Internet Search (COSC 488) Nazli Goharian nazli@cs.georgetown.edu - - PDF document

1

1

Internet Search

(COSC 488) Nazli Goharian

nazli@cs.georgetown.edu

Nazli Goharian, 2005, 2012 2

Outline

• Web: Indexing & Efficiency

– Partitioned Indexing – Index Tiering & other early termination techniques – Index in Dynamic Environment

• Improving effectiveness of Web search engines

– Web page ranking

• Query log, anchor text, authority/hub, page rank, sponsored search, localized search,

social search

– Result snippets

• Social Search

– tagging, collaborative search/filtering, recommender system – Real-time search

• Peer-to-Peer Search

2

2

3

The Web

• Document collections are scattered across many

geographical areas.

• Constraints prohibiting the centralization of data

include:

– Data security – Volume – Rate of change – Political and legal constraints – Other proprietary motivations

4

Web Search

• Parallel and distributed processing
• Web search tools access data distributed on servers

worldwide but indexed centrally.

• Most of these systems have a partitioned index on

large clusters of servers with a centralized control.

• They store pointers in the form of hypertext links to

various Web servers.

3

5

Partitioned Indexing

• Partitioning of index across multiple machines, based
• n either:
• Terms (Global index organization)
• Each node holds posting list for some terms
• Using content-index, query terms sent to nodes having the terms
• Higher concurrency level, but larger postings lists
• Documents (Local index organization)– more common
• Each node holds a complete index (shorter PLs)
• Query terms sent to all nodes
• Top k results from each node merged
• Global statistics (e.g.. idf) must be calculated
• A Hybrid approach in Tiered Indexing may be used

6

Index Tiering

• A popular early termination technique to improve

the efficiency of query processing

• Dividing nodes into two tiers to allocate the index
• f most popular documents on tier 1 and the rest on

tier 2.

• Search tier 1 first, if not enough results then search

tier 2.

• Note: other popular early termination techniques (top-doc and query pruning)

were discussed earlier in the semester!

6

4

Distributed Index Construction

• Not possible on a single machine
• Various architecture for distributed indexing
• MapReduce architecture (a term-partitioned

index)

• Master node assigns tasks to worker nodes (map

workers & reduce workers) to split up the computing jobs:

• Map Phase: Parsing & building localized <term, doc> pairs
• Reduce Phase: Combining/merging posting pairs for each term

7

MapReduce (Cont’d)

• Map & reduce phases can be done in parallel on many machines
• A map machine can be a reducer machine in the process
• Data broken into pieces (shards)…generally 16M-64 M [128M]

and send to map workers as they finish their job

• Map workers work on one shard at a time (generally), unless

having more than one CPU, parse and generate <term,doc> pair (can be combined to <term,doc,tf>

• Sort based on term, and then secondary key (doc_id)
• The same keys (terms) are assigned to the same reduce worker
• Load should be balanced on the reducers

8

5

MapReduce (Cont’d)

9

Taken from: C. Manning, P. Raghavan, H. Schutze, Introduction to Information Retrieval, Cambridge University Press, 2008

10

Query Servers

• Each server has its own disk holding a portion of

index

• Queries are distributed, via a centralized control, to

servers that contain the related posting lists

• Common terms may map to many servers
• No single point of resource contention (efficient)
• If a server crashes, that portion of index is not

available

6

11

Index in Dynamic Environment

• Data collection is not static
• Reconstruct the index periodically from scratch

(many search engines use this)

• Maintain an auxiliary index to store new document

& remerge with existing index

• Maintain multiple indexes - complicated in

maintaining collection statistics

12

Outline

• Web: Indexing & Efficiency

– Partitioned Indexing – Index Tiering & other early termination techniques – Index in Dynamic Environment

• Improving effectiveness of Web search engines

– Web page ranking

• Query log, anchor text, authority/hub, page rank, sponsored search, localized search,

social seacrh

– Result snippets

• Social Search

– tagging, collaborative search/filtering, recommender system – Real-time search

• Peer-to-Peer Search

12

7

Definitions….

