Social networking websites have been increasingly popular in the recent years. The users create and maintain their social networks by themselves in these websites through establishing or removing the connections to friends and sites of interests. The smart phones not only create a high availability for social network applications, but also serve various ways of digital communication such as voice or video calls, e-mails and texts to form or maintain our social network. In this paper, we tackle the problem of automatically generating and organizing social networks by analyzing and assessing mobile phone usage and interaction data. We assign weights to the different types of interactions. The interactions among users are then evaluated based on these weight values for certain periods of time. We then use these values to rank the friends of users by a sports ranking algorithm, which recognizes the changes in the collected data over time.

1. Social Network Generation and Friend Ranking
Based on Mobile Phone Data
Social Network Generation and Friend Ranking
M. ˙I. Akba¸s, R. N. Avula, M. Bassiouni and D. Turgut
Department of Electrical Engineering and Computer Science
University of Central Florida - Orlando, FL
June 10, 2013
Akba¸s, Avula, Bassiouni, Turgut (UCF) ICC 2013 June 10, 2013 1 / 16

2. 1 Problem definition
Akba¸s, Avula, Bassiouni, Turgut (UCF) ICC 2013 June 10, 2013 2 / 16

3. 1 Problem definition
2 Application overview
Akba¸s, Avula, Bassiouni, Turgut (UCF) ICC 2013 June 10, 2013 2 / 16

4. 1 Problem definition
2 Application overview
3 Ranking and grouping friends
Akba¸s, Avula, Bassiouni, Turgut (UCF) ICC 2013 June 10, 2013 2 / 16

5. 1 Problem definition
2 Application overview
3 Ranking and grouping friends
4 Simulation study
Akba¸s, Avula, Bassiouni, Turgut (UCF) ICC 2013 June 10, 2013 2 / 16

6. 1 Problem definition
2 Application overview
3 Ranking and grouping friends
4 Simulation study
5 Conclusion
Akba¸s, Avula, Bassiouni, Turgut (UCF) ICC 2013 June 10, 2013 2 / 16

7. Problem definition
Problem
Online social networks are successful in providing digital social
networks of friends.
The explicit interactions in virtual social network give users the
impression of full control over their ego network.
Time-consuming, not accurate and the users are confronted with the
need to organize.
Objective
Automatically generating and organizing social networks by analyzing
and assessing mobile phone usage and interaction data.
Akba¸s, Avula, Bassiouni, Turgut (UCF) ICC 2013 June 10, 2013 3 / 16

8. Data available on smart phones
Various smart phone applications allow us to trace multiple types user
interactions:
Text messages, call logs, e-mail, chat messages.
Smartphones are equipped with many sensors recording additional
information
Acoustic sensors, Bluetooth, GPS receivers.
Usage of all available data enables the interpretation of ego network.
Akba¸s, Avula, Bassiouni, Turgut (UCF) ICC 2013 June 10, 2013 4 / 16

9. Ego networks
Dunbar et al.† showed:
People can have a limited number of friends.
Social networks are hierarchically organized in discretely sized groups.
†R. Dunbar, “How Many Friends Does One Person Need?: Dunbars Number and Other Evolutionary Quirks” Faber and Faber,
Feb. 2010.
Akba¸s, Avula, Bassiouni, Turgut (UCF) ICC 2013 June 10, 2013 5 / 16

10. System dynamics
Weights assigned to different types of interactions.
Interactions among users are evaluated based on weight values for
certain periods of time.
Weight values over time are used to rank friends of users by a sports
ranking algorithm.
Ranking algorithm recognizes the changes in the collected data over
time.
Akba¸s, Avula, Bassiouni, Turgut (UCF) ICC 2013 June 10, 2013 6 / 16

11. Overview of the system
Overview of social network generation for users.
Akba¸s, Avula, Bassiouni, Turgut (UCF) ICC 2013 June 10, 2013 7 / 16

12. Nodobo dataset
An experiment from the University of Strathclyde to gather
communications metadata from a group of high school students.
Our approach is applied to the dataset of Nodobo project.
Android phones with a special OS is used to collect interaction
information.
Our approach concentrates not on the social network among the users
but on formation and organization of these personal networks.
Akba¸s, Avula, Bassiouni, Turgut (UCF) ICC 2013 June 10, 2013 8 / 16

13. Interaction evaluation
The interaction data derived from mobile devices and sensors need to
be evaluated in terms of their importance.
Interaction data are analyzed and assigned with weights according to
interaction information:
Identifications of the parties in the communication.
Time of the communication.
Type of the communication method.
Weight assignment results in the ability to change the priority
according to experience or context.
iA,B = α · F(T) + β · V (T) + γ · nP(T) + δ · E(S)
Akba¸s, Avula, Bassiouni, Turgut (UCF) ICC 2013 June 10, 2013 9 / 16

14. Ranking friends
Ranking is used to find the discrete groups of friends in ego network.
The interactions between friends correlate in number with the
strength of friendship.
We define a sports competition type relationship among users.
Interaction values are used to determine wins or losses for defined
periods of time.
The friend with a larger interaction value in a defined period of time
has a win against the friend with lower interaction value.
Akba¸s, Avula, Bassiouni, Turgut (UCF) ICC 2013 June 10, 2013 10 / 16

15. Selection of ranking method
The most common sports ranking method utilizes the winning
percentage, which is based on the number of wins and number of
comparisons.
Popular methods using schedule and current rankings: Colley and
Massey.
Colley and Massey methods are insensitive to small changes†.
Insensitivity is a desirable property for ranking.
Colley method is chosen: Colley ranking is based only on the results
from the comparisons while Massey utilizes actual game scores and
homefield advantage.
†T. P. Chartier, E. Kreutzer, A. N. Langville, and K. E. Pedings, “Sensitivity and stability of ranking vectors,” SIAM J. on
Scientific Computing, vol. 33, pp. 10771102, May 2011.
Akba¸s, Avula, Bassiouni, Turgut (UCF) ICC 2013 June 10, 2013 11 / 16

16. Colley ranking
Colley ranking can be defined by a linear system: Cr = b
b, is defined as follows: bi = 1 + (wi − li )/2
Cn×n is called the Colley coefficient matrix and defined as follows:
Cij =
2 + (wi + li ) i = j
−nij i = j
where nij : number of times friends i and j are compared.
Then the rating of a friend of a user is defined as follows:
ri =
1 + (wi −li )
2 +
k∈Fu
rk
2 + (wi + li )
Akba¸s, Avula, Bassiouni, Turgut (UCF) ICC 2013 June 10, 2013 12 / 16

17. Ranking friends of Nodobo users
Ratings of friends with Colley method for nine users:
Akba¸s, Avula, Bassiouni, Turgut (UCF) ICC 2013 June 10, 2013 13 / 16

18. Colley vs. Winning Percentage
Friend ratings for two users:
Akba¸s, Avula, Bassiouni, Turgut (UCF) ICC 2013 June 10, 2013 14 / 16

19. Change in rankings
Change in ratings of a user’s friends is for each month during the four
month interval:
Akba¸s, Avula, Bassiouni, Turgut (UCF) ICC 2013 June 10, 2013 15 / 16

20. Conclusion
Outlined an approach to infer ego networks by assessing smartphone
data.
Proposed an interaction evaluation function, a ranking and grouping
method.
Possible future work:
Using multiple real life smartphone data with collection times spread
over longer periods.
Implementation of a smartphone application, collecting and integrating
all available interaction data to extend our aproach.
Akba¸s, Avula, Bassiouni, Turgut (UCF) ICC 2013 June 10, 2013 16 / 16