response time
– traffic volume hitting the backend
“bin laden dead”
BACKEND
miss (compulsory)
hit
135

Problem: dynamicity!
• Key observations:
– Caches can be very large (thanks to cheap hw)
– Web changes rapidly (thanks to users)
“bin laden dead”
BACKEND
Minutes, days, or
weeks, based on Q
136

A deeper look: Capacity vs. Hits
• Query log :Yahoo! Search
engine
– Few cache nodes during 9 days
• Cache capacity
– can be very large
– eviction policy is less of a
concern
• Hit rate: fraction of queries
that are hits
• Higher capacity  more hits!
[Cambazoglu et al., WWW 2010] (Slide provided by the authors)
137

Capacity vs. age
• Hit age: time
since the last
update
• Higher capacity
 higher age!
138

Strategies for cache freshness
• Decoupled approaches: cache does not
know what changed
• Coupled approaches: cache uses clues
from the index to predict what changed
[Cambazoglu et al., WWW 2010]
139

Decoupled approaches: Flushing
• Solution: ﬂush periodically
– Coincides with new index
– Bounds average age
– Impacts hit rate negatively
140

Decoupled approaches:TTL
• Time-to-live (TTL)
– Assign a fixed TTL value to each cached result
– Pros: Practical, almost no implementation cost
– Cons: Blind strategy, may be sub-optimal
…
qi
Result Cache
q1 R1 TTL(q1)
q2 R2 TTL(q2)
qk Rk TTL(qk)
141

Decoupled approaches:TTL
• TTL per query
– Stale results  Acceptable: search engines
do not have a perfect view!
– Stable hit rate; average age is still bounded!
142

Intelligent Refreshing
• Enhancement to the TTL mechanism
• Updates entries
– Re-execute queries
– Uses idle cycles
• Ideally
– update: low activity
– use: high activity
• How to select queries for refreshing?
143

Intelligent refreshing
Temperature
• A new query goes
to young and cold
bucket!
• Fixed “query
interval” for shifting
age
• Lazy update of
temperature
• Refresh hot
and old first! 144

Refresh-rate adjustment
• Idle cycles?
• Critical mechanism
– Prevent overloads
• Feedback from the
query processors
– Latency to process
query
• Track recent latency
– Adjust rate accordingly
Query
Query Processors
Cache
Results,Latency
145

Design summary
• Large capacity: millions of entries
• TTL: bounds the age of entries
• Refreshes: updates cache entries
• Refresh rate adjustment: latency feedback
146

Cache evaluation
• Simulation
– Yahoo! query log
• Baseline policies
– No refresh
– Cyclic refresh
147

Cache evaluation
• Simulation
– Yahoo! query log
• Baseline policies
– No refresh
– Cyclic refresh
148

Performance in production
Refreshes on
Refreshes off
Refreshes on
Refreshes off
149

Degradation
• Under high load
– Processors degrade results
– Shorter life
• TTL mechanism
– Prioritize for refreshing
– AdjustsTTL when degraded
• Refreshes
– Replace with better results
150

Adaptive TTL
• Up to now we considered fixed TTL values
• Are all queries equal?
– “quantum physics”
– “barcelona FC”
– “ecir 2012”
• Another promising direction is assigning
adaptive TTL values
[Alici et al., ECIR 2012] (Slide provided by the authors)
151

Adaptive TTL
• Up to now we considered fixed TTL values
• Are all queries equal?
– “quantum physics”
– “barcelona FC”
– “ecir 2012”
• Another promising direction is assigning
adaptive TTL values
One size does not fit all!!!
152

Adaptive I: Average TTL
• Observe past update frequency of for
top-k results of a query and compute
average
• Simple, but
– Needs history
– May not capture bursty update periods
5 1 6 4
NOW: q submitted again!
TTL =( 5+1+6+4)/ 4= 4PAST: Update history
153

Adaptive II: Incremental TTL
• Adjust the new TTL value based on the
current TTL value
• Each time a cached result Rcached with an
expired TTLcurr is requested:
compute Rnew
if Rnew ≠ Rcached // STALE
TTLnew  TTLmin // catch bursty updates!
else
154

Adaptive III: Machine-learned
• Build an ML model
– Query and result-specific feature set F
– Assume a query result changes at time point ti
• Create a training instance with features Fi
• Set target feature (TTLnew) as the time period from ti
to the time point tj where query result changes
again: tj-ti
5 1 6 4
<F1, 5> <F2, 1> <Fn, ?>
t1
t2
155

Features for ML
156

Analysis of Web Search Results
• Setup:
– 4,500 queries sampled from AOL log
– Submitted to Yahoo! API daily from Nov 2010
to April 2011 (120 days)
– For days di and di+1
If Ri ≠ Ri+1 (consider top-10 urls and their order), we
consider the result as updated!
157

