Understanding User Behavior at Scale in a Mobile Video Chat Application

Understanding User Behavior at Scale in a Mobile Video
Chat Application

Online video chat services such as Chatroulette and Omegle
randomly match users in video chat sessions and have become
increasingly popular, with tens of thousands of users

online at anytime during a day. Our interest is in examining
user behavior in the growing domain of mobile video, and
in particular how users behave in such video chat services
as they are extended onto mobile clients. To date, over four
thousand people have downloaded and used our Androidbased
mobile client, which was developed to be compatible
with an existing video chat service. The paper provides a
first-ever detailed large scale study of mobile user behavior
in a random video chat service over a three week period.
This study identifies major characteristics such as mobile
user session durations, time of use, demographic distribution
and the large number of brief sessions that users click
through to find good matches. Through content analysis of
video and audio, as well as analysis of texting and clicking
behavior, we discover key correlations among these characteristics,
e.g., normal mobile users are highly correlated
with using the front camera and with the presence of a face,
whereas misbehaving mobile users have a high negative correlation
with the presence of a face.
AUTHOR KEYWORDS
Mobile video chat; behavior analysis; effective matching;
misbehavior detection
ACM CLASSIFICATION KEYWORDS
H.5.1 [Information Interfaces and Presentation]: Multimedia
Information Systems: Video, Evaluation/methodology
INTRODUCTION
Realtime, interactive video-based services are fast becoming
an integral part of the Internet user experience. Video chat
services from providers such as Skype, Google+ and Facebook
are now commonplace. Jump started by Apple’s FacePermission
to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than
ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
UbiComp’13, September 8–12, 2013, Zurich, Switzerland.
Copyright
c 2013 ACM 978-1-4503-1770-2/13/09...$15.00.
http://dx.doi.org/10.1145/2493432.2493488
time, these services are even rapidly becoming mainstream
for mobile devices such as phones and tablets as well. A
common theme is that these video chat services link users
who have previously established friend relationships.
On the other hand, random video chat services such as Chatroulette
[5], Omegle[18] and MeetMe [16] have recently
started to gain in popularity. In random video chat, strangers
are randomly paired together for video-based conversations.
The appeal of these services for users is a classic one: the
ability to meet new people face to face, now shifted to the
virtual domain. These services have become extremely popular
over the last few years. For example, both Chatroulette
and Omegle have tens of thousands of users at any given time
actively using their systems. As random video chat continues
to gain in popularity, we expect mobile random video
chat to increase in volume and frequency as well.
While standard and even random video chat services have
been studied extensively (see, e.g., [11, 1, 24, 6, 25]), prior
work has not shed light on the behavior of users in mobile
random video chat services. Introducing the new paradigm
of mobile interaction into what had formerly been primarily
a desktop-based interaction paradigm with webcam-driven
online random video chat raises new questions: How are
mobile users using random video chat? Is physical mobility
fundamentally altering the nature of interacting with a
virtual acquaintance?
There are two important motivations for understanding the
behavior of mobile users in random video chat. First, from
a social science perspective, better insight into why users
engage in these services can aid in designs that improve
user experience. For example, matching users quickly with
other users with whom they would prefer to chat would lead
to higher user satisfaction and presumably extended session
durations. An understanding of mobile user behavior can
inform us as to salient characteristics that would lead to better
matching beyond purely random pairings. Second, safeguarding
regular users from misbehaving users is equally
important. Standard (non-mobile) random video chat services
are especially prone to objectionable content such as
unauthorized advertisements and sexually explicit flashers
who expose all or parts of their bodies. Prior work has shown
that intelligent filtering that leverages user behavior patterns
can help to safeguard regular users from such content [24, 6,25]. An understanding of misbehavior in the mobile setting
can help us extend prior protections to the mobile domain.
However, studying mobile user behavior in random video
chat services is challenging due to two compounding factors.
First, the very nature of mobility means that users have frequent,
untethered and low-overhead opportunities to interact
with the video chat service. Therefore, outside observation
of active mobile users is a myopic undertaking. Second, as
we will show in detail, users are very selective about with
whom they interact. Small scale user studies are thus illsuited
to capturing the aggregate trends of random user-touser
interactions, and the formation of new social bonds.
In this paper, we address these important challenges by conducting
a large-scale study of mobile-client user behavior
in online video chat services in the real world. We have
developed MVChat, an Android-based mobile client compliant
with the popular Omegle random video chat service
(Figure 1). The first version was released in November of
2012, and has been used by more than four thousand users
so far. In this paper, we analyzed in detail three weeks (January
25th - February 14th) of user behavior on the mobile
video chat client.1 Our key findings are that: the majority
of sessions are brief, as users search for a more meaningful
partner to converse with for a longer duration session; normal
mobile users are highly correlated with using the front
camera and with the presence of a face, whereas misbehaving
mobile users have a high negative correlation with the
presence of a face; females are highly popular, in that users
with a large enough fraction of sustained sessions are disproportionately
female, but surprisingly gender is not correlated
with misbehavior, i.e., females were just as likely to misbehave
as males; mobile content is more diverse than Web
online content; and groups typically imply normal behavior.
To our knowledge, this is the first-ever study of mobile video
chat behavior at scale. In the following, we describe related
work, our methodology, the system used for data collection,
and our key findings in more detail.
Figure 1. A screenshot of the MVChat application.
RELATED WORK
Prior work by Jana et al. [14] has discussed the issues and
challenges in designing successful mobile video chat applications.
Scholl et al. [20] designed and built an online
1This study has been approved by our institution’s Human Research
& the Institutional Review Board.
video chat application, reporting results from 53 users in a
social setting, focusing on bandwidth issues and view navigation.
