19-02-2013, 04:47 PM
How do Facebookers use Friendlists
How do Facebookers use Friendlists.pdf (Size: 3.22 MB / Downloads: 84)
Abstract—
Facebook friendlists are used to classify friends into
groups and assist users in controlling access to their information.
In this paper, we study the effectiveness of Facebook friendlists
from two aspects: Friend Management and Policy Patterns by
examining how users build friendlists and to what extent they
use them in their policy templates. We have collected real
Facebook profile information and photo privacy policies of 222
participants, through their consent in our Facebook survey
application posted on Mechanical Turk. Our data analysis shows
that users’ customized friendlists are less frequently created and
have fewer overlaps as compared to Facebook created friendlists.
Also, users do not place all of their friends into lists. Moreover,
friends in more than one friendlists have higher values of node
betweenness and outgoing to incoming edge ratio values among
all the friends of a particular user. Last but not the least, friendlist
and user based exceptions are less frequently used in policies as
compared to allowing all friends, friends of friends and everyone
to view photos.
Index Terms—Policy, Access Control, Grouping, Privacy, Social
Network
I. INTRODUCTION
Facebook (FB) is the largest online social network used today
with over 900 million active users. The friendlist (FL) feature
was introduced in 2007, in order to help FB users in organizing
a large friend network into groups [2]. The FB privacy
controls allow users to use the FLs for customizing their
sharing preferences. Recently, FB improved the FL feature
by standardizing lists into the following three categories:
Close Friends: The user can put his top priority friends in
this list. Mostly, these are the friends with whom the user
interacts the most.
Acquaintances: The friends in this list are the ones that user
keeps with a mute button pressed. Their updates hardly appear
in the homepage news feed.
Smart Lists: These are lists that appear with lightening icon
and are automatically created and populated for each new
workplace, city or school that the user adds to his profile.
People can now use the FB standardized lists in addition to
their own customized lists. But, research shows that creating
FLs is a secondary task completed by only a few FB users.
An average FB user has about 120 friends [9], which makes
the process of friend grouping quite tedious. Most users are
either not aware of the FL feature or consider managing lists
a difficult task which requires remembering who is in which
list [12]. In a small thesis study, it was found that none of
the 10 study participants used FLs to control their privacy
settings [13]. Strater et al. [14] studied 18 FB users and found
that 72% of them either made their profile completely public or
restricted it to their friends only. Only 5 users used customized
settings to incorporate FLs. These studies lead to the following
questions about the effectiveness of the current FL feature on
FB:
Do users use the FL feature frequently?
How do users build FLs?
How are FLs used to compose policy patterns?
In this paper, we try to answer the above questions by
studying real user profile data and photo privacy policies.
We collected the policy data of 222 Facebookers through our
FB survey application distributed as Human Intelligent Task
(HIT) on Mechanical Turks. Using this data, we analyze the
effectiveness of FLs from two aspects: 1)Friend management,
by finding out the number and size of FB and user created
lists, percentage of users who do not fall in any lists, overlaps
between FLs and their comparison with various network
metrics and 2)Usage in policy patterns for setting exceptions.
The rest of the paper is organized as follows: In Section II,
we discuss the related work on studying FB FLs. Section
III explains how we collected the user data for our analysis.
Section IV details the various statistics and metrics that we
have used in our analysis. Section V presents and discusses the
results. Finally, we wrap up the paper with our conclusions.
II. RELATED WORK
Kelley et al. [7] have done preliminary work towards investigating
how users create friend groups in FB. They have
examined four different methods of friend grouping and their
results show that the type of mechanism used, affects the
groups created. Their study shows that 30% of the users had
FLs out of which 40% did not use them to control privacy
settings. Those who had FLs never updated them.
