ArticlePDF Available

Understanding and controlling the filter bubble through interactive visualization: A user study

Authors:

Abstract and Figures

The "filter bubble" is a term which refers to people getting encapsulated in streams of data such as news or social network updates that are personalized to their interests. While people need protection from information overload and maybe prefer to see content they feel familiar or agree with, there is the danger that important issues that should be of concern for everyone will get filtered away and people will lack exposure to different views, living in "echo-chambers", blissfully unaware of the reality. We have proposed a design of an interactive visualization, which provides the user of a social networking site with awareness of the personalization mechanism (the semantics and the source of the content that is filtered away), and with means to control the filtering mechanism. The visualization has been implemented in a peer-to-peer social network, called MADMICA, and we present here the results of a large scale lab study with 163 crowd-sourced participants. The results demonstrate that the visualization leads to increased users' awareness of the filter bubble, understandability of the filtering mechanism and to a feeling of control over their data stream.
Content may be subject to copyright.
Understanding and Controlling the Filter Bubble through
Interactive Visualization: A User Study
Sayooran Nagulendra and Julita Vassileva
Department of Computer Science
University of Saskatchewan,
Saskatoon, Canada
{sayooran.nagulendra, julita.vassileva}@usask.ca
ABSTRACT
“The filter bubble” is a term popularized by Eli Pariser which
refers to people getting encapsulated in streams of data such as
news or social network updates that are personalized to their
interests. While people need protection from information overload
and maybe prefer to see content they feel familiar with and
viewpoint that they agree with, there is the danger that important
issues that should be of concern for everyone will get filtered
away and people will live in “echo-chambers”, blissfully unaware
of reality, and exposure to different views. We have proposed a
design of an interactive visualization, which provides the user of a
social networking site with awareness of the personalization
mechanism (the semantics and the source of the content that is
filtered away), and with means to control the filtering mechanism.
The visualization has been implemented in a peer-to-peer social
network and we present here the results of a qualitative and a
quantitative evaluation. The quantitative study with 163
participants demonstrates that the visualization leads to
increased users’ awareness of the filter bubble, understandability
of the filtering mechanism and to a feeling of control over the data
stream they are seeing.
Categories and Subject Descriptors
D.2.8 [Information Storage and Retrieval]: Information Search
and Retrieval – information filtering
General Terms
Design, Experimentation, Human Factors
Keywords
Visualization, Filter Bubble, Recommender Systems, Online
Social Networks
1. INTRODUCTION
Today, social networks provide a global platform for people to
share and collaborate with their friends and families. Facebook,
Twitter and Google+ are currently the most widely used social
networks. With the growth of mobile and web technologies, these
social networks are growing rapidly and millions of users are
sharing data with their friends and families. As of September
2013, Facebook has 1.15 billion users and 699 million daily
active users [28]. Nearly a quarter (24 %) of the content that is
shared on the internet is shared on Facebook [27] and more than
3.5 billion pieces of content shared each week [26], creating a
stream of data that can overload any user. The social data
overload problem is commonly solved by filtering out the
irrelevant data. Personalized stream filtering mechanisms aim at
reducing information overload by presenting the user with only
the content deemed to be the most relevant. Social media sites,
such as Facebook, Digg and YouTube, have already implemented
personalized stream filtering.
Paradoxically, the main problem with information filtering is that
they could be “too good”. The high level of optimization to the
interest of the user that typical algorithms lead to items that
remain in the data fit the user’s scope of interest that has been
inferred by the system from the user’s previous behavior, users
tend to becoming encapsulated in a “bubble” of their comfort,
seeing only content related to their interests, and being spared of
anything else. This is referred as “the filter bubble” problem.
We proposed an approach to make the user aware of the filtering
mechanism and take control over it. It is based on an interactive
visualization that shows the filter bubble and some features of the
hidden filtered data (its semantics and origin). The intention is to
make the user aware of the user model that the recommender
system has developed, so that they can consciously decide to
explore items that are filtered away by changing interactively her.
But showing what is hidden and filtered away from the stream can
increase the social data overload problem again. Therefore, the
main challenge is to find an effective visualization technique that
can be seamlessly integrated into the activity stream without
contributing additionally to the social data overload. In addition
to that the visualization design has to take into account of the
right amount of detail to expose in the hidden filtered social data
display the hidden social data stream in an understandable way to
the user.