VideoPal explored the use of video to facilitate asynchronous
communication between six children and their close
friends [11]. Other work focuses on understanding video
chat system usage between teenagers [4, 21] and in the context
of families, especially the facilitation of communication
between grandparents and grandchildren [1, 19, 17]. Our
work differs from these prior works in one or more of the
following respects: its larger scale; its publicly available deployment;
its focus on the mobile context; and the study of
random video chat.
Prior research has described the design of classifiers to detect
and remove misbehaving users from online random video
chat sessions [24, 6, 25]. These papers do not explore the
mobile context nor do they provide a detailed understanding
of typical and misbehaving user behavior and their correlation
factors.
In mobile video research, MoVi explored the use of collaborative
sensing on mobile phones to trigger the video recording
of a social event by one of the participants’ camera phone,
as well as the generation of video highlights of the event [2].
MicroCast sought to share video streaming amongst a local
group of smartphones, who also share their partial results
with one another [15]. Mobile video encoding for wireless
links has recently introduced cross-layer encoding (SoftCast)
[13], and reliable coding (ChitChat) [23] techniques.
A study measuring the energy and bandwidth costs of streaming
video from popular websites such as YouTube to six different
smartphones has been conducted [10]. More generally,
there have been a number of studies examining user
behaviors and usage of mobile phones [9, 7, 22, 8, 12, 3].
METHODOLOGY
In seeking to understand user behavior of random mobile
video chat at scale, we are interested in answering a variety
of questions, starting with demographic questions. Who
uses these kinds of services (country of origin, male/female,
single/group, etc.)? How well equipped is the software and
hardware of smartphones for such users? And which users
participate prominently in these services?
Behavioral questions are also of great interest. What is the
length of a typical random video chat session? How often
do users seek new pairing sessions while using the application?
What (day/night/hour) are the most popular times for
random video chat? How often do mobile users terminate
their sessions compared to the other side terminating the session?
What role does texting play during a typical mobile
video chat session? What fraction of mobile users behave
in an unsafe manner, e.g., flash or reveal themselves? And
what are the key differences in user behaviors between mobile
users and other online video chat users? Finally, from
a contextual standpoint, to what extent can mobile sensor
data help us understand the overall environment, e.g., background
sound, location (indoor/outdoor), motion, etc. for
mobile video chat?Once we have the basic taxonomic classifications, we ask
whether there are any strong correlations (positive or negative)
between various characteristics, such as gender, presence/absence
of a face, presence/absence of a person, front/
back camera usage, normal behavior/misbehavior, and audio
silence/voice/music? Is there any correlation between acceleration
data and user behavior?
To answer these questions, we have developed an Androidbased
mobile client (called MVChat) for random online video
chat services through which we can collect user data. This
client allows mobile users to connect with online Omegle
users for random video chat. In addition, MVChat logs on
our server multi-dimensional, user-related sensor data from
each participating mobile client. This data includes image
snapshots, audio, camera position (front or back), accelerometer
and gyroscope data, texting data, nexting clicks, and
times when new chat sessions begin and end. This data is
analyzed and classified on an image, session, and user basis.
Image-based analysis examines user behavior based on individual
snapshot images and the corresponding sensor data,
e.g., audio and accelerometer, nearest in time to when that
snapshot was taken. Session-based analysis examines user
behavior across all data collected during a single chat session,
which contains multiple image snapshots. User-based
analysis groups together all data collected during all sessions
associated with a single user for behavioral analysis.
Images were categorized depending on whether the snapshot
contained one person, more than one person, or no person
at all. For images that contain at least one person, we further
subdivide them into several different categories: images
that contain a face or no face; images that contain a male,
female, or mixed; images that contain a normal user or misbehaving
user, where a misbehaving user is defined as one
that flashes. For audio analysis, we listened to the audio
samples and classified them as containing either silence or
background noise, human voice, or music & sound. For text
analysis, we classified them as either containing some text
or not containing any text. Camera position was classified
as either front position or back position. Finally, each session
was classified as either terminated by the mobile user
by clicking the Next button or terminated by the other side.
SYSTEM FOR DATA COLLECTION
In order to understand user behavior at scale, we designed
our MVChat system to collect data on the scale of thousands
of mobile video chat users, millions of video chat sessions,
and gigabytes of image snapshots, audio, and smartphone
sensor data. Building a mobile client application that
is compatible with an existing online video chat service allows
us to quickly scale our study to a large number of users
of our application. We chose to make MVChat compatible
with Omegle, which is a popular video chat service that
has tens of thousands of randomly paired users at any given
time and is fairly similar to Chatroulette. Unlike Chatroulette,
Omegle’s user population is unfiltered, which allows
MVChat to gain insight about misbehaving users and
thereby capture a more representative sample of the true proportions
and behaviors of mobile video chat users.
We chose the Android mobile development platform because
the application approval process for iPhone essentially precludes
any applications such as random video chat that have
some indecent content or misbehaving users. Our system
consists of three major components: an Android mobile
video chat client that is Omegle compliant, a data collection
server to store multi-dimensional mobile user data, and the
Omegle server. Figure 2 illustrates the overall architecture
of the MVChat system.
Adobe Stratus
Server
Omegle
Server
MVChat Data
Collection Server
1. Stratus Registration
2. App Behavior Control
3. Video Session Establishment/
 Chatting (RTMFP + P2P)
4. Data Collection
Acc Gyro Audio Image
Camera Position
2
2
2
1
1
1
3 3
4
4
Text Next
MVChat client
MVChat client
Figure 2. System architecture of the MVChat system.