Recently, researchers have developed tools in order to assist a
user in grouping his/her friends efficiently and simultaneously
enabling him to create better privacy policies and introduce
exceptions. Adu-Oppong et al. [1] have proposed partitioning
a user’s friends into lists based on communities extracted
automatically from the network, as a way to simplify the
specification of privacy policies. Jones and O’Neill [6] created
PUBLICATION YEAR 2012
an automatic method to group FB users’ friends using the
SCAN clustering algorithm and compared these groups against
the user created groups, achieving 70% accuracy. Mazzia et
al. [11] built a policy visualization tool that extracts and
presents the user’s communities to help him in managing his
group based privacy policies. Egelman et al. [4] built a Venn
diagram based interface to cope up with semantic errors that
the current users make in their access control settings resulting
in over-sharing or under-sharing of information.
III. DATA COLLECTION
In order to collect real profile data and privacy policies, we
developed a FB survey application using the FB APIs. The
survey comprised of questions for gathering users’ privacy
concerns. The user consent was requested and users were
informed of the collected data. To recruit participants, we
published our survey application as an HIT on Amazon
Mechanical Turk. Amazon Mechanical Turk is a crowd sourcing
marketplace that pairs Requesters of work and Workers.
Requesters formulate work into HITs which are individual
tasks that workers complete. A total of 222 participants’ profile
and privacy policy data was collected. 173 out of these were
male and 49 were females. 57% of them had ages in the range
15-25, 37% in the range 25-35 and 6% in the range 35-70.
17.5% had grad school as their highest education. 10% had
college and 62.6% had high school as their highest education.
IV. DATA ANALYSIS
In this section, we describe the various metrics that were used
to analyze the collected data.
A. Friend Management
In rest of the paper, we refer to FB’s standardized list
categories i.e., acquaintances, close friends, work, education
and current city etc as FB created friendlists (FCFLs) and the
user’s customized FLs as User created friendlists (UCFLs).
For example, Alice has 200 friends. She creates three FLs with
names Family and relatives, High school friends and Undergrad
friends. After she updates her work with WellsFargo, FB
automatically creates a list with the name of this organization
and populates it. Therefore, the three FLs that she created
herself fall into UCFLs while WellsFargo falls into FCFLs.
To study how well users manage their friends through FLs,
we set out to extract the following statistics and determine
how users build their FLs:
1) FL creation statistics
Frequency w.r.t FL type: We calculate the average number
of FCFLs and UCFLs.
FL size: We measure the average list size.
Friend coverage: This metric specifies how many of the
user’s friends fall in at least one FL and is defined as
No: of friends in FLs
Total No: of friends .
2) Grouping strategy statistics
! Friends in more than one FLs: Sometimes, a user wants
some of his/her friends in one list to have access to additional
information as compared to the other members in this list. This
can be done by placing this friend in another list which has
access to this additional information, resulting in FL overlap.
FL overlap is therefore, an interesting metric to investigate the
number of roles a friend can have and to understand what type
of friends fall in more than one lists.
In order to find out specific features of these common friends,
we measure two network metrics:
! Node Betweenness: This metric measures how central a
node is to the network by calculating the number of shortest
paths in the network that contain this node [10].
! Node outgoing to incoming edge ratio: Node outgoing
to incoming edge ratio is another metric to analyze the
type of friends that belong to more than one FLs. This is
calculated with respect to the cluster in which a friend lies.
We use the Clauset Newman Moore (CNM) network clustering
algorithm [3] to cluster a network and then find out the ratio
between the number of connections of a node with nodes in
other clusters and the number of connections that this node
has with other nodes in its cluster.
Comparison of FLs with CNM clustering: We study
the correlation between UCFLs and FCFLs and the clusters
generated by CNM network clustering algorithm [3]. For this
purpose, we use the Adjusted Rand Index to measure the
alignment between CNM clusters and user defined FLs [5].
The Adjusted Rand Index compares the predicted labels (CNM
clusters) with the actual labels (user defined FLs) and produces
an index between 0 and 1. To cater for the FL label of friends
falling in more than one UCFL and FCFL, we picked a random
label for our comparison.
B. Policy Patterns
A policy pattern represents a combination of allow and deny
rules for access to information by friends. It can range from
being public to moderately private i.e., allowing all friends
or denying specific FLs to extremely private i.e., allowing or
denying specific users only. The collected privacy policies and
FL membership information was used to extract the policy
patterns. We divide these policy patterns into two categories :