Connecting Distributed People and Information on the Web

1
Connecting Distributed People and Information on
the Web

• Jennifer Golbeck
• College of Information Studies
• Human-Computer Interaction Lab
• University of Maryland, College Park

2
Information Access on the Web

• Find an mp3 of a song that was on the Billboard
  Top Ten that features a cowbell.

The Cowbell Project - http//www.geekspeakweekly.c
om/cowbell/
3
Finding Trusted Information

• How many cows in Texas?

http//www.cowabduction.com/
4
The Social Solution

• People are the sources of information
• Social relationships give us information about
  people
• Use relationships to understand the information
  people produce.

5
Current State

• 250-ish social networks
• 850,000,000 users
• Ning claims 185,000 networks

6
My Research Questions

• How do users behave and relate to one another in
  web-based social networks?
• How do social connections, like trust, relate to
  information?
• How can we estimate relationships (like trust)
  between people who do not know each other?
• How can we use social networks to build
  intelligent systems to improve information
  access?

7
Social Relationships and Information

• How Trust Relates to Similarity

8
A Study

• People create information on the web
• An expression of their opinions and view of the
  world
• Focus on quantitative information (e.g. ratings)
• People express trust in social networks
• How does trust relate to the similarity of two
  people

9
The Idea

• We know trust correlates with overall similarity
  (Ziegler and Golbeck, 2006)
• Does trust capture more than just overall
  agreement?
• Two Part Analysis
• Controlled study to find profile similarity
  measures that relate to trust
• Verification through application in a live system

10
Experimental Outline

• Phase 1 Rate Movies - Subjects rate movies on
  the list
• Ratings grouped as extreme (1,2,9,10) or far from
  average (4 different)
• Create profiles of hypothetical users
• Profile is a list of movies and the hypothetical
  users ratings of them
• Subjects rate how much they would trust the
  person represented by the profile
• Vary the profiles ratings in a controlled way

11
Phase 1 Rating Movies

• Movies most subjects would have seen - (100
  worldwide top grossing films of all time)
• Cover a broad spectrum of genres -
• Top 10 rated movies from each genre as listed in
  the Internet Movie Database (IMDB) Action,
  Adventure, Animation, Family, Comedy, Crime,
  Documentary, Drama, Fantasy, Film-Noir, Horror,
  Independent, Musical, Mystery, Romance, Science
  Fiction, Thriller, War, and Western.
• Include bad movies -(IMDB 100 worst rated movies
  with at least 1,000 ratings)
• 283 total films
• Ratings on 1 (bad) to 10 (great) scale

12
Generating Profiles

• Each profile contained exactly 10 movies, 4 from
  an experimental category and 6 from its
  complement
• E.g. 4 movies with extreme ratings and 6 with
  non-extreme ratings
• Control for average difference, standard
  deviation, etc. so we could see how differences
  on specific categories of films affected trust

13
Example Profile
14
Subjects

• 59 subjects
• Age 20 to 52
• Education
• 6 high school, 11 bachelors, 23 masters, 11 PhD,
  8 unreported
• Movie Experience
• Watch 1-2 times per week on average
• Movie media (web, magazines, etc.) every week or
  two

15
Results

• Reconfirmed that trust strongly correlates with
  overall similarity (?).
• Agreement on extremes (??)
• Largest single difference (r)
• Subjects propensity to trust (?)

16
Propensity to Trust (?)
17
Validation

• Gather all pairs of FilmTrust users who have a
  known trust relationship and share movies in
  common
• 322 total user pairs
• Develop a formula using the experimental
  parameters to estimate trust
• Compute accuracy by comparing computed trust
  value with known value

18
In FilmTrust

• Use weights (w1,w2, w3, w4, w?) (7,2,1,8,2)

19
Experimental Conclusions

• Social trust relationships are stronger between
  people who are similar in certain ways
• First observed in controlled experiments
• Verified through application in a real system

20
Applications

• Using social trust for improved information access

21
Social Information Access

• Use social relationships (e.g. trust) for
• Aggregating Information
• Sorting and Ranking Information
• Filtering and Assessing the Quality of
  Information
• FilmTrust

22
FilmTrust

• Use Trust for information access
• Recommender system
• Review ordering
• 1200 users

23
Information Aggregation Using Trust

• Trust-based Recommender System
• Generates predictive movie ratings based on trust
• Weighted average of everyones ratings of the
  film,where trust is the weight

24
(No Transcript)
25
Difference between known user rating and
recommended rating (measured in number of stars
difference)
Minimum difference between known user rating and
average rating
26
(No Transcript)
27
(No Transcript)
28
Conclusions andFuture Directions
29
Conclusions - Social Information Access

• Use understanding and analysis of social behavior
  in web-based social networks to improve
  information access
• Shown a connection between social trust and
  similarity
• Shown how trust can be used for aggregating,
  sorting, and filtering information

30
Future Directions - General

• Improved understanding of behavior in web-based
  social networks
• How different types of social connections relate
  to information
• How to improve information access using new
  social analyses

31
Future Directions - Specific

• Ad hoc information and social networks for micro
  news
• E.g. I have evacuated for natural disaster
  (earthquake, hurricane, flood). I want to know
  whats going on at my house.
• Distributed information (satellite photos,
  ground, video, photos, blog entries, local news
  reports, message board text)
• Needs
• Provenance - is this information unique, or is it
  all derived from the same source?
• Trust - should I trust the source of this
  information?