Distribution of result updates
• On average, query results are updated in
every 2 days!
158

Effect of query frequency
• More frequent  results updated more frequently
• Less frequent  scattered[Alici et al., ECIR 2012] (Slide provided by the authors)
159

Effect of query length
• No correlation between query length and
update frequency
160

Simulation
• Assume all 4,500 queries are submitted
daily for 120 days
– On day-0, all results are cached
– On the following days, whenever the TTL
expires, the new result is computed and
replaces old ones (i.e., result for that day from
Yahoo! API)
161

Simulation setup
• Evaluation metrics [Blanco 2010]
– At the end of each day, we compute:
False Positive Ratio =
Redundant query executions
Number of unique queries
Stale Traffic Ratio =
Stale results returned
Number of query occurrences
162

Simulation setup
• While computing ST ratio for a given query
at day i:
Strict policy:
– if Rcached is not strictly equal to Rday-i
staleCount++
Relaxed policy:
– if Rcached is not strictly equal to Rday-i
staleness += 1 – JaccardSim(Rcached, Rday-i)
163

Performance: Incremental TTL
Strict policy Relaxed policy
164

Less-often updated queries More-often updated queries
Performance: Incremental TTL
165

Performance: Average and Machine
Learned TTL
Strict policy Relaxed policy
166

Coupled approaches: CIP
• Cache invalidation policy (CIP)
• Key idea: Compare all
added/updated/deleted documents to
cached queries
– Incremental index update  updates reflected
on-line!
[Blanco et al., SIGIR 2010]
all queries in cache(s) all changes in the backend index
CIP module
167

Coupled approaches: CIP
Incremental index update: updates reflected on-line!
Content changes sent to CIP
module to invalidate queries (offline)
168

Synopses
• Synopsis: A vector of document’s top-
scoring TF-IDF terms
• η: length of synopsis
• δ: modification threshold  consider a document d
as updated only if diff(d, dold) > δ
169

Invalidation Approaches
• Invalidation logic (for added docs)
Basic: If the synopis matches the query
(i.e., involve all query terms)
Scored: Compute Sim(synopsis, query)
to the score of k-th result
Example
170

Invalidation Approaches
• Invalidation logic (for deleted docs):
– Invalidate all query results including a deleted
document
• Update is a deletion followed by an
addition
• Further apply TTL to guarantee an age-
bound
• Scored + TTL
171

Evaluation
• Really hard to evaluate if you are not
sitting at the production department in a
real search engine company!
172

Simulation Setup
• Data: English wikipedia dump
– snapshot at Jan 1, 2006 ≈ 1 million pages
– All add/deletes/updates for following 30 days
• Queries: 10,000 from AOL log
[Blanco et al., SIGIR 2010; Alici et al. SIGIR 2011]
173

Simulation setup
• Evaluation metric
– The query result is updated if two top-10 lists
are not exactly the same
174

Performance
• Setup
– Score + TTL
– Full synopsis
• Smaller δ or TTL:
– Lower ST, higher FP
TTL: 1≤t≤5
t=1
t=2
t=3
t=4
t=5
δ=0, 2≤t≤10
δ=0.005, 3≤t≤10
δ=0.01, 5≤t≤20
175

TIF: Timestamp-based Invalidation
Framewrok
• Devise a new invalidation mechanism
– better than TTL and close to CIP in detecting
stale results
– better than CIP and close to TTL in efficiency
and practicality
[Alici et al., SIGIR 2011] (Slide provided by the authors)
176

Timestamp-based Invalidation
• The value of the TS on an item shows the
last time the item was updated
• TIF has two components:
– Offline (indexing time) : Decide on term and
document timestamps
– Online (query time): Decide on the staleness of the
query result
177

TIF Architecture
…
Doc.
parser
documents assigned
to the node
SEARCH
NODE
index
updates
…
t1
t2
tT
178

TIF Architecture
…
Doc.
parser
documents assigned
to the node
SEARCH
NODE
index
updates
…
t1
t2
tT
…
Document timestamps
document TS
updates
TS(d1) TS(d2) TS(dD)
Document timestamps
179

TIF Architecture
…
Doc.
parser
documents assigned
to the node
SEARCH
NODE
index
updates
…
Document timestamps
document TS
updates
Document timestamps
180

TIF Architecture
…
Doc.
parser
documents assigned
to the node
SEARCH
NODE
index
updates
term TS updates
…
TS(t1)
TS(t2)
TS(tT)
…
Document timestamps
document TS
updates
Document timestamps
181