In this paper we present a qualitative and a quantitative evaluation
of an interactive visualization which metaphorically visualizes the
filter bubble and provides awareness, understanding and control
of personalized filtering to alleviate the filter bubble problem.
2. RELATED WORK
Recommender Systems (RSs) are software tools and techniques
which adapt to the needs of an individual user and provide
personalized suggestions of most relevant information [1]. The
personalized suggestions help users to make decisions on various
types of items, such as what news items are interesting, what book
to read, what movie to watch and so on. Information filtering
systems can be considered as a type of recommender systems,
which select from a stream of data (e.g. news, certain events,
social updates, etc.) those that fit the scope of interest of the user.
Permission 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. To copy
otherwise, or republish, to post on servers or to redistribute to lists,
requires prior specific permission and/or a fee.
Conference’10, Month 1–2, 2010, City, State, Country.
Copyright 2010 ACM 1-58113-000-0/00/0010 …$15.00.
The difference between filtering and recommendation is that in
filtering the irrelevant data is simply not displayed, i.e. remains
hidden from the users, while in recommendation the relevant data
is highlighted in some way (e.g. shown first in a list of search
results, highlighted in a stream of data, etc.), but the irrelevant
data is still available for the user to see.
Recommendation techniques have been applied to personalize the
streams in online social networks such as Facebook, Google+ and
Twitter [2, 3]. Facebook’s edge rank algorithm is one such
filtering technique which presents a personalized stream of news
and friends’ status updates to the user by ranking every interaction
on the site [4]. While all these social networks are centralized,
Tandukar & Vassileva [5] have developed an interest-based
filtering model for a decentralized online social network (OSN),
which enables each peer to learn the user’s interests and to filter
away messages received from the user’s friends,.
Many researchers have worked on developing new RSs and
improving the accuracy of their filtering algorithms. However the
ultimate measure of success in this area is the user acceptance and
trust of the recommendations [6]. The way recommendations are
presented is critical for the user acceptance of recommender
systems. Visualization techniques can be deployed to provide an
intuitive “at a glance” explanation for recommendations and can
also motivate the user to accept the recommendation.
Presenting
the recommendations in a ranked list according to their
recommendation score is the most simple and commonly used
visualization technique.
Features like colour and font-size can be
used to emphasize recommended items in a stream or list or items
[7].
iBlogViz is a system to visualize blog archives. It uses many
visual cues to represent the blog content and social interaction
history with the blog entry which help to navigate the blog archive
quickly and easily. Particularly, visual cues about the social
response (comments) to the news can be used to help users
navigate stream data quickly to find interesting news [8]. Webster
& Vassileva [9] proposed an interactive visualization of a
collaborative filtering that allows the user viewer to see the other
users in her “neighborhood”, who are similar to her, and also to
change manually to degree of influence that any of the other users
can have on the recommendations of the viewer.
Rings is a
visualization of social data stream developed by Shi [10]. It is
organized around the people who post in the user’s Facebook
stream and empathizes users who have posted many and
influential posts recently, without filtering any posts. It helps the
users of OSN to browse social data efficiently and find out the
active users and the time pattern of their social updates.
As the activity stream in Online Social Network is personalized
according to the user’s interests, the user will ultimately only see
activities related to her interests and will thus have no opportunity
of discovering items not related to her current interests, or
developing new interests. “The filter bubble” is a term introduced
by Eli Pariser [11] to denote a limited scope of information
defined by the user’s interests and isolated from anything that
doesn’t belong to this scope.
Isolating the user in a filter bubble has its advantages and
disadvantages. The main advantage is that it can help users get
relevant information a lot faster while not causing social data
overload. On the other hand, there are number of problems
outlined by Eli Pariser [11]. The first one is the problem of
distortion of the content posted on the site or by the user’s friends
and the user does not know in what way the way is biased. Users
become less likely to be recommended information that is
important, but not “likeable”. The second problem is the
information equivalent of obesity.
Because of the users’ tendency
to give positive feedback, they will give feedback only to
information items they are most compulsively attracted to. Using
an analogy from food, users will be eating candy all the time, and
the filter bubble leave users locked in a world consisting of
information junk food. As a result the users are increasingly
surrounded by the ideas with which they are already familiar and
agree, while being protected from surprising information, or
information that challenges their views, the filter bubble threats
people’s open-mindedness and prevents learning. Psychologist
Lowenstein mentions that the “curiosity is aroused when we are
presented with an ‘information gap’” and Pariser suggests that the
existence of curiosity, is based on the awareness that something
exists that is hidden or unknown [11]. The third problem is the
matter of control i.e. the growth of user knowledge will be greatly
influenced by the algorithms and systems giving excessive power
to the computer scientists who develop the personalization
techniques.