• Web graph: each page is a node and links are directed edges

from one node to other node

• Out-links (out-degree) A: links from page A to B
• In-links (in-degree) A: links from other pages to A
• Sink: if out-links = 0
• Source: if in-links=0
• Static page: pages that are generated prior to any request
• Dynamic page: pages that generated as the result of a request
• Hidden/deep web: pages with no links/password protected/via a

Form,…

• Indexable Web: union of pages indexed by major search engines

13 14

Evaluation of Web Search Engines: High Precision Search

• Traditional IR systems are evaluated based on

precision and recall.

• Web search engines are evaluated based on top N

documents.

• Recall estimation is very difficult
• Precision is of limited concern, as many users do not look

beyond 1st screen. => How fast and accurate the first results screen is generated?

8

15

Web Page Ranking

• Considering both query dependant and query

independent scores (captured during indexing), a global score is generated for each page:

• Query dependant score
• Similarity measures such as Cosine, BM25, proximity,…
• Query independent score
• Link analysis (anchor text, popularity metrics such as:

authorities and hub, page rank,…)

• Sponsored search
• Localized search
• Query log analysis
• etc.

15 16

Query Log Analysis

• Using user query patterns on certain days and

time of day, week, month, and year, many

• ptimizations are possible:
• Pre-cache likely Web pages in anticipation of user

queries to reduce page access delays; increasing system throughput (efficiency optimization)

• Possible to adjust relevance ranking to tune for

certain user queries (accuracy optimization)

9

17

Anchor Text

• Short, 2-3 terms, describe the linked/destination page.
• May/may not be a different point of view than the

author’s.

• Anchor text of links to a doc di included in index for di
• Extended anchor text (text surrounding anchor text) may

also be used

• Generally weighted based on frequency (notion of idf)
• Spamming problem

18

Page Rank

• A scoring mechanism in Web search (trade marked by Google and patented by Stanford)
• Generally calculated at the time of crawling
• Using incoming and outgoing links as an indicator of

popularity, adjusts Web page score

• Popular page is defined as a page that
• Many Web pages link to it (inlinks)
• Important (popular) pages link to it
• May be affected by link spam

10

19

Page Rank

∑

+ − =

n

D D i i

D C D PageRank d N d A PageRank

...

1

) ( ) ( ) 1 ( ) ( C(Di) : number of links out from page Di d : damping factor (from 0-1; commonly 0.85) N: total number of pages

An Iterative Algorithm:

Initially all pages are assigned an arbitrary page rank (1/n), summing to 1 Iteratively calculate the scores until the new scores do not change significantly To converge faster, may initialize page ranks based on number of inlinks, log info,….

20

Authorities and Hub

• Various algorithms based on assigning each retrieved

web page two scores: Authority and Hub scores. (HITS: Hyperlink-Induced Topic Search, 1999)

• Authority page: an authoritative source on a given topic
• Hub page: page listing pointers to authority pages on a topic
• Authority score: summation of scores of all the hubs pointing

to that authority page

• Hub score: summation of scores of all authority pages the hub

is pointing to

11

21

Computing Authority and Hub Scores

• Retrieve all pages containing the query term t. This is

called root set. (~200 pgs)

• Create a set including union of root set pages, pages that

point to root set pages, and pages that root set pages point

• to. This is called base set.
• Using the base set to compute the hub and authority

scores.

• An iterative algorithm:
• Initialize hubs and authorities a score of 1
• Update s(H) and s(A)

22

Sponsored Search

• Search system vendors sell advertisers keywords so

that whenever such words are issued in a query, the advertiser’s desired homepage link is returned.