Compatibility of the MVChat clients with the Omegle system
required careful engineering to mimic the behavior of
online Web clients. To start an Omegle video chat session, a
typical Web browser/client contacts the Adobe stratus server
to register and receive a unique peer ID. Next, that client establishes
a connection with the Omegle server by providing
it the peer ID. The connections between the Omegle server
and all the Web clients are used for establishing random pairing
of online Omegle clients for chatting. Each time a client
requests to chat with someone, the Omegle server replies
with the peer ID of a randomly chosen client that is currently
idle. The Omegle server also notifies the randomly
chosen client with the requesting client’s peer ID. The two
clients then proceed to establish a peer-to-peer video session
between them using each other’s peer IDs and the Real Time
Media Flow Protocol (RTMFP).
After the video session has been established, each MVChat
mobile client establishes a separate connection to our data
collection server and periodically posts to it the user’s image,
audio, camera position, accelerometer and gyroscope data as
well as any text data. When a client wants to end its video
chat session, it notifies the Omegle server and at the same
time stops transmitting its video stream, releases resources,
and stops posting data to the data collection server.
Mobile Client Application
Figure 1 shows a screenshot of the MVChat mobile client.
Unlike the Omegle Web client, where text messaging dominates
the screen and the videos of the sender and receiver are
shown in less than half the screen, our MVChat mobile clientemphasized video first, as there was limited screen real estate
and we felt most mobile users would interact most easily
with video and audio. After invocation and connection with
a remote user, the mobile client displays the remote client’s
video (captured via remote user’s device camera) in the large
window on the right. The local camera view that is being
sent to the remote user is shown in the lower left window.
However, to maintain compatibility with the user experience
of Omegle chatters, we included text messaging. Text messages
are composed in the upper left dialog box and the most
recent messages are shown in the middle left box.
To maintain compatibility with Omegle’s session flow, where
one click ends a session and a second click requests a new
session, we added a Disconnect button that is displayed at
the top when a video chat session is in progress. A user
may press this button to end the current video chat session,
which changes the button to a Next button, and then a user
may press the Next button to request a new client to chat
with. Unlike Web clients, mobile devices have both front
and back cameras, so we added a pull-down menu that allows
the user to specify whether they wish to use their front
or back camera for a chat, and this can be changed anytime
during a video chat session. The other pull-down menu allows
the user to specify the microphone that they wish to
use, should there be more than one supported.
Data Collection Server
The data collection server is comprised of three components:
an Apache server, a MySQL server and a Flash Interactive
Media Server. The Apache server stores image snapshots
directly on our server file system and cooperates with the
MySQL server to store most of the rest of the mobile sensor
data. The Flash Interactive Media Server is responsible for
storing the audio file and for easy binding with the mobile
client via the RTMFP protocol.
For each mobile device, a random device ID (not the Adobe
Stratus peer ID) is generated, and for each video session,
the device generates a session ID. Every set of data posted
from devices to the data collection server is tagged with the
captured time stamp, camera position as well as the device
and session IDs for easy segmentation.
The default sampling frequency for accelerometer and gyroscope
is 5Hz. Snapshots are captured every 30 seconds and
audio is captured in the first 10 seconds of every 40-second
interval. On average, each image is about 35 KB (120*160)
and each audio file (10 second duration) is around 110 KB.
So the total amount of data transmitted from mobile client
to the data collection server is less than 5 KB per second
per user. This aggregate sampling rate in theory allows us
to scale to thousands of concurrent users, though in practice
we have thus far seen a maximum of 80 concurrent users.
In order to accommodate unforeseen workloads and help the
system scale, we designed the data collection server to adjust
flexibly or even turn off the sensor reporting streams
from the mobile clients. For each mobile client, the server
can vary the sampling frequencies, or even stop data collecFigure
3. (a) User country distribution and (b) installation device OS
distribution.
tion altogether. This ability allows us to throttle clients if
the overall load on the data collection server is too high, and
also to suspend data collection while allowing the clients to
continue to video chat normally. When the client starts, it
first contacts the data collection server, which responds with
one of three options: disabling the client completely for this
session so that it cannot access the Omegle service at all; disabling
just the posting of sensor data to the collection server,
whereupon the client can access the Omegle service without
any remote logging; or enabling the posting of sensor data as
the user accesses the Omegle service. In the third case, the
server also (optionally) specifies the sampling frequencies
for each sensor modality.
Data Collection Experiment
We deployed our MVChat mobile application on the Google
Play marketplace for Android applications and collected data
spanning about 3 weeks from January 25th to February 14th,
2013. In total, 4,632 distinct users of our MVChat application
were identified, generating 1,703,837 pairing sessions.
To protect remote users’ privacy, data was only collected for
our mobile users, and not for the remote users paired with
them. The total amount of data collected was about 170 GB,
comprised of 70 GB of image snapshots, 90 GB of audio
snippets, 8.5 GB of mobile sensor, texting, and clicking data.
Figure 3(a) shows the distribution of users by country of origin
who have downloaded and installed our application. We
see that the vast majority are from the United States and that
all other countries comprise at most single digit percentages
of the downloads. As shown in Figure 3(b), many users of
our application still rely on older versions of Android 2.3,
but that a fairly large fraction have Android 4.0 or higher.
We also found that about three-fourths of our users’ smartphones
were equipped with both front and back cameras, but
that a fourth still lacked the front camera.
USER BEHAVIOR ANALYSIS
We conducted a comprehensive analysis of our data to understand
and identify key behavioral characteristics of mobile
video chat users at scale. Our data analysis consists
of five specific components: (1) Global data analysis studies
the overall statistical distributions and identifies common
user behaviors. (2) Mobile-online video chat comparison
highlights the differences between mobile and online video
chat, and new features associated with mobile video chat.