TIF Architecture
…
Doc.
parser
documents assigned
to the node
SEARCH
NODE
index
updates
qi
results
Result cache
0/1
Invalidation
logic
miss/stale
qi, Ri, TS(qi)
…
q1 R1 TS(q1)
q2 R2 TS(q2)
qk Rk TS(qk)
qi
results
Result cache
0/1
Invalidation
logic
miss/stale
qi, Ri, TS(qi)
…
q1 R1 TS(q1)
q2 R2 TS(q2)
qk Rk TS(qk)
term TS updates
…
TS(t1)
TS(t2)
TS(tT)
…
Document timestamps
document TS
updates
Document timestamps
182

TS Update Policies: Documents
• For a newly added document d
– TS(d) = now()
• For a deleted document d
– TS(d) = infinite
• For an updated document d
– if diff(dnew, dold) > L
TS(d) = now()
– diff(di, dj): |length(di) – length(dj)|
183

TS Update Policies: Terms
t
• Frequency based update
t
TS(t) = T0, PLLTS= 5
Number of added postings > F x PLLTS
TS(t) = now()
PLLTS= 6
184

TS Update Policies: Terms
• Score based update
t p1 p2 p3 p4 p5
TS(t) = T0, STS = Score(p3)p4 p3 p2 p5 p1
t p6
Score of added posting > STS TS(t) = now()
STS = re-sort & compute
p1 p2 p3 p4 p5
sort w.r.t. scoring function
185

Result Invalidation Policy
• A search node decides a result stale if:
– C1: ∃d ϵ R, s.t. TS(d) > TS(q)
(d is deleted or revised after the generation of query
result)
or,
– C2: ∀t ϵ q, s.t. TS(t) > TS(q)
(all query terms appeared in new documents
after the generation of query result)
• Also apply TTL to avoid stale accumulation
186

Simulation Setup
• Data: English wikipedia dump
– snapshot at Jan 1, 2006 ≈ 1 million pages
– All add/deletes/updates for following 30 days
• Queries: 10,000 from AOL log
[Alici et al., SIGIR 2011]
187

Simulation setup
• Evaluation metrics [Blanco 2010]
– The query result is updated if two top-10 lists
are not exactly the same
188

Performance: all queries
Frequency-based term TS update Score-based term TS update
189

Performance: single-term queries
Frequency-based term TS uıpdate Score-based term TS update
190

Invalidation Cost
Send <q, R, TS(q)> to
the search nodes
Send all <q, R> to CIP
Send all docs to CIP
TIF CIP
Data
transfer
Invalidation
operations Compare TS values
Traverse the query index
for every document
191

Performance: TIF
• A simple yet effective invalidation approach
• Predicting stale queries
– Better than TTL, close to CIP
• Efficiency and practicality
– Straightforward in a distributed system
192

Grand Summary
• Fixed TTL
– With refreshing
• Adaptive TTL
• TIF
• CIP
Decoupled
Decoupled
Coupled
Coupled
Hit rate
Hit age
ST Ratio
FP Ratio
ST Ratio
FP Ratio
ST Ratio
FP Ratio
STandFPratiodecreases
Invalidationcostdecreases
Yahoo!
Web sample
AOL
Wikipedia
AOL
Wikipedia
AOL
193

Open Research Directions
Investigate:
• User satisfaction vs. freshness
• Complex ranking functions
• Alternative index update strategies
• Combinations
• Adaptive TTL + TIF or CIP
• Adaptive TTL + refresh strategy
194

Evolution of Web Search Results
• Earlier works consider the changes in the
content of the
–Web
–queries
• How do real life search engines react this
dynamicity?
• Compare results from Yahoo! API
–for 630K real life queries (from AOL log)
–obtained in 2007 and 2010
[Altingovde et al., SIGIR 2011] (Slide provided by the authors)
195

What is Novel?
• Queries are real, not synthetic
• Query set is large
• Results from a search engine at two very
distant points in time
• Focus on the properties of results, but not
the underlying content
196

We Seek to Answer:
• How is the growth in Web reflected to top-
ranked query results?
• Do the query results totally change within
time?
• Are results located deeper in sites?
• Is there any change in result title and
snippet properties?
197

Avg no. of results
• Almost tripled (from 16M to 52M), but not
uniformly
198

No. of unique URLs
• 20% of the URLs returned at the highest rank in
2010 were at the same position in 2007!
199

No. of unique domains
• The increase in unique domain names in 2010 is
more emphasized in comparison to the increase
in the number of unique URLs (diversity?
coverage?)
• Even higher overlap for top-1 domains
200

Research Directions
• How are the results are diversified at
these two different time points?
–Can we deduce these from snippets?
• How does the level of bias changes in
query results?
201