32
Questions

• Jennifer Golbeck
• golbeck_at_cs.umd.edu
• http//trust.mindswap.org

33
(No Transcript)
34
Generating Profiles

• Pre-defined rating differences
• Subjects rated 54 total profiles
• Six categories
• Three ? values
• Three profiles in each ?-category combination

35
The Provenance Challenge

• Researchers in many areas
• Storage systems
• Databases
• Grid computing
• Data mining
• A challenge provides a standard for comparing
  approaches
• Given a scientific workflow and nine challenge
  queries
• Represent all data that we consider relevant
  about the history of each file
• Answer as many queries as possible

36
FilmTrust Results

• FilmTrust compared trust from the social network
  with overall similarity (via collaborative
  filtering algorithms) as a weight in recommender
  systems.
• Trust outperformed overall similarity in some
  cases, suggesting that trust captures something
  more than overall similarity does

37
Ten Largest WBSNs

• MySpace 150,000,000
• ChinaRen Xiaonei 60,000,000
• Adult Friend Finder 26,000,000
• Bebo 25,000,000
• Friendster 21,000,000
• Cyworld 21,000,000
• Tickle 20,000,000
• Black Planet 18,000,000
• Hi5 14,000,000
• LiveJournal 12,000,000

38
Example Queries

• Find everything that caused a given Graphic to be
  as it is.
• Find all invocations of procedure align_warp
  using a twelfth order nonlinear 1365 parameter
  that ran on a Monday.
• Find all images where at least one of the input
  files had an entry global maximum4095.
• A user has annotated some images with a
  key-value pair centerUChicago. Find the outputs
  of align_warp where the inputs are annotated with
  centerUChicago.

39
Semantic Web Approach

• Ontology represents information about the
  execution of services and the dependencies among
  files
• Logical inferences connect objects to their
  ancestors
• Role hierarchy separates direct lineage from
  ancestry
• Semantics of transitive roles imply connections
  among files connected through ancestral
  relationships
• Additional reasoning with Semantic Web Rules

40
Evaluation through Query Answering

• SPARQL, a W3C standard, is used to formulate
  queries
• We were easily able to answer all nine queries
  for the challenge (one of only 3 teams from 15
  entries)
• Have already completed the second phase of the
  challenge, importing data from other systems and
  applying our techniques

41
Definition

• A Web-based Social Network (WBSN) must meet these
  criteria
• Accessible over the web with a web browser
• Users must explicitly state their relationship
  with other people qua stating a relationship
• Must have explicit built-in support for users
  making these connections.
• Relationships must be visible and browsable

(back)
42
Why the Difference?

• Ranges of disconnected members
• Dogster and HAMSTERster have lowest rates
• Ecademy
• FilmTrust
• Mobango and Worldshine
• As the non-social networking purpose of the
  website becomes stronger, the number of
  friendless and outsiders increases

(back)
43
Using Web-BasedSocial Networks (WBSNs)

• If we are going to use social networks for
  information access we must understand
• How do users behave in social networks?
• How do social relationships relate to information?

44
Daily Trends
45
Implications

• The trust we have in people can inform how we
  treat information provided by those people
• This and other studies suggest trust will work
  well for aggregating, filtering and sorting
  information
• Important when working on the web

46
Outline

• Motivation
• Understanding Relationships in Web-based Social
  Networks
• Behavior
• Trust
• Using Social Relationships for Information Access
• Conclusions and Future Directions

47
Understanding Social Behavior

• In Web-Based Social Networks

48
Behavior and Dynamics

• Social networks are not static.
• Relationships constantly change, are formed, and
  are dropped.
• New people enter the network and others leave
• Do people behave the same way in social networks
  on the Web?

49
Questions

• How do these networks grow (and shrink)?
• How are relationships added (and removed)?
• What affects social disconnect?
• What affects centrality?

50
Methodology

• 24 month study
• Automatically collected adjacency lists (everyone
  and who they know), join dates, and last active
  dates for all members.
• December 2004
• December 2006
• For 7 networks, I collected adjacency lists every
  day for 7 weeks.
• Who joined or left
• What relationships were added or removed

51
Network Growth

• People do not leave social networks
• On sites with a clear simple process, less than a
  dozen members leave per day
• In some networks, essentially no one has ever
  left
• Lots of people join social networks
• For ten networks we knew the date that every
  member joined the network
• Networks tend to show linear growth
• The slope can shift
• Usually occurs suddenly
• Explained by some event

52
(No Transcript)
53
Relationships

• Forming relationships is the basis for social
  networking
• Almost all networks are growing denser
• Relationships grow at approximately 1.7 - 2.7
  times the rate of membership
• There is a strong social disincentive to remove
  relationships

54
FilmTrust Network
55
Friendless and the Outsiders

• Friendless have no social connections
• Outsiders have social connections but are
  independent from the major connected component of
  the network
• Important because if we are using the social
  network for information access, these people will
  get little benefit.

56
(No Transcript)
57
Centrality

• Other than having lots of friends, what makes
  people more central?
• Average shortest path length as centrality
  measure
• Activity
• Consider join date, last active date, and length
  of activity (last active date - join date)
• Compute rank correlation with centrality
• Medium strength correlation (0.5) between
  duration and centrality

58
Conclusions

• Networks follow a linear growth pattern, where
  the slope shifts in response to events
• People rarely leave networks
• Networks grow denser, with relationships added
  more frequently than members
• People will delete relationships, but orders of
  magnitude less frequently than they add them
• Websites with more non-social features tend to
  have more friendless and disconnected users
• Users with longer periods of activity tend to be
  more central to the network

59
Example Profile

• Movies m1 through m10
• User ratings r1r10 for m1m10
• r1r4 are extreme (1,2,9, or 10)
• r5r10 are not extreme
• Profile ratings pi ri?i