(3) Taxonomy correlation analysis identifies user behavioral
attributes that have strong positive or negative correlationsFigure 4. CDF of session duration: 99.5% of the 1.7 million
video chat sessions were shorter than 60 seconds.
Figure 5. CDF of number of sessions per user: 42% of the
4,632 users had more than 100 video chat sessions.
Figure 6. Time of use: (L) day of week and (R) hour of day.
and directional association rules with strong confidence. (4)
Meaningful user analysis aims to characterize users who are
popular and are more effective in maintaining longer video
chat sessions. (5) Accelerometer data analysis studies how
accelerometer sensor data can help identify certain user behavioral
characteristics in mobile video chat.
Global Data Analysis
We first conduct a global data analysis to understand key
characteristics of user behavior in mobile video chat, including
session duration, time of use, local stop behavior, use of
text messaging, and a taxonomy distribution.
Session duration refers to the length of users’ video chat
sessions. Figure 4 shows the cumulative distribution function
(CDF) for session duration. Among the 1.7 million
video chat sessions we have collected, 80% of the sessions
were less than 5 seconds; only 1% of the sessions were longer
than 30 seconds; and only 0.5% of the sessions were longer
than 60 seconds. Figure 5 shows the CDF of number of
sessions for each user. Among the 4,632 users we have
observed, 80% of the users participated in at least 10 sessions,
42% of the users participated in more than 100 video
chat sessions; and 6% of the users participated in over 1,000
video chat sessions during our 3-week data collection period.
A hypothesis consistent with these findings is that video chat
users spend a lot of effort going through many random pairing
sessions in order to find someone interesting to chat with
for a longer duration. This suggests that random pairing is
ineffective in the sense of generating a lot of “noise” until the
“right” pairing shows up in a video chat screen. While some
video chat services have recently introduced interest-based
pairing, it is not clear if interest is the only and most effective
matching metric. To better understand this problem, we
study “meaningful sessions” and “meaningful users” later in
this section.
Time of use refers to the time when users participate in
video chat sessions. We consider both day of week and hour
of day using the local time reported by users’ smartphones.
As shown in Figure 6(L), the average number of users is
Figure 7. PDF of local stop probability: Mobile users are more likely
to end video chat sessions locally.
similar throughout a week except for Fridays and Saturdays.
This is interesting as we expect that more users have free
time on Fridays and Saturdays. One possible explanation
is that, because more users are free on Fridays and Saturdays,
they go out and participate in other social events (i.e.,
“party time”) instead of using our mobile video chat application.
Figure 6(R) shows the number of users during different
hours of the day. We can see that early morning (5am) has
the fewest number of users, and the average number of users
increases steadily throughout the day, with quick jumps at
3pm and 10pm, and reaches the highest number around midnight.
The fact that we observe more users during the late
evening hours is potentially related to the private nature of
mobile video chat, i.e., when users are by themselves and
would like to meet strangers in the virtual world.
Local stop probability refers to the probability of a user
ending his/her session locally instead of the remote party
ending the session. Note that each user can click the “Disconnect”
and “Next” buttons at any time to end the current
video chat session and start a new pairing session. If we
assume that for each video chat session, the two parties (local
and remote) have equal probability of ending the session,
then across all users and their sessions, we would expect to
see a distribution for local stop probability with a mean value
of 0.5. However, the distribution shown in Figure 7 is different,
with a mean value close to 0.6. Note that we only
collected data from mobile video chat users and not the online
video chat users at Omegle with whom our mobile users
chat. The higher local stop probability of our mobile users
means that mobile users are more likely to end a video chat
session than the online users. We think such deviation is possibly
due to the difference in network connection between
mobile users and online users. Video chat is relatively expensive
and requires good network connection. Our mobile
users are typically connected through 3G or WiFi networks,
while the online users usually have better network connection
with higher bandwidth and more stability. As a result,
mobile users are more likely to end a session if there is delay
in starting the session or the quality is low.Table 1. Taxonomy Distribution of Meaningful Session Images
Taxonomy Distribution
Person No Person Person
594 1519
Group No Person Single Group
594 1446 73
Normal No Person Misbehaving Normal
594 314 1205
Face No Person No Face Face
594 425 1094
Gender∗ No Person Male Female Both
594 688 721 3
Camera Back Front
Position 653 1460
Audio No Data Silence/Noise Voice Music
687 735 594 94
∗There are 107 images for which we could not determine the gender.
Text messages can also be used in users’ video chat sessions.
Although 58.3% of the users had used text messages during
video chat, only 2.8% of all video chat sessions contained
text messages. We think this is due to the fact that most
of the sessions were short sessions, when users are quickly
clicking through many random pairing sessions in order to
find the right person to talk to. These short sessions generally
contained no text.
GPS data collection is a function incorporated in our application.
However, the data we collected contain almost
no GPS data. This, together with our examination of the
snapshot images, indicates that almost all video chat sessions
occurred indoors. This is reasonable as video chatting
with strangers is considered a private activity and people
prefer to participate in such activities in private indoor
environments. Moreover, based on our observation of the
snapshot images, many of our mobile video chat users are
young people and tend to use the application in their homes
or dorms. The indoor locations vary from living room to
bedroom and even bathroom. Also, mobile user posture is
distributed across sitting, lying down, and standing, and appears
to be more diverse than online Webcam-based images
captured from desktop clients [24], where users typically are
located in the bedroom and are sitting.