References
• [Alici et al., SIGIR 2011] Sadiye Alici, Ismail Sengör Altingövde, Rifat Ozcan,
Berkant Barla Cambazoglu, Özgür Ulusoy: Timestamp-based result cache
invalidation for web search engines. SIGIR 2011: 973-982
• [Alici et al., ECIR 2012] Sadiye Alici, Ismail Sengör Altingövde, Rifat Ozcan, Berkant
Barla Cambazoglu, Özgür Ulusoy: Adaptive Time-to-Live Strategies for Query Result
Caching in Web Search Engines. ECIR 2012: 401-412
• [Altingovde et al., SIGIR 2011] Ismail Sengör Altingövde, Rifat Ozcan, Özgür
Ulusoy: Evolution of web search results within years. SIGIR 2011: 1237-1238
• [Blanco et al., SIGIR 2010] Roi Blanco, Edward Bortnikov, Flavio Junqueira, Ronny
Lempel, Luca Telloli, Hugo Zaragoza: Caching search engine results over
incremental indices. SIGIR 2010: 82-89
• [Cambazoglu et al., WWW 2010] Berkant Barla Cambazoglu, Flavio Paiva
Junqueira, Vassilis Plachouras, Scott A. Banachowski, Baoqiu Cui, Swee Lim, Bill
Bridge: A refreshing perspective of search engine caching. WWW 2010: 181-190
• [Fagni et al., TOIS 2006] Tiziano Fagni, Raffaele Perego, Fabrizio Silvestri,
Salvatore Orlando: Boosting the performance of Web search engines: Caching and
prefetching query results by exploiting historical usage data. ACM Trans. Inf. Syst.
24(1): 51-78 (2006)
• [Lester et al., IPM 2006] Nicholas Lester, Justin Zobel, Hugh E. Williams: Efficient
online index maintenance for contiguous inverted lists. Inf. Process. Manage. 42(4):
916-933 (2006)
202

Indexing the Past
(1) Indexing and searching methods for
versioned document collections
204RuSSIR 2012 204

Time Travel?
• Time travel is easy (in IR)! You only need:
– The data
– The mechanisms to search
• indexing + query processing
– (And fasten your seat belts!)
205

Why Search the Past?
• Historical information needs (an analyst
working on the social reactions on the net
after 9/11)
• To find relevant resources that do not exist
anymore
• To discover trends, opinions, etc. for a
certain time period (what people think
about UFOs at the beginning of
millenium?)
206

Data: Preserving the Web
• Non-profit organizations:
– Internet archive, European archive,...
• Several EU projects
– Planets, PROTAGE, PrestoPRIME, Scape,
ENSURE, BlogForever, LiWA, LAWA,
ARCOMEM...
http://archive.org/
207

Data
• There are also other “versioned” data
collections
– Wikis (Wikipedia)
– Repositories (code, document, organization
intranets)
– RSS feeds, news, blogs etc. with continuous
updates
– Personal desktops
Collections including multiple versions of
each document with a different timestamp
208

Indexing
• Various earlier approaches:
– [Anick et al., SIGIR 1992 ], [Herscovici et al, ECIR 2007]
• Focus on recent work from two different
perspectives
– Indexing schemes that are more concentrated
on the partitioning-related issues
• [Berberich et al., SIGIR 2007; Anand et al. CIKM 2010; Anand et al.
SIGIR 2011]
– Indexing schemes that are more concentrated
on the index size
• [He et al., CIKM 2009; He et al., CIKM 2010; He et al., SIGIR 2011]209

Time-travel Queries
• “Queries that combine the content and temporal
predicates” [Berberich et al., SIGIR 2007]
• Interesting query types
– Point-in-time
• “euro 2012 articles” @ 1/July/2012
– Interval
• “euro 2012 articles” between 01.06-01.07 2012
– Durability in a time interval [Hou U et al., SIGMOD 2010]:
• “search engine research papers” that are in top-10 results for
75% of the time between years 2000 and 2012
210

Time-point Queries: Indexing
• Formal model:
– For a document d, each version has
begin/end timestamps (validity interval):
• For the current version, de = ∞
• For d, all validity times are disjoint
t0
t1 t2
t3 ∞
d0
: [t0, t1) d3
: [t3, ∞)
[Berberich et al., SIGIR 2007]
211

Indexing
• Key idea
– Keep “validity intervals” within posting lists
v <d, tf> <d, tf, db, de>
t0
t1 t2
t3 ∞
tf(v)=1 tf(v)=0 tf(v)=2 tf(v)=2
v <d, 1, t0, t1>
From
<d, 2, t2, t3> <d, 2, t3, ∞>
• Problem: Index size explodes!
– Even for Wikipedia, 9*109
entries
212

Remedy: Coalescing
• Combine adjacent postings with similar
pay-loads
v <d, 1, t0, t1> <d, 2, t2, t3> <d, 2, t3, ∞>
v <d, 1, t0, t1> <d, 2, t2, ∞>
213