The importance of these three problems increases rapidly, as an
increasing proportion of users are using OSN to get all their
information and news; and nearly all OSN deploy information
filtering to personalize their streams to users. Yet, most of the
personalization systems do not create awareness about what is
being hidden from the user.
Resnick et al. [12] outline some strategies for promoting diverse
exposure. They discuss two approaches: the first one is to build
diversity aware recommender systems and filtering mechanisms.
As an example of this approach, Tandukar and Vassileva [13]
developed an interest-based stream filtering technique, which
allows for diversity exposure by allowing popular items to pass
through the filter to ensure some serendipity. The second
approach is to provide tools and techniques that encourage users
to consider themselves searching for diverse exposure. Munson
has implemented a browser extension which displays the bias in a
user’s online news reading over time, which encourages users to
seek the diverse exposure of news [14].
Though algorithmic personalization approaches can certainly find
the most relevant content related to what users are already
interested in a more efficient manner human curators and
especially the user herself is probably the most appropriate agent
to take the responsibility for ensuring a diverse exposure, to
address the third problem outlined by Pariser. This means
enabling the users to select what they want to see as well as what
they do not want to see over the personalization presented by the
algorithms. To enable them to do this, it is necessary first to make
them aware of their filter bubble, as well as understanding of how
they got inside it, and how they can control it to let different kind
of information in and out, enlarge it or make it smaller… To our
best knowledge there is currently no existing work that aims to
create this kind of awareness and control in users. This is the aim
of our work.
3. VISUALIZATION DESIGN
The visualization of filter bubble has been designed and
implemented based on the personalized stream filtering used in
MADMICA
[15] - an implementation of a privacy-aware
decentralized (peer-to-peer) OSN using the Friendica open source
framework [16]. MADMICA implements an approach to filtering
social data, according to a model of the strength of the user’s
interests in different semantic categories overlaid over a model of
their social relationships, which was originally developed and
evaluated in a simulation [13].
The intuition behind the filtering
approach is that two people can be friends, but not share the same
level of interest in different topics or categories and not trust each
other’s judgment with regard to these categories. In essence, the
filtering approach is based on a model of the user’s interest in a
finite set of categories of social data that is overlaid with a model
of the strength of user interpersonal relationships (over each
category).
The visualization design is based on a bubble metaphor to make
the effect of the personalized stream filtering in OSNs more
understandable for the users (see Figure 1). The main goal of the
visualization is to creating awareness, understanding, and control
of personalized stream filtering in an OSN to alleviate the filter
bubble problem and increase the users’ trust in the system. It
divides the space of the screen in two parts - outside and inside
the bubble. The items that are inside the bubble are visible for the
user, those outside the bubble are those that have been filtered
away and are invisible in the stream (but they are shown in the
visualization). The visualization is personalized to the user
viewing it (let’s say Anna), and provides two alternative points of
view: one focusing on the user’s (Anna’s) friends (see Figure 2)
and one focusing on the semantic categories of the social data
originating from them in the OSN (see Figure 1). We assume that
there is a finite, enumerable set of sematic categories in which the
content can be classified. For practical reasons, these are
categories of higher level of generality, e.g. “news”, “technology”,
“health”, “sport”, similar to the categorization used by news
websites, Google, Yahoo, etc.
The category view shown in Figure 1 represents all the categories
of posts shared by Charlie during last week that were shown in
Anna’s newsfeed or filtered out by the system. All the category
circles inside the bubble represent the categories of posts that are
shown in Anna’s newsfeed; they represent the common categories
of interest between Anna and her friend Charlie. But Charlie has
more interests, which are outside Anna’s filter bubble and are
therefore being filtered out by the filtering mechanism based on
the past history of actions that Anna performed on the posts
shared by Charlie in the category “health”.