Taxonomy analysis aims to characterize key user behaviors
when using the mobile video chat application. Due to the
large scale of the data we have collected (4,632 users and
1.7 million sessions), it is infeasible to label all the data. In
addition, we have observed earlier that the majority of the
sessions were short sessions. Therefore, we decided to focus
our taxonomy analysis on meaningful sessions sampled
from the overall data set, i.e., sessions that lasted 60 seconds
or longer. Our reasons for focusing on meaningful sessions
are three-fold: (1) We want to understand what user behavioral
characteristics promote longer and potentially more effective
video conversations; (2) Sessions that last at least 60
seconds contain at least 3 snapshot images and 2 audio samples,
which provide adequate information to label each session;
and (3) There are almost 8,000 meaningful sessions in
our data set, which are sufficient for our analysis. Among all
the meaningful sessions, we randomly sampled 1/40 of the
sessions. After removing noisy sessions whose snapshots are
black and have no content, we have a set of 218 meaningful
Table 2. Taxonomy Distribution of Meaningful Sessions
Taxonomy Distribution
Person No Person Person
62 156
Group No Person Single Group
62 147 9
Normal No Person Misbehaving Normal
62 57 99
Face No Person No Face Face
62 50 106
Gender No Person Male Female Both
62 94 50 1
Text No Text Text
90 128
Stop No Data Local Stop Remote Stop
49 79 90
Audio No Data Silence&Noise Voice Music
4 122 79 13
sessions with 2,113 images in total. Using the taxonomy we
have defined, we then manually label each individual image.
Since each meaningful session contains multiple snapshot
images and the variance of these images’ labels is generally
small, we use majority voting of image-based labels to determine
the taxonomy labels for each session. There are two
exceptions. First, camera position has a higher variance in
sessions, so we ignore them in session-based analysis. Second,
to label normal and misbehaving sessions – if any image
in a session is labeled as misbehaving, that the whole
session is labeled as misbehaving. For each session, we also
considered text and local vs. remote stop information.
The taxonomy distributions of meaningful session images
and meaningful sessions are summarized in Table 1 and Table
2, respectively. We can see that most meaningful sessions
and their images contained a person, single user, normal
user, user face, silence/noise or voice. Also, the front
camera is used much more often than the back camera. It is
interesting to note that among the meaningful sessions, there
are fewer female sessions than male sessions, but more female
images than male images. The reason is that female
users are possibly more popular and tend to have longer
video chat sessions, thus each session contains more images.
Mobile-Online Video Chat Comparison
We compared both mobile and online user behavior in the
Omegle application, with the caveat that the information we
obtained from online users was less extensive than our mobile
data collection. This is because we relied on a randomized
image data set provided to us by Omegle, from which
we could derive such properties as gender proportion, but not
session duration for example. While we could have built our
own instrumented Web application compliant with Omegle
to measure these factors, we felt the adoption rate would not
have been strong since Web browsers already connect with
Omegle. In contrast, our mobile application introduced new
capabilities - namely mobile video chat - that drove adoption.
Figure 8 compares the gender distributions of mobile
and online Omegle users. As shown in the figure, mobile
users were about 49% male and 12% female, while online
users were 68% male and 17% female. Note mobile users
had a much higher fraction of other types of content. Due to
the portable feature of mobile video chat, users can engageFigure 8. Comparison of user gender distribution: (L) mobile video
chat and (R) online video chat. “Others” refer to other types of content.
Figure 10. Comparison of time of use (day of week and hour of day):
(L) mobile video chat and (R) online video chat.
in video chat sessions at varying locations and with different
posture. Many smartphones are also equipped with both
front and back cameras, allowing users to switch between the
two cameras and show different content during video chat
sessions. As a result, with mobile video chat, we expect to
see more diverse other types of content than that of online
video chat, such as the sample images shown in Figure 9.
We also compare the time of use, i.e., day of week and hour
of day, between mobile and online Web users. The results
are shown in Figure 10. Omegle is a popular online video
chat service, and its number of concurrent users vary between
15,000 and 45,000. For our application, the number
of users at any given hour of day and given day of week
varies between 15 and 80. The general trend for hour of day
is similar for both mobile and online users. Overall, online
users are more uniformly distributed across different time,
while mobile usage sees more fluctuation (partly due to the
smaller number of mobile users) and higher usage in the afternoon
and late evening hours. One significant difference
is on day of week: online usage is much higher on Friday
evenings and most of Saturdays, while mobile usage actually
sees lower activity during these two days. This may be
due to different user populations for the mobile and online
worlds, e.g., mobile users are more likely to be out partying
on Fridays and Saturdays.
Taxonomy Correlation Analysis
Our taxonomy aims to characterize users’ mobile video chat
behavior from multiple dimensions. Given the labeled taxonomy
information, one important question we want to answer
is which user characteristics are correlated. In other
words, we want to identify behavioral characteristics which
are likely or unlikely to occur together. For instance, do
male and female users behave differently, or how do normal
users behave compared to misbehaving users. Understanding
such taxonomy correlations can offer useful insights into
designing better user pairing strategies, misbehavior detection
mechanisms, etc.
Since our taxonomy contains categorical rather than numerical
attribute values, we utilize four correlation metrics that
are typically used for categorical correlation analysis: χ
2
,
lif t, all conf idence (or all conf), and cosine. Let A and
B be two attributes (e.g., gender and camera position) with
values ai(1 ≤ i ≤ c) and bj (1 ≤ j ≤ r) respectively, these
four metrics are defined as follows:
χ
2 =
Xc
i=1
Xr
j=1
(nij − eij )
2
eij
(1)
eij =
count(A = ai) ∗ count(B = bj )
N
(2)
liftij = nij/eij (3)
all confij =
nij
max{count(A = ai), count(B = bj )}
(4)
cosineij = p
nij
count(A = ai) ∗ count(B = bj )
(5)
Here N is the total number of samples, nij is the number of
samples with both A = ai and B = bj , count(A = ai) and
count(B = bj ) are the numbers of samples with A = ai
and B = bj respectively.