Optimization Problem
214

ATC
• Linear-time approximate algorithm
215

It works!
216

Temporal Partitioning (Slicing)
• Even after colescing, wasted I/O for
redundant postings that does not overlap
with query’s time point
v, [t0, t1) <d, 1, t0, t1>
<d, 2, t2, ∞>
Remark: Validity intervals still reside in the postings!
v, [t2, ∞)
217

Trade-off
• Optimal approaches
– Space-optimal
– Performance-optimal
• Trading off space and performance:
– Performance-Guarantee apparoach
– Space-Bound apparoach
218

Computational Complexity
• Performance-Guarantee Approach
– DP: time and space complexity O(n2
)
• Space-Bound Approach
– DP solution is prohibitively expensive
– Approximate solution using simulated
annealing
• Time O(n2
), Space O(n)
• Notice n is the number of all unique time interval
boundaries for a given term’s posting list
219

Time-point Queries: Result
• Given a space-bound of 3 (i.e., 3x of the
optimal space), close-to-optimal
performance is achievable!
– (Reminder: Partitioning is not very cheap!)
220

Time-point Queries: Result
• Time-point queries can be handled like
this:
• Example
221

Time-interval Queries?
• When more than one partitions should be
accessed, wasted I/O due to repeated
postings! (e.g., 3x more postings in SB!)
• Example
[Anand et al., CIKM 2010]
222

Solution Approaches
Partition selection [Anand et al., CIKM 2010]
Can we avoid accessing all partitions related to a given query?
Document partitioning (sharding) [Anand et al., SIGIR 2011]
Can we partition postings in a list in a different way
i.e., other than using time information?
223

Partition Selection
• Problem: Optimize result quality by
accessing a subset of “affected partitions”
without exceeding a given I/O upper-
bound
Optimization criterion: maximize the fraction of
original query results (relative recall).
Two types of constraints:
a) Size-based: allowed to read up to a fixed
number of postings (focus on sequential access)
b) Equi-cost: allowed to read up to a fixed
number of partitions (focus on random access)
224

Assumption
• Assume an oracle exists “providing us the
cardinalities of the individual partitions as
well as the result of their intersection/union”
• Later, KMV synopses will be employed as
an approximation
225

Single-term queries
• Size-based partition selection
• Equi-cost partition selection
• DP solutions exist, but expensive!
• Approximation
– Reduce the problem to budgeted max
coverage
N
226

GreedySelect Algorithm
• Cost(P):
– Size-based: no. of postings in the partition
– Equi-cost: 1
• Benefit(P):
– no of unread postings in P
• At each step:
– Select the partition with highest B(P) / C(P)
– Update benefits of the unselected partitions
Recall: we assume an oracle
providing these numbers!
227

Multi-term Queries
• Conjunctive semantics intersect
partitions of query terms
• Optimization objective: increase the
coverage of postings that are in this
intersection space of partitions
Size-based
N
Equi-cost
229

Simple Math does the job!
• Compute “union of intersections”
• x: a tuple of intersection from each term
• x is a tuple from the Cartesian product of
partitions for each query term
– X = Pterm1 x Pterm2 x..... X Ptermk and then,
– x = {(Pterm1,i, Pterm2,j, ..., x Ptermk, l)}
Choose a partition from this term’s partitions
230

GreedySelect, again!
• Apply GreedySelect over X to pick x ϵ X
• C(x)
– Sum of the sizes of unselected partitions in x,
or, number of unselected partitions in x
• B(x)
– No of new documents that appear in the
intersection of the partitions in x
• Modify algorithm to update benefits and
costs after each picking a tuple x.
231

Cartesian Product or Join?
• Problem: Set X might be very large!
• Remedy: Observe that partitions of
different terms that have no temporal
overlap cannot have any intersecting
docs!
Apply the algorithm
over the t-join set!
232

Performance of Partition Selection
• Reletive recall: 50% of affected partitions
might be enough!
– 3 datasets, Wikipedia seems harder!
Wikipedia UKGOV NY Times
233

Performance of Partition Selection
• Compare query run times for:
– Unpartitioned index,
– No partition selection,
– Partition selection with I/O bound (all index
files are compressed!)
234

Real Life: KMV Synopses
• Instead of assuming an “oracle” for
cardinality estimation, use KMV synopses
• A KMV synopsis for a multiset S:
– Fix a hash function h
– Apply h to each distinct value in S
– k-smallest of the hashed values form KMV
[Beyer et al., SIGMOD 2007]
Effective sketches for sets supporting arbitrary
multiset operations: ⋃, ∩, /
235

Partition Selection with KMV
• For each partition of each term:
– create a flat file storing KMV synopses (5%
and 10% samples)
236

Partition Selection with KMV
• KMV is promising
– 5% is enough!
Wikipedia UKGOV NY Times
237

