NSF grant

1
NSF grant 0429452

• Life in the Network
• The Coming Era of Computational Social Science

David Lazer Harvard University Netsci 2007
2
Life in the Network

• Much of human civilization has been about
  building network infrastructure (or taking
  advantage of naturally occurring infrastructure,
  such as rivers)
• Benefits of trade
• Economies of scale
• Human drive to connect
• Cities, roads, railroads, telephone lines
• There has been a proliferation of various types
  of ICT networks connecting people in the last few
  decades

3
Digital traces of our networked lives

• E-mail
• Instant messaging
• Text messaging
• Telephone logs
• Link structure among websites
• Facebook
• Web surfing http//www.iq.harvard.edu/blog/netgo
  v/
• What can data like these tell us?

4
Existing approaches to studying social networks

• Growth in interest in these kinds of relational
  phenomena
• Generally rely on self reports
• Static generally based on snapshots
• Shaky reliability what is being measured by self
  reports?
• Small scale mostly systems in the hundreds or
  less
• ? Inferential challenges in existing research
• ? Many important phenomena are neglected

5
What can data like these tell us?

• How do things spread through a network?
• Ideas?
• Avian flu?
• How do people/organizations work together?
• Collaboration and coordination?
• Who is in key positions in the network?
• Form an empirical basis for various types of
  policy recommendations
• Possibly even real-time feedback for effective
  interventions

6
Computational social science

• The capturing and analysis of human activity
  represented in digital form
• Increased computational capacity to manipulate
  data
• Incidental, vast archives of human activity
  (e.g., Internet, e-mail)
• Instrumentation of human behavior (e.g., cookies,
  GPS devices)
• Creation of virtual worlds to experiment with
• What are the implications for our understanding
  of collective human behavior?

7
What can be done with data like these? Four
studies

• Call log analysis
• Instrumentation of human behavior
• Natural language processing
• Building virtual worlds

8
Study 1 Call log analysis

• Structure and tie strengths in mobile
  communication networks (just came out in PNAS,
  with J.-P. Onnela, J. Saramäki, J. Hyvönen, G.
  Szabó,K. Kaskil, J. Kertész, A.-L. Barabási )
• Examination of call log data from mobile phone
  company in moderate sized European nation a
  total of approximately 7,000,000 users, 49
  trillion dyads
• What does network structure look like?
• Small world? (six degrees of separation Watts
  Strogatz)
• Scale free? (hub-spoke structure Barabasi and
  Albert)
• Strength of weak ties? (Granovetter)

9
Call log network data
10
Results

• Hub-spoke structure (scale free)
• Small world (on average, 13 degrees of
  separation)
• But poorly structured for dissemination Strong
  ties tend to be clustered, and weak ties bind
  clusters together (consistent with Granovetter)
• Simulations suggest that weak(est) ties are not
  effective at spreading (inconsistent with
  Granovetter)
• Potentially powerful tool for studying evolving
  social structures of communities
• Possible use of data for a variety of policy
  purposes, from criminal investigations to early
  warning system for avian flu
• But what does a phone call between two phones
  mean??

11
Study 2 Instrumentation of human behavior

• Paper Revealing Social Relationships using
  Contextualized Proximity and Communication Data
  (with Nathan Eagle and Sandy Pentland)
• Collaboration with Media Lab
• Program mobile phones of 100 students for 9
  months
• Call log data
• Physical proximity (using Bluetooth)
• Location (using cell tower triangulation)
• Also collected self report data on friendship,
  satisfaction
• What is the information in these data?
• Compare observations to self reports

12
Self reported vs observed proximity

• Substantial recency effects recent interactions
  weighted more heavily
• Reciprocal non-friends 99.5 accurate at
  reporting 0s
• Reciprocal friends 35 accurate at reporting
  0s
• Friends more accurate at non-0s

13
Is friendship observable?

• Friendship is important at individual and
  collective levels due to the resources that flow
  among friends
• Purely cognitive relationship in principle,
  you could be friends with someone with whom you
  do not interact.
• But generally we all make inferences about who is
  friends with whom based on our observations
• Can the types of information that inform our
  inferences be captured via our mobile phones?
• Certainly, one anticipates that (for ex) friends
  will tend to be proximate to each other
• If high accuracy is possible, then possible to
  look at evolution of friendship structure in
  larger populations over time (as well as other
  cognitive relationships, such as advice)

14
Predicting friendships

• Relational scripts culturally-embedded patterns
  of relational behavior
• We generated seven relational variables
• Frequency of phone calls
• Proximity at home
• Proximity at work
• Proximity outside work
• Proximity on Saturday nights
• Proximity with no signal
• Number of unique locations
• Interactions broke into two factors