χ
2 measures the difference between observed values nij and
expected values eij (if A and B are not correlated). So a
small χ
2 value (close to 0) means non-correlation while a
high χ
2 value indicates possible correlation. Similarly, a
lif t value of 1 means no correlation (nij = eij ), and a lif t
value > 1 (or < 1) indicates positive (or negative) correlation.
However, both χ
2
and lif t are sensitive to skewed
distribution of the attribute values (e.g., most user are normal
or very few users chat as a group). By focusing on ai
and bj values and ignoring other values (i.e., null-invariant),
all conf and cosine can tolerate different data set scales
and skewed attribute value distributions. Generally, all conf
and cosine values that are close to 1 indicate strong positive
correlation and values that are close to 0 indicate strong
negative correlation. In our analysis, we leverage χ
2
and
lif t to confirm non-correlation (i.e., χ
2
is close to 0 and
lif t is close to 1), and leverage all conf and cosine to
identify strong positive correlation (i.e., both all conf and
cosine are > 0.85) and strong negative correlation (i.e., both
all conf and cosine are < 0.1).
Note that correlation is bi-directional: if ai and bj are positively
(negatively) correlated, seeing ai means that bj is
more (less) likely to occur, and vice versa. However, some
relations between user behaviors can be unidirectional, e.g.,
if ai occurs, bj is likely to occur but the reverse may not be
true. We use association rules (e.g., ai ⇒ bj ) to capture
such one directional relations, and the conditional probability
P r(bj |ai) is referred to as the confidence of an association
rule.
Figure 11 summarizes the key results of our taxonomy correlation
analysis, including strong positive (negative) correla-Content in bathroom Quilt Trademark Partial legs
Figure 9. Sample images collected via our app demonstrating more diverse image content in mobile video chat than that in online video chat.
No
Person Person Group Single
Normal
Silence/
Noise Voice Music
Misbehaving
No Face
Face
Back
Camera
Front
Camera
Male
Female
Both
1. Strong Positive
 Correlation
2. Strong Negative
 Correlation
3. Association Rule
Local
Stop
Remote
Stop
No
Person Person Group Single
Normal
Misbehaving
No Face
Face
Male
Female
Both
No
Text Text
1. Strong Positive
 Correlation
2. Strong Negative
 Correlation
3. Association Rule
Silence
/Noise Voice Music
Figure 11. Taxonomy correlation analysis: (L) image-based graph and (R) session-based graph. Highlighted in the graphs are strong positive
correlations (both all conf and cosine are > 0.85, strong negative correlations (both all conf and cosine are < 0.1), some non-correlations (χ
2
is
close to 0 and lif t is close to 1), and association rules with conf idence > 0.8.
tions, some (surprising) non-correlations, and one-directional
association rules with high confidence values. Our correlation
analysis has been conducted at both the image level
and session level. Image-based analysis broadly refers to the
multi-modal set of data closest in time to an image snapshot,
including the audio snippet and sensor readings immediately
preceding a snapshot. Note that our sampled meaningful session
data set contains 218 sessions and 2,113 images, so the
image-based analysis is more stable, but the session-based
analysis still offers some important insights. Next, we describe
in detail the image-based correlation analysis results,
then discuss the differences in the session-based correlation
analysis results.
Camera position, i.e., front or back camera, is a key feature
that allows users to show different content. As shown in
Figure 11(L), Front Camera has strong positive correlations
with Person (seeing person in the image), Face (seeing face
in the image), and Normal (normal user), and strong negative
correlation with No Face (not seeing face in the image);
while Back Camera has strong positive correlation with No
Face and strong negative correlation with Face. This can
be explained by the notion that people typically use the front
camera to show their own faces and the back camera to show
some other content. In addition, the observation that normal
users and front camera have a strong positive correlation can
be used to differentiate normal users from misbehaving ones.
Since checking camera position on smartphones is an inexpensive
operation, this can be particularly useful for misbehavior
detection in mobile video chat services.
Face appearance in video chat is another important factor
to consider. In video chat services, people seek other interesting
people to chat with, and showing their faces help
keep people engaged in video chat sessions. As shown in
Figure 11(L), Face has a strong positive correlation with
Normal and a strong negative correlation with Misbehaving.
In other words, normal users tend to show their faces
while misbehaving users tend not to show their faces. We
believe the explanation is three-fold: (1) Normal users show
their faces so they can chat more effectively with their partners;
(2) Misbehaving users tend to hide their faces to avoid
being identified; and (3) Due to the limited aperture angle
of mobile phone cameras, it is difficult for a misbehavior
to show both his/her face and private body parts. Given
the strong positive (negative) correlations between Face and
Normal (Misbehaving), image-based face detection can be
quite effective for differentiating normal users from misbehaving
ones [24, 25].
Group chatting is rare in video chat services and most users
choose to chat alone with their remote partners (Table 1, Table
2). Still, when group chatting does occur, we observe that
Group has strong negative correlations with Back Camera,
Music & Sound, Silence/Noise, and Male. In other words,
when people chat as a group, they are less likely to use back
camera, have background music/sound or silence/noise. It isalso interesting to observe that male users tend not to chat
in groups. In addition, based on the association rules shown
in Figure 11(L), when users chat as a group, they are very
likely to show their faces and are very likely to be normal.
Female users also have some interesting behavioral characteristics:
They are likely to use the front camera and show
their faces (Figure 11(L)). Specifically, based on our sampled
data set, the probability for a female user to use the
front camera is 92% and the probability for a female user to
show her face is 84%. The intuition is that female users tend
to be popular in video chat services, i.e., more people are
interested in talking to female users. Therefore, using the
front camera and showing their faces can help female users
to chat more effectively with their remote partners.