• Sponsored search results are biased towards

advertisers with higher bids, click frequency of Ads,…

• Significant revenue is generated to search engine

vendors via such search approach (ex.: per click (50 sents to

15 dollars)

12

23

Sponsored Search

• Search engines maintain an advertisement database

(Description of advertisement, link to that page, bids, popularity,…)

• Searching the advertisement database for a match to:
• query terms
• keywords extracted from retrieved result page

(pseudo-relevance feedback, page features, …)

• Ranking advertisements based on bids (on keywords)

and advertisement popularity (using clickthrough data logs)

24

Localized Search

• Using geographic information to modify the ranking of

results (in addition to SC scores, link based scores,…).

• Geographic information maybe derived from:
• Location of device sending the query
• Context of query
• restaurant near Al Capone’s home’s town
• restaurant Near White Sox stadium
• Geographic location in the query
• Chicago restaurants
• Geographic location in a document metadata

24

13

25

Result Snippets

• Providing users a short summary (snippet) of page (title,

url, link to cached page, snippet).

• Static snippets
• Query independent
• Created at indexing time and cached
• Containing title, n number of sentences/words, (NLP can be used)
• Dynamic snippets
• Query dependent
• Created at the time of results scoring
• “Windows” of the document - also called KWIC (keyword in context)

25 26

Result Snippets

• Index maintains sentence level information
• Snippet sentences can be picked:
• Based on query term(s):
• heading
• Location in document (nth sentence)
• Closeness of query terms in sentence
• Ratio of query terms in sentence
• Unique query terms in sentence
• From page metadata

26

14

27

Result Snippets

• An effective snippet should:(Clarke et al 2007’s clickthrough analysis)

– have all the query terms (unless already included in title) – Use the page metadata, if needed – Display URL and mark the query terms – Provide meaningful snippets vs. only some keywords

27 28

Outline

• Web: Indexing & Efficiency

– Partitioned Indexing – Index Tiering & other early termination techniques – Index in Dynamic Environment

• Improving effectiveness of Web search engines

– Web page ranking

• Query log, anchor text, authority/hub, page rank, sponsored search, localized search,

social search

– Result snippets

• Social Search

– tagging, collaborative search/filtering, recommender system – Real-time search

• Peer-to-Peer Search

28

15

29

Social Search

• Social search introduces new aspects to search engines

– Village paradigm (Collaborative) [Horowitz & Kamvar, WWW’10]

• Crowd/ Social network /friends vs. Corpus-based
• Routing questions to potential answerers
• Community of users, sharing goal or interest,

participate in search and interact with each other online

– YouTube, Twitter, Flickr, Facebook, Myspace, LinkedIn, forums, blogs, online games,…

From Wikipedia: “Social search or a social search engine is a type of web search that takes into account the Social Graph of the person initiating the search query. When applied to web search this Social-Graph approach to relevance is in contrast to established algorithmic or machine-based approaches where relevance is determined by analyzing the text of each document or the link structure of the documents”

29 30

Real-Time Search

• Traditional search indexes the crawled pages
• Real-time search results of search engines such as

Google, Bing, Yahoo come from variety of real-time search services such as twitter, flicker, your-tube, etc.

– Receive data directly from various social media and blogs (subscribed to social networking sites) – A filtering engine identifies spams – Measuring relevance -- The ranking is based on:

• Time,
• relevance to query,
• number of followers of authors,
• reputation of a link defined by the frequency of forwarding (re-tweets), …

First real time search: “Summize” in 2007 with real time trend analysis– later

• n merged with twitter 2008)

30

16

31

Social Search

• Documents or websites are deemed relevant if

searcher’s social network were also interested in it.