Document partitioning (Sharding)
• Another solution to avoid reading and
processing repeated postings with cross-
cutting validity intervals in temporal
partitioning
• Example
[Anand et al., SIGIR 2011]
238

Sharding
• Key idea:
– Instead of partitioning temporally, partition
postings based on doc-ids
– No repetition as in slicing
– Example
239

Sharding
• Entries in a shard are ordered according to
their begin time
• For each shard, an axuliary data structure:
– List of pairs: (query begin times, offset in shard)
– Maintain for practical granularities of begin
times (like, days)  can fit in memory!
– For a query with a begin time
• Seek to the offset position in the shard and then read
sequentially
240

Why several shards?
• There can be still wasted disk I/O while reading a
shard:
• Problem is the postings with long validity intervals
(i.e., subsuming lot of other postings)
Query begin time
Earliest posting that includes
the query begin time!
This causes reading 5 useless
postings!
241

Idealized Sharding
• For a given posting list L, find a minimal set of
shards that satisfy staircase property (SP)
if begin(p) ≤ begin(q) → end(p) ≤ end(q)
• Greedy solution:
– For each posting p in L (in begin-time order)
• Add p in an exiting shard s if SP is satisfied
• Otherwise, start a new shard with p
242

Merging Shards
• Idealized sharding can yield several
shards
– random access per shard is expensive!
• A greedy merging algortihm
– Penalty <= CostRan /CostSeq
– Penalty (pairwise): wasted sequential reads
for merging two shards
– Merge in ascending choice order and then
smallest size order
More on this later!
243

Performance
• Sharding improves query processing
times w.r.t. temporal partitioning or no
partitioning et al.
• Merging shards results shorter QP times
• Time measurements are taken on warm
caches!!!
244

A Grand Summary
• A full posting list with validity intervals
– High sequential access + CPU cost
- 1 random access per query term
• Temporal partitioning (slicing)
– Reduce seq access (but, repetitions)
– Reduce CPU cost
– ≥1 random accesses per query term (with partition
selection)
• Document partitioning (sharding)
– Reduce seq access (no repetitions + time map)
– Reduce CPU cost (staircase property)
– ≥1 random accesses per query term (read all
shards?!)
TRADE-OFF: May increase random
accesses
for reducing sequantial accesses and CPU
cost
245

An Alternative Perspective
• Work from Polytechnique Ins. of NYU
• Focus on the index size
Approaches up to now consider each version
of a document separately: no special attention
on the overlap between versions
246

Index Compression
• Key ideas
– Small integers can be represented with
smaller codes
– Doc ids are not so small: instead, compress
the gaps between the ids
– Term frequencies are already small
– Example
247

Indexing Versions
• Assume validity-intervals are stored separately
 Time-constraints @ post-processing
• Versions of di are represented as di,j
t0
t1 t2
t3 ∞
tf(v)=1 tf(v)=0 tf(v)=2 tf(v)=2
v <d1, 1, t0, t1> <d1, 2, t2, t3> <d1, 2, t3, ∞>
v <d11, 1> <d1,2, 2> <d1,3, 2>
Question: How can we reduce the storage space for such a representation?
[He et al., CIKM 2009]
248

Versions Are Similar
• There is a high overlap between versions!
Snapshot@
May 15
Snapshot@
June 15
Snapshot@
July 15
249

How to Exploit?
• Simplest idea:
– Assign consecutive doc IDs to consecutive
versions of the same document
– Allows small d-gaps for overlapping terms
among the versions
<d1000, 1> <d3000, 2> <d6000, 2>
russir <d1, 1> <d2, 2> <d3, 2>
d1: http://romip.ru/russir2012/, [May 15, June 15]
d2: http://romip.ru/russir2012/, [June 15, July 15]
d3: http://romip.ru/russir2012/, [July 15, Aug 15]
russir Random ids
Consecutive ids
250

MSA Approach
• Multiple Segment Alignment [Herscovici et al., ECIR’07]
– Given d with some versions, a virtual
document Vi,j is all terms occuring in (only)
versions i through j
– Reduces the number of postings but
increases the document space! (theoretically,
up to N2
virtual documents!)
251

MSA: Example
Herscovici et al., ECIR 2007]
252

Indexing the Differences
• DIFF approach [Anick et al., SIGIR’92]
– For every pair of consecutive versions di and
di+1; create a virtual document that is the
symmetric difference between these versions.
253

DIFF: Example
• Example:
• QP
254

Performance
• Two index compression schemes
– Interpolative coding (IPC)
– PForDelta with optimizations (PFD)
• Two datasets: Wiki, Ireland
Wikipedia Ireland
Random-IPC 4957 2908
Random-PFD 5499 3289Index sizes for doc ids
Query processing times (in-memory): Sorted is the best!
255