15
Predicting friendships

• Relational scripts culturally-embedded patterns
  of relational behavior
• We generated seven relational variables
• Frequency of phone calls
• Proximity at home
• Proximity at work
• Proximity outside work
• Proximity on Saturday nights
• Proximity with no signal
• Number of unique locations
• Interactions broke into two factors
• In-role communication
• Extra-role communication

16
Reported friendships
17
Inferred friendships (based on extra-role factor)
18
Self reported versus observed friendships

• We were able to categorize correctly 95 of
  reciprocated friendships and reciprocated
  non-friendships
• Unreciprocated friendships came from high
  scores in-role communication, perhaps capturing
  cultural ambiguity
• Created continuous construct from dichotomous
  self report perhaps a more valid measure of
  friendship?
• Second layer of validation predicting
  satisfaction based on (a) actual friendships and
  (b) inferred friendship. Second model does
  slightly better.
• We follow culturally embedded programs with
  respect to our relationships
• Results suggest potential for inferring
  friendship on much larger scale.

19
Current data collection the sociometer

• More collaboration with the Media Lab
  sociometers
• Study of teams (with Nancy Katz)
• When do team members talk to each other, and who
  does the talking?
• What difference does this make at individual and
  collective level?

20
Study 3 Automated content analysis of
Congressional websites

• Acknowledgement NSF grant 0429452, Stephen
  Purpura
• Advances in computational linguistics and voice
  recognition software
• Importance amplified by simultaneous improvements
  in voice recognition software
• Code human coders categorize blocks of text
  (training set and test set)
• Train classification algorithms
• Test against test data

21
Official Congressional websites

• Every House member has a website at www.House.gov
• Example http//www.house.gov/capuano/
• Strategic calculus of what message to send to
  constituents
• Comparable set of websites, create panel data
  set, some things varying over time, some not

22
(No Transcript)
23
Official Congressional websites

• Every House member has a website at www.House.gov
• Example http//www.house.gov/capuano/
• Strategic calculus of what message to send to
  constituents
• Comparable set of websites, create panel data
  set, some things varying over time, some not

24
Official Congressional websites

• Track evolution of language usage
• Examples positive/negative mentions of Bush?
• For example, what was/is strongest single word
  predictor of partisanship of Member of Congress?
• In 2001 terror (Republicans used more than
  Democrats)
• In 2006 Iraq (Democrats used more than
  Republicans)
• Allows us to see what is flowing through networks

25
Study 4 Building virtual deliberation world

• Acknowledgements Kevin Esterling, Michael
  Neblo, Curt Ziniel NSF grant 0429452
  
• Creation of virtual space (use of Macromedia
  Breeze)
• Twenty deliberative sessions with Members of
  Congress regarding immigration
• Allows complete control and recording of
  interactions, recruitment of representative
  sample
• Pretest, various control groups, during session
  questions, post-test a week later, and
  post-election survey
• Analysis still pending (dramatic effects on
  approval and vote intention)

26
Virtual deliberation world
27
Study 4 Building virtual deliberation world

• Acknowledgements Kevin Esterling, Michael
  Neblo, Curt Ziniel NSF grant 0429452
  
• Creation of virtual space (use of Macromedia
  Breeze)
• Twenty deliberative sessions with Members of
  Congress regarding immigration
• Allows complete control and recording of
  interactions, recruitment of representative
  sample
• Pretest, various control groups, during session
  questions, post-test a week later, and
  post-election survey
• Analysis still pending (dramatic effects on
  approval and vote intention)

28
Computational social science

• Orders of magnitude increase in data being
  collected about human behavior over last decade
• Constant increase in computational power
• Shift in social science research over the next
  generation
• Thinking relationally what is flowing among
  people? How are people working together?

29
The big picture

• The capturing of massive amounts of digitalized
  information about human behavior (especially
  relational behavior)
• The capacity to manipulate those data
• New insights into collective human behavior

30
Challenges, Caveats, and Conundrums

• Overcoming silos of academia, particularly wide
  between the sciences and social sciences
• The need to develop new infrastructures within
  social sciences
• Substantial human subjects issues
• Partnerships with those that are guardians of the
  network
• Concerns about use of knowledge that is produced
  (ex of NSA wiretaps, private sector data mining)
• Figuring out what those insights are time to
  shift the paradigm

31
NSF grant 0429452

• Computational Social Science

Partially supported by NSF grant 0429452
32
NSF grant 0429452

• Computational Social Science

Partially supported by NSF grant 0429452
continue at http//www.iq.harvard.edu/blog/netgo
v/
33
(No Transcript)
34
(No Transcript)