• Nature of queries

– Many cases opinionated, subjective – Query length (Many cases longer queries than Web’s)

• Index

– Storing user’s behavior ( responsiveness, answer quality, expertise) – Mapping users to topics

31 32

Social Search

• Social Search Ranking – based on combination of:

– Query-dependent (prob. of a good answer to query q by user u)

• Similarity of results to query (various ranking: cosine, bm25,

proxomity,…)

• Relatedness of query/results to user

– Query-independent

• How many users bookmarked x
• Social Trust
• Similarity of asker to answerer -- user profiles similarity, users’

connectedness

32

17

33

Social Search

• Mapping users to topics. An example: [Horowitz &

Kamvar, WWW’10]

– User specifies interest /expertise in topics – Friends of users indicate the expertise of user u in topics – Automatically identified topics from

• User’s existing online profiles
• User’s homepages, blogs
• User’s status messages (Twitter, Facebook, IM,…)

33 34

Social Search

• Measuring connectedness using cosine similarity
• ver various features, such as: [Horowitz & Kamvar, WWW’10]

– Social connection (common friends and affiliations) – Demographic similarity – Profile similarity (e.g., common favorite movies) – Vocabulary match (e.g., IM shortcuts) – Chattiness match (frequency of follow-up messages) – Verbosity match (the average length of messages) – Politeness match (e.g., use of “Thanks!”) – Speed match (responsiveness to other users)

34

18

35

Social Search

Sample approach: [Karweg, et.al, CIKM’ 11]:

• Social Relevance Score (SRS) ranks the result elements of a query

according to their social relevance for the user. It is calculated based on 2 factors: – Engagement Intensity: how intense the users interacted with the result

• Engagement: Interaction in terms of recommendation, rating, status messages
• Intensity: effort of textual feedback vs. rating score /thumps up

– Trust Score : level of trust to those who recommend a link

• Assigned by users & refined by social network analysis using page-rank on

social graph

35

∑

∈

=

i

E x x s

i e x t i SRS ) ( ). ( ) (

• SRS(i): social rank score of document/page i
• X: a user in social network interacted/recommended page i

36

Social Search -- Trust

• Trust has been discussed for years in sociology and social

psychology

• [Marsh , Ph.D. dissertation,1994] formalized trust as a computational

concept (agents that keep history of behaviors)

• Trust in peer-to-peer, EigenTrust [Kamvar et al. 2004] (corrupt vs.

valid files)

• Various efforts in formalization of trust in recommender systems

and social network [Swearingen and Sinha ,2001], [Ziegler and Golbeck [2006].

– The more similar two people were, the greater the trust between them [Ziegler

and Golbeck [2006].

– “Trust in a person is a commitment to an action based on a belief that the future actions of that person will lead to a good outcome.” Example: “Alice trusts Bob regarding email if she chooses to read a message (commits to an action)

that Bob sends her (based on her belief that Bob will not waste her time”

19

37

Tagging

• Social media sites allow users to tag the data
• User tags act as manual indexing of data – in addition

to automatic indexing

• User tags serve as folksonomy
• Tags are used to organize and search data
• Challenges with the tagged data:

– Vocabulary mismatch – Noisy or Spam tags – Missing tags

37 38

Searching Tagged Data:

Vocabulary Mismatch problem

• Tag keywords describe textual or non-textual data

and are used to search for items

– Tags are very sparse (only few keywords) – Boolean (conjunctive, disjunctive) search can lead to high precision/low recall or high recall/low precision – To reduce the vocabulary mismatch perform stemming, or pseudo-relevance feedback

38

20

39

Searching Tagged Data:

Noisy and Spam Tags

• Spam/misspelled/non-relevant tags mislead search

– Some incentive must be provided to users to report spam tags, and to enter good quality tags. – Log and statistical information may help to identify spam tags

39 40

Searching Tagged Data:

Missing Tags Automatically generate tags for items with missing tags, using:

– Term weight of textual representation of item – Classification of item to a label (i.e.. Tag)

40

21

41

Tag Clouds

• The most popular tags are represented to users to

provide a more wide view of collection

• Tag cloud displays the tags as a weighted list
• The font size is proportional to the weight

Thanks to: www.wordle.net for tagcloud generator & F. Silvestri, CNR, Italy, S. Orlando, U. of Venice, Italy

41

Recommender Systems