Two-level Indexing: Roots
• Assume we have clusters of documents
t1
t2
d1
d2 2 d4 2
1 d2 2 d3 6
C1
d1
d2
d3
d4
C2
C3
• We have queries restricted to clusters:
– {t1, t2} in C3  intersect lists  d2, d4  d4
d4 1
256

Cluster-skipping index
• Skip redundant postings during QP
– Reduce decompression cost [Altingovde et al., TOIS 2008]
• Skipping is widely used for QP in practice
t1
t2
C1 d1 1 C2 d2 2 d3 6
C2 d2 2 C3 d4 2
C3 d4 1
[Altingovde et al., TOIS 2008]
257

Two-level indexing
• Apply the same idea for versions:
– First level index: document ids
– Second level index: a bitvector for versions
No need for indexing seperate ids for each
version  implicit from the bicvector.
258

Two-level indexing
• In actual implementation, group blocks of
doc ids and bitvectors of versions
• Bitvectors best compressed by hierarchical
Huffman coding!
russir <d1, 1> <d2, 2> <d3, 2>
russir <d> 1 1 1 0 0/1: First three versions include “russir”
russir <d> 1 2 2 0 TFs: Actual term frequencies
259

Two-level indexing: HUFF
• Query Processing: Decompress bitvectors
of documents that are in the intersection
of query terms
russir <d1> 1 2 2 0 <d2> 1 1 2 0 <d4> 1 0 1 0
schedule <d3> 1 2 2 1 <d4> 1 2 2 0 <d5> 1 3 2 1
Final result: d4,1 and d4,3
260

Two level indexing
• Same idea can be applied to MSA and
DIFF
– For virtual documents, create bitvectors as
before
– Compress first level IPC, second level PFD
• Even better compression performance
Index sizes for doc ids
HUFF 213 577 261

Two level indexing: QP
• Queries without temporal constaints
[He et al., SIGIR 2011]
262

• Queries with temporal constaints
– Post-processing
263

• Queries with temporal constaints
– Partitioning, again!
264

QP Performance
265

Grand Summary
Partition-focused
– Time information kept
in postings
– Redundancy solved
by partitioning
• Process only relevant
partitions
– Lossy compression of
versions (coalescing)
Size-focused
– Time information kept
separately
– Redundancy solved by
2-level compression
• Process compact blocks
• Hierarchical partitions
– Lossless compression
of versions
266

References
• [Altingovde et al., TOIS 2008] Ismail Sengör Altingövde, Engin Demir, Fazli Can, Özgür Ulusoy:
Incremental cluster-based retrieval using compressed cluster-skipping inverted files. ACM Trans.
Inf. Syst. 26(3): (2008)
• [Anand et al., CIKM 2010] Avishek Anand, Srikanta J. Bedathur, Klaus Berberich, Ralf
Schenkel: Efficient temporal keyword search over versioned text. CIKM 2010: 699-708
• [Anand et al., SIGIR2011] Avishek Anand, Srikanta J. Bedathur, Klaus Berberich, Ralf Schenkel:
Temporal index sharding for space-time efficiency in archive search. SIGIR 2011: 545-554
• [Anick et al., SIGIR 1992] Peter G. Anick, Rex A. Flynn: Versioning a Full-text Information
Retrieval System. SIGIR 1992: 98-111
• [Berberich et al., SIGIR 2007] Klaus Berberich, Srikanta J. Bedathur, Thomas Neumann,
Gerhard Weikum: A time machine for text search. SIGIR 2007: 519-526
• [Beyer et al., SIGMOD 2007] Kevin S. Beyer, Peter J. Haas, Berthold Reinwald, Yannis
Sismanis, Rainer Gemulla: On synopses for distinct-value estimation under multiset operations.
SIGMOD Conference 2007: 199-210
• [He et al., CIKM 2009] Jinru He, Hao Yan, Torsten Suel: Compact full-text indexing of versioned
document collections. CIKM 2009: 415-424
• [He et al., CIKM 2010] Jinru He, Junyuan Zeng, Torsten Suel: Improved index compression
techniques for versioned document collections. CIKM 2010: 1239-1248
• [He et al., SIGIR 2011] Jinru He, Torsten Suel: Faster temporal range queries over versioned
text. SIGIR 2011: 565-574
• [Herscovici et al, ECIR 2007] Michael Herscovici, Ronny Lempel, Sivan Yogev: Efficient
Indexing of Versioned Document Sequences. ECIR 2007: 76-87
• [Hou U et al., SIGMOD 2010] Leong Hou U, Nikos Mamoulis, Klaus Berberich, Srikanta J.
Bedathur: Durable top-k search in document archives. SIGMOD Conference 2010: 555-566
267