Non-correlations can sometimes indicate something interesting
as well. For instance, we observe no negative correlation
between No Person and Voice, i.e., even when there
is no person shown in the video chat images, there can be
human voice in the audio recordings. This may indicate that
users sometimes show other content (not themselves) to their
remote partners while talking. Another example is Gender
versus Normal/Misbehaving. We originally expected to see
Male being correlated with Misbehaving, i.e., male users are
more likely to misbehave and misbehaving users are more
likely to be male. However, we did not find such correlation
in our data set. In other words, a misbehaving user can be
male or female, and knowing the gender of a user does not
increase or decrease his/her probability of misbehaving.
Session-based correlation analysis reveals fewer relations than
image-based correlation analysis (Figure 11). This is mainly
due to the fact that there are fewer number of sessions than
images. Most of the session-based relations are similar to
that of image-based relations. One correlation that is specific
to session-based analysis is the strong negative correlation
between Group and Remote Stop, i.e., video chat sessions of
group users are unlikely to be stopped by the remote party.
The fact that people do not hang up when talking to a group
of users on the remote side is interesting. Although group
chatting is rare in our current data set, they do seem to keep
people more engaged in a video chat session.
Meaningful User Analysis
Besides understanding the characteristics of meaningful sessions
(> 60 seconds), we also want to characterize “meaningful
users”, i.e., users who are more successful in participating
in longer video chat sessions. We define meaningful
users as the users who have more than 10% probability
to have a video chat session that is more than 30 seconds
long. Given that most users have less than 5% probability
to have longer than 30 seconds sessions, 10% is considered
significant. We pick 30 seconds instead of 60 seconds as
the session duration threshold, since the former allows us
to study 114 meaningful users while the latter has only 30
meaningful users. For comparison purposes, we also sampled
a similar number of “non-meaningful users”. And for
each meaningful or non-meaningful user, we randomly sampled
about 20 images from the user’s sessions. As a result,
Figure 12. Acceleration distribution of meaningful sessions.
we have obtained 2,714 images for 114 meaningful users
and 2,832 images for 123 non-meaningful users. We labeled
these images using the same taxonomy as defined before.
By comparing the taxonomy distributions of these two different
user sets, we find that gender is the most dominating
factor. In the set of meaningful users, we have 10 males
and 69 females (12.7% vs. 87.3%), while in the set of nonmeaningful
users, we have 56 males and 9 females (86.2%
vs. 13.8%). Note that the remaining 35 ‘meaningful users’
and 58 ‘non-meaningful users” images are absent of person
content so we cannot identify their genders. A similar gender
distribution occurs when using the 60 second definition,
where the skew is even more significant: 2 males and 20
females (9.1% vs. 90.9%). To summarize, compared with
male users, female users are much more likely to have long
meaningful sessions. This is partially due to the fact that
there are a lot more male users than female users in video
chat services (Figure 8), and male users appear to be more
interested in talking with female users.
Acceleration Analysis
For mobile video chat, acceleration information can be easily
collected and can offer useful contextual information such
as how the phones are positioned/oriented during video chat.
Figure 12 shows the acceleration distribution of meaningful
sessions. Note that the Adobe AIR API reverses the x and y
axes used in the Android API, and refers to the long side of
the mobile phone panel as the x axis (and the short side as
the y axis). Since MVChat organizes its video chat screen
using the landscape layout, users tend to hold the phone horizontally
and potentially rotate along the x axis. As a result,
acceleration centers around 0 for the x axis, centers around
G for the y axis, and spreads between −0.5G and G for
the z axis. Based on the acceleration distribution, we de-
fine four different classes to identify acceleration orientation
based on the degree of rotation along the x axis (Figure 13).
In other words, the phone can be placed horizontally facing
up, can be rotated along the long-side phone panel (tilting)
until facing down. We found that (1) Class 1 and Class 2
occur much more frequently than the other two classes, (2)
normal users are more likely to use the front camera with
orientation classes 1 and 2, (3) misbehaving users are more
likely to use the back camera with acceleration orientation
classes 1 and 2. This can be potentially leveraged to design 60
Horizon
(y=0, z=G)
Class 2 Class 1
Class 3
Class 4
Up
Down
 60 60 Figure 13. Acceleration
orientation classes
defined based on the degree
of rotation along the
x axis (long side plane of
smartphones).
more effective misbehavior detection mechanisms for mobile
video chat services.
CONCLUSIONS & FUTURE WORK
In this paper, we have presented the results of a large scale
study of mobile users of an Android video chat application,
which we designed and deployed for random video chat with
Omegle online users. We summarize our findings as follows.
We found that most sessions were short as users sought interesting
people to talk to, which motivates future work for
better user matching strategies. Normal users are highly correlated
with using the front camera and showing their faces,
whereas misbehaving users tend to hide their faces – which
suggests the exploration of camera position and face detection
for distinguishing normal users from misbehaving ones.
Users with a large enough fraction of sustained sessions are
disproportionately female, but surprisingly females were just
as likely to misbehave as males. Mobility introduces more
diversity than Web online content, while groups typically
imply normal behavior.
In the future, we plan to increase the scale of our dataset further,
incorporate analysis of more sensing modalities such
as the gyroscope and text messaging context, and better understand
what factors characterize meaningful user behavior,
with a view towards potentially more effective matching.
We also hope to use our findings to design an accurate misbehavior
classifier for mobile video chat users. Our analysis
reveals some good indicators for that issue like group users
will not misbehave, and different camera positions and orientations
inferred by acceleration result in different distributions
for normal and misbehaving users.
