Bridging the Gap Between Physical Location and Online Social Networks

1. Bridging the Gap Between Physical
Location and Online Social Networks
J. Cranshaw, E. Toch, J. I. Hong, A. Kittur, and N. Sadeh.
In Proceedings of the 12th ACM International Conference on Ubiquitous
Computing, Copenhagen, Denmark, September 2010.
Presented by Beibei Yang
UMass Lowell 91.650, Spring 2011
Tuesday, April 12, 2011 1

2. Overview
• Examines the location traces of 489 users
• Introduces location-based features for analyzing
geographic regions
‣ location entropy
• Provide model for predicting friends
• Identify relationships between users’ mobility patterns and
structural properties of their underlying social network
• Potential design and research of online social networks on
offline mobility
Tuesday, April 12, 2011 2

3. Motivation
• Difficult distinction of online and offline social networks
• Open ended debate:
‣ “online social networks are contributing to the
isolation of people in the physical world”--Deresiewicz
‣ “online social networks have a positive impact on
social relations in the physical world”--Pew Internet
and American Life
• Distinction further blurred by ubiquity of location-
enabled smartphones
Tuesday, April 12, 2011 3

4. Motivation
• Difficult distinction of online and offline social networks
• Open ended debate:
‣ “online social networks are contributing to the
isolation of people in the physical world”--Deresiewicz
‣ “online social networks have a positive impact on
social relations in the physical world”--Pew Internet
and American Life
• Distinction further blurred by ubiquity of location-
enabled smartphones
Tuesday, April 12, 2011 3

5. Motivation
• Difficult distinction of online and offline social networks
• Open ended debate:
‣ “online social networks are contributing to the
isolation of people in the physical world”--Deresiewicz
‣ “online social networks have a positive impact on
social relations in the physical world”--Pew Internet
and American Life
• Distinction further blurred by ubiquity of location-
enabled smartphones
Tuesday, April 12, 2011 3

6. Challenges
• Infer properties of user social behaviors from
their location trails.
- Measure user similarity based on mobility to
infer user social structures [Eagle et al. (2009) and Li et al. (2008)]
- Co-location of two users insufficient to
determine their relationship, especially in
urban areas, where co-location among
strangers is frequent. [Miklas et al. (2007)]
• In reality, location tracking is inherently partial
and inexact.
Tuesday, April 12, 2011 4

7. Contribution
• Evaluate on two main tasks
- Predicting whether two co-located users are friends
on Facebook
- Predicting number of friends a user has
• Contributions:
1. Establish model of friendship by co-location
2. Find relationship between mobility pattern and
number of friends
3. Show diversity of location can be used to analyze
the context of social interactions
Tuesday, April 12, 2011 5

8. Related Work
• Statistical modeling of mobility patterns
- Examined features of mobility
- Tracked phone conversations
- Number of unique locations
- Proximity at work, Saturday night, etc.
- Self report of important factors
• Most work relied solely on co-location without
digging further
Tuesday, April 12, 2011 6

9. Methodology
Tuesday, April 12, 2011 7

10. Locaccino
• Web-application for Facebook
• Developed by Mobile Commerce Lab
at CMU
• Allows users to share location
‣ Facebook controlled privacy rules
• Contains two components
• Web Application
‣ Query friends’ locations
‣ Review Privacy rules
• Locator Software
‣ Updates user location http://locaccino.org/
‣ Run on laptops and mobile phones
‣ Update locations every 10 minutes
Tuesday, April 12, 2011 8

11. Locaccino
• Locator software uses combination of:
- GPS if applicable (Accurate ~10m-15m)
- WiFi lookup service (Accurate
~10m-20m)
- IP geolocation (city or neighborhood
level granularity)
• Sends time, latitude and longitude to
Locaccino
Tuesday, April 12, 2011 9

12. User Demographics
• 489 users of Locaccino
• Ranging from 7 days to several months (Average
74 days, median of 38 days)
• Use at different times and for different reasons
• Mostly from university campus
Tuesday, April 12, 2011 10

13. Data Collection
• 3 million location observations
- 2 million in Pittsburgh
- ignore IP geolocation
- 93.7% from laptop locator software
• Divide lat. and lon. into 30m x 30m grid
• Use 10 min. interval for time coordinate
• Co-location = same grid + same time
Tuesday, April 12, 2011 11

14. Network Data
• Social Network (S) –
Friends in Facebook
• Co-location Network
(C) – Co-located at
least once
• Co-located Friends
Network (S ∩ C) –
Friends and co-
located
Tuesday, April 12, 2011 12

15. Location Diversity Measurement
• Frequency – Raw count of observations
• User Count – Total unique visitors
• Entropy – Number of users and proportions of their observations
Tuesday, April 12, 2011 13

16. Co-location Features
• Intensity and Duration – Size and spatial and
temporal range. How long and how actively users
have embraced the system.
• Location Diversity – Frequency, user count and
entropy
• Specificity – How specific a location is to a given
co-location [TFIDF (l)]
u1,u2
• Structural Properties – Measures the strength of
the relationship between two co-located users
Tuesday, April 12, 2011 14

17. Other Measured Features
• Regularity of a user’s routine: {L, D, H}
• User mobility features
- Intensity and duration
- Location diversity: Location observations
of a single user
- Mobility regularity
Tuesday, April 12, 2011 15

18. Results
• 6 classifiers, 50-fold cross validation
• Performance:
• AdaBoost > Random Forest > SVM
• Overall accuracy of AdaBoost: 92%
- Guess better on non-friendship than friendship
Tuesday, April 12, 2011 16

19. y
n sit
e
t int
o u
ith
W l
e
od
m
Full
Number of co-
locations!
!
tu res
ea
yf
sit
en
Int
Tuesday, April 12, 2011 17

20. Inferring Number of Friends
• Look to relate number of Facebook friends
to mobility patterns
• Expectations:
- Users who have used the system longer
have more friends
- Users who visit “high diversity” locations
have more friends
- Users with irregular schedules may have
more friends (require help from Locaccino)
Tuesday, April 12, 2011 18

21. Pearson Correlation of User Mobility Features
• Worst: Intensity and duration
• Best: Location diversity
• MaxEntropy, MaxUserCount, MaxFreq best
Tuesday, April 12, 2011 19

22. Conclusions
• Co-location network 3x larger than social network (edge-wise)
- Social network better connected
• Properties of location are crucial
- Especially Entropy
- Difference between high and low entropy
- Help define both relationships and number of friends
• Created set of features to help classify social network friends
- Better than by simple co-location observations
• Found interesting patterns
- Co-location without friends
- Friends without co-location
Tuesday, April 12, 2011 20

23. Future Work
• Use classifiers for social network friend recommendation
system
• Augment and expand current friend-link system in place
• Could help provide insight into strength of relationship
• Still requires more research and validation
• Develop system for segregating and categorizing friends
• Help with privacy rules
• Build off relationship between online and offline social behavior
• Using things such as entropy of a location
• Use of location patterns of users
• Suggest similar locations to friends
• Suggest similar locations to non-friends with similar behavior
Tuesday, April 12, 2011 21

24. Appendix
Tuesday, April 12, 2011 22

25. Tuesday, April 12, 2011 23