Retrieval and Ranking Models
(1)Searching the past
(2)Searching the future

RECAP
Two time dimensions
270

RECAP
Two time dimensions
content time
publication time
271

Searching the past
• Historical or temporal information needs
– A journalist working the historical story of a particular
news article
– A Wikipedia contributor finding relevant information
that has not been written about yet
272

Searching the past
• Historical or temporal information needs
– A journalist working the historical story of a particular
news article
– A Wikipedia contributor finding relevant information
that has not been written about yet
273

• Time must be explicitly modeled in order to
increase the effectiveness of ranking
– To order search results so that the most relevant
ones come first
• Time uncertainty should be taken into account
– Two temporal expressions can refer to the same time
period even though they are not equally written
– E.g. the query “Independence Day 2011”
• A retrieval model relying on term-matching only will fail to
retrieve documents mentioning “July 4, 2011”
Challenge
274

Query- and document model
• A temporal query consists of:
– Query keywords
– Temporal expressions
• A document consists of:
– Terms, i.e., bag-of-words
– Publication time and temporal expressions
275

Query- and document model
• A temporal query consists of:
– Query keywords
– Temporal expressions
• A document consists of:
– Terms, i.e., bag-of-words
– Publication time and temporal expressions
276

Time-aware ranking models
• Follow two main approaches
1. Mixture model [Kanhabua et al., ECDL 2010]
• Linearly combining textual- and temporal similarity
1. Probabilistic model [Berberich et al., ECIR 2010]
• Generating a query from the textual part and temporal part
of a document independently
277

Mixture model
• Linearly combine textual- and temporal similarity
– α indicates the importance of similarity scores
• Both scores are normalized before combining
– Textual similarity can be determined using any term-
based retrieval model
• E.g., tf.idf or a unigram language model
278

Mixture model
• Linearly combine textual- and temporal similarity
– α indicates the importance of similarity scores
• Both scores are normalized before combining
– Textual similarity can be determined using any term-
based retrieval model
• E.g., tf.idf or a unigram language model
How to determine temporal similarity?
279

Temporal similarity
• Assume that temporal expressions in the query are
generated independently from a two-step
generative model:
– P(tq|td) can be estimated based on publication time
using an exponential decay function [Kanhabua et al.,
ECDL 2010]
– Linear interpolation smoothing is applied to eliminates
zero probabilities
• I.e., an unseen temporal expression tq in d
280

Temporal similarity
• Assume that temporal expressions in the query are
generated independently from a two-step
generative model:
– P(tq|td) can be estimated based on publication time
using an exponential decay function [Kanhabua et al.,
ECDL 2010]
– Linear interpolation smoothing is applied to eliminates
zero probabilities
• I.e., an unseen temporal expression tq in d
Similarityscore
Time distance
d1 d2
281

• Five time-aware ranking models
– LMT [Berberich et al., ECIR 2010]
– LMTU [Berberich et al., ECIR 2010]
– TS [Kanhabua et al., ECLD 2010]
– TSU [Kanhabua et al., ECLD 2010]
– FuzzySet [Kalczynski et al., Inf. Process. 2005]
Comparison of time-aware ranking
282

• Experiment
– New York Times Annotated Corpus
– 40 temporal queries [Berberich et al., ECIR 2010]
• Result
– TSU outperforms other methods significantly for most
metrics
• Conclusions
– Although TSU gains the best performance, but only
applied to a collection with time metadata
– LMT, LMTU can be applied to any collection without time
metadata, but time extraction is needed
Discussion
283

Searching the future
• People are naturally curious about the future
– What will happen to EU economies in next 5 years?
– What will be potential effects of climate changes?
284

Previous work
• Searching the future
– Extract temporal expressions from news articles
– Retrieve future information using a probabilistic
model, i.e., multiplying textual similarity and a time
confidence
• Supporting analysis of future-related
information in news and Web
– Extract future mentions from news snippets obtained
from search engines
– Summarize and aggregate results using clustering
methods, but no ranking
[Baeza-Yates SIGIR Forum 2005; Jatowt et al., JCDL 2009]
285

Recorded Future
[http://www.recordedfuture.com/]
286

Yahoo! Time Explorer
[Matthews et al., HCIR 2010]
287

Ranking news predictions
• Over 32% of 2.5M documents from Yahoo! News
(July’09 – July’10) contain at least one prediction
• Retrieve predictions related to a news story in news
archives and rank by relevance
[Kanhabua et al., SIGIR 2011a]
288

Dynamics of Web: Analysis and Implications from Search Perspective

More Related Content

Viewers also liked (18)

Similar to Dynamics of Web: Analysis and Implications from Search Perspective (20)

Recently uploaded (20)

Dynamics of Web: Analysis and Implications from Search Perspective

Editor's Notes