ACKNOWLEDGMENTS
We would like to acknowledge the support of NSF grants
CSR-1162614 and CIC-1048298. We would like to thank
Leif K-Brooks, the CEO of Omegle, for helping to enable
our MVChat client to be compliant with Omegle’s server,
and in providing data for our research. We would also like
to thank Xinyu Xing, Hanqiang Chen and Yuli Liang for
their advice on image analysis. We also sincerely thank the
anonymous reviewers for their comments and suggestions.
REFERENCES
1. M. G. Ames, J. Go, J. J. Kaye, and M. Spasojevic. Making love in the
network closet: the benefits and work of family videochat. In Proc of
the 2010 ACM conf. on Computer supported cooperative work,
CSCW ’10, pages 145–154, 2010.
2. X. Bao and R. Roy Choudhury. Movi: mobile phone based video
highlights via collaborative sensing. In Proc. of 8th Intl. ACM Conf.
on Mobile Systems, Applications, and Services, MobiSys ’10, pages
357–370, 2010.
3. M. Bohmer, B. Hecht, J. Sch ¨ oning, A. Kr ¨ uger, and G. Bauer. Falling ¨
asleep with angry birds, Facebook and kindle: a large scale study on
mobile application usage. In Proc. of the 13th Intl. Conf. on Human
Computer Interaction with Mobile Devices and Services, MobileHCI
’11, pages 47–56, 2011.
4. T. Buhler, C. Neustaedter, and S. Hillman. How and why teenagers
use video chat. In Proc. of the 2013 conf. on Computer supported
cooperative work, CSCW ’13, pages 759–768, 2013.
5. Chatroulette web site. http://www.chatroulette.com/.
6. H. Cheng, Y.-L. Liang, X. Xing, X. Liu, R. Han, Q. Lv, and S. Mishra.
Efficient misbehaving user detection in online video chat services. In
WSDM ’12, pages 23–32, 2012.
7. K. Church and B. Smyth. Understanding mobile information needs. In
Proc. of the 10th intl. conf. on Human computer interaction with
mobile devices and services, MobileHCI ’08, pages 493–494, 2008.
8. Y. Cui and V. Roto. How people use the web on mobile devices. In
Proc. of the 17th intl. conf. on World Wide Web, WWW ’08, pages
905–914, 2008.
9. J. Froehlich, M. Y. Chen, S. Consolvo, B. Harrison, and J. A. Landay.
Myexperience: a system for in situ tracing and capturing of user
feedback on mobile phones. In Proc. of the 5th intl. conf. on Mobile
systems, applications and services, MobiSys ’07, pages 57–70, 2007.
10. M. A. Hoque, M. Siekkinen, J. K. Nurminen, and M. Aalto.
Investigating streaming techniques and energy efficiency of mobile
video services. Technical report, 2012.
11. K. Inkpen, H. Du, A. Roseway, A. Hoff, and P. Johns. Video kids:
augmenting close friendships with asynchronous video conversations
in videopal. CHI ’12, 2012.
12. http://blog.appsfire.com/infographic-ios-apps-vs-web-apps/.
13. S. Jakubczak and D. Katabi. A cross-layer design for scalable mobile
video. In Proc. of the 17th annual intl. conf. on Mobile computing and
networking, MobiCom ’11, pages 289–300, 2011.
14. S. Jana, A. Pande, A. Chan, and P. Mohapatra. Mobile video chat :
Issues and challenges. IEEE Communications Magazine, Consumer
Communications and Networking Series, 2013.
15. L. Keller, A. Le, B. Cici, H. Seferoglu, C. Fragouli, and
A. Markopoulou. Microcast: cooperative video streaming on
smartphones. In Proc. of the 8th intl. conf. on Mobile systems,
applications, and services, MobiSys ’10, pages 57–70, 2012.
16. MeetMe web site (formerly MyYearbook). http://www.meetme.com/.
17. M. Milliken, S. ODonnell, K. Gibson, and B. Daniels. Older adults
and video communications: A case study. The Journal of Community
Informatics, 8(1), 2012.
18. Omegle web site. http://www.omegle.com/.
19. H. Raffle, G. Revelle, K. Mori, R. Ballagas, K. Buza, H. Horii,
J. Kaye, K. Cook, N. Freed, J. Go, and M. Spasojevic. Hello, is
grandma there? let’s read! storyvisit: family video chat and connected
e-books. CHI ’11, pages 1195–1204, 2011.
20. J. Scholl, P. Parnes, J. D. McCarthy, and A. Sasse. Designing a
large-scale video chat application. In Proc. of the 13th ACM intl. conf.
on Multimedia, MULTIMEDIA ’05, pages 71–80, 2005.
21. Study: 37% of U.S. teens now use video chat, 27% upload videos.
http://techcrunch.com/2012/05/03/study-37-of-u-s-teens-now-usevideo-chat-27-upload-videos/.
22. H. Verkasalo. Contextual patterns in mobile service usage. Personal
Ubiquitous Comput., 13:331–342, 2009.
23. J. Wang. Chitchat: Making video chat robust to packet loss. Master’s
thesis, Massachusetts Institute of Technology, 2010.
24. X. Xing, Y.-L. Liang, H. Cheng, J. Dang, S. Huang, R. Han, X. Liu,
Q. Lv, and S. Mishra. Safevchat: Detecting obscene content and
misbehaving users in online video chat services. In Proc. of the 20th
intl. conf. on World Wide Web, WWW ’11, pages 685–694, 2011.
25. X. Xing, Y.-l. Liang, S. Huang, H. Cheng, R. Han, Q. Lv, X. Liu,
S. Mishra, and Y. Zhu. Scalable misbehavior detection in online video
chat services. In Proc. of the 18th ACM SIGKDD conf. on Knowledge
discoery and data mining, KDD ’12, pages 552–560, 2012.