The “friends view” of Anna’s bubble visualization is shown in
Figure 2. It represents all the friends who shared some posts in the
“health” category during the last week that were shown in Anna’s
newsfeed or filtered out by the system. The position of each friend
circle relative to the big bubble is intended to create awareness
about the filtering i.e. whose posts the user (Anna) can see in her
newsfeed. Moreover, the filter bubble shape itself metaphorically
creates the awareness that the user is encapsulated in a bubble and
that there are friends outside of the bubble who have posted on
the topic but the user has not seen these posts.
As mentioned earlier, providing some understanding about the
personalized stream filtering is one of the main goals of this
visualization. Organizing posts by categories and friends gives
some understanding about the personalized filtering: that there is
a relationship between the categories of posts and the post origin
(the friends who shared them), and the underlying filtering
mechanism. In addition to that, it visualizes the common interests
between user and her friends i.e. what is shown inside the big
bubble are common interests between the user and her friends.
Providing control of the personalized stream filtering to the users
i.e. users can manually override the filtering system is another
main goal of this visualization. This is achieved by allowing users
to drag and drop the circles in and out of the big bubble. For
example, if Anna drags and drops the circle representing the
“games” category (see in Figure 1) from inside the big bubble to
its outside the user effectively tells the system that she does not
like to see that category of posts in her newsfeed in the future.
Similarly, the user could also drag and drop a friend from within
her “friends-view” bubble to the outside and it signals the system
to filter out the posts shared by that friend in the future. In the
reverse situation, when the user realizes that she is interested in
posts in category “health” shared by a friend (say, Glen), who is
outside her “friends-view” bubble in Figure 2 and wants to see his
posts in her newsfeed homepage in the future, she will drag and
drop that particular friend inside the big bubble. Apparently, this
Figure 1. Anna’s “category view” of her filter bubble
related to Charlie’s posts
Figure 2. Anna’s “friends view” of her filter bubble
related to a certain category (“health”) of posts
action is equivalent of the Anna coming out of her filter bubble
and explore new interests. If Anna wants to see all posts by Glen
in any category, she will select the “Friends” view and the generic
category “All” from the “Categories” menu and drag him in her
bubble.
Nagulendra and Vassileva presented justification of the
visualization design decisions and a pilot user study to evaluate
the usability and user acceptance of the visualization and whether
it achieves its goals of providing awareness, control and trust in
the filtering mechanism in MADMICA in [17]. Eleven (11)
graduate students from the MADMUC research lab used the
MADMICA system with the filter bubble visualization instead of
Facebook and shared interesting and research-related links over a
period of three weeks in March 2013.
The results of the study
showed that the filter bubble visualization makes the users aware
of the filtering mechanism, engages them in actions to correct and
change it, and as a result, increases the users’ trust in the system
[17]. Next we present the results of two more studies: a qualitative
study with 5 participants and a larger scale quantitative study with
163 Mechanical Turk participants.
4. EVALUATION
4.1 Qualitative User Study
A qualitative study was carried out to understand in-depth the user
perception of the filter bubble visualization i.e. what do users
think about the visualization. Five (5) participants from different
departments in the university took part in this study.
4.1.1 Experimental Setup
The study was carried out in a lab environment where users were
given computers to use the MADMICA system and the
visualization. The subjects were 5 university students from
different fields of study such as public education, public health
and statistics. They were recruited through a mailing list of
potential subjects for HCI studies. First, the users were given
some introduction to MADMICA and then about the filter bubble
problem. After the introduction, users were given instructions to
get familiar with the MADMICA newsfeed homepage and the
filter bubble visualization for 10 minutes. Once they have
explored the system, an interview was conducted. The interview
consists of a set of tasks related to with 15 different views that are
generated using the filter bubble visualization. They were asked to
interact with the systems and think aloud, the users’ actions were
observed and recorded and the users’ voice responses were
recorded. The views in the questionnaire were generated to collect
the perceptions about the visualization’s main goals: providing
awareness, understanding and control. Moreover, the views
included both the category view and friends view.
4.1.2 Methods
The recorded users’ voice responses were imported into NVivo
software [18] which is a platform for qualitative research analysis.
Then the voice responses were transcribed into text. With the help
of the NVivo software, thematic analysis was carried out to
identify the desirable and undesirable perceptions of the
visualization.
Thematic analysis categorizes qualitative data into themes. It
encodes the qualitative information into codes that act as labels
for sections of data [19]. The users’ responses were coded and the
codes were grouped into three: position of circle, size of circle
and drag action. While coding, the number of references for each
code was also recorded i.e. the frequency of that code in the
transcript of users’ responses. Then based on the three criteria, the
number of desirable references and undesirable references as
calculated. The three criteria were:
1. Circles inside the big bubble represent content or friends that
was shown in the user’s newsfeed.
2. Circles outside the big bubble represent content of friends that
were filtered out by the system
3. The visualization only shows the newsfeed shared by friends
organized into categories and friends.
4.1.3 Results
The thematic analysis results are summarized in Table 1. The
desirability percentage for a perception category is calculated as
the number of references that are desirable in that perception
category divided by the total number of references for the position
of circle visual representation multiplied by 100.
Regarding the position of circle visual representation, 108 total
references were made i.e. users mentioned 108 times in all of their
responses together that the position of circles relative to the wall
of the big bubble represents the user’s interest.
This is the most referred perception category (16.67%) about the
position of circle that is desirable. Some excerpts from the
transcript for the user’s interest perception category follow:
“categories outside the bubble represent the posts that the user
doesn’t want to see”, “categories inside the bubble represent my
interests”, “categories inside the bubble represent users main
interests for the selected duration”, “All the categories outside
the bubble represent that none of user’s friends posts are
related”, “categories outside the bubble represent the areas
outside of my interest for that period”, and “categories inside the
bubble represent that the user wants to focus on them”. The least
referred (1.85%) desirable perception category regarding the
position of circle is relationship. Some excerpts from the
transcript for the least referred desirable perception follows:
“friend circle outside the bubble for a category doesn’t mean that
the user unfriended with that friend”, “having some categories
inside the bubble for last month for a friend might mean an
acquaintance relationship”, and “friend relationship is
maintained regardless of user’s friends are outside the bubble”.
Some excerpts from the transcript for the least referred desirable
perception follows: friend circle outside the bubble for a
category doesn’t mean that the user unfriended with that friend”,
“having some categories inside the bubble for last month for a
friend might mean an acquaintance relationship”, and “friend
relationship is maintained regardless of user’s friends are outside
the bubble”. Some excerpts from the transcript for the least
referred desirable perception follows: “friend circle outside the
bubble for a category doesn’t mean that the user unfriended with
that friend”, “having some categories inside the bubble for last
month for a friend might mean an acquaintance relationship”,
and “friend relationship is maintained regardless of user’s
friends are outside the bubble”.
Table 1. Thematic analysis results
Feature/Visual
Representation
Perception
Category
Sources
(number
of users)
References
(desirable:
undesirable)
Desirability
percentage (%)
Undesirability
percentage (%)
Position of circle
(friend/category)
Common interest 4 13 (10:3) 9.26 (10/108) 2.78
Friends’ interest 4 40 (16:24) 14.81 22.22
Friends’ sharing 5 25 (18:7) 16.67 6.48
Interaction with
newsfeed
1 3 (3: 0) 2.8 0
User’s interest 5 23 (19:4) 17.59 3.7
Relationship 3 4(2:2) 1.85 1.85
Size of circle Number of posts 5 7 (6:1) 37.5(6/16) 6.25
Frequency of sharing 2 2 (2:0) 12.5 0
Friends’ interest 2 5 (4:1) 3.7 6.25
Common interest 1 2 (0:2) 0 12.5
Drag action Common interest 4 5 (4:1) 57.14 (4/7) 14.29
Relationship 1 2 (0:2) 0 28.57
The most referred (22.22%) undesirable perception category is the
friend’s interest. But here only 4 users have referred this whereas
in the most desirable perception all the 5 users referred it at some
point in the transcript. Following are some excerpts from the
transcript regarding the most undesirable perception: “categories
inside the bubble represent friend’s interest and outside
represents not interested” and “friend circle more in the middle
more interest in the category selected”. Like the least referred
desirable perception, the least referred undesirable perception is
relationship and the excerpt follows: “friend circle outside the
bubble represents unfriending”.
The number of posts related perception category is the most
referred (37.5%) desirable perception for the size of the circle.
Users perceive it as follows: “bigger circle for friend/category
represents more number of posts and small for less number of
posts”. As for the least referred (3.7%) desirable perception for
the size of the circle, users perceive it as friend’s interest i.e.
“larger circle for category means the selected friend has more
interest on that category”. In case of undesirable perception
category, the most referred (12.5%) one is common interest
(“small friend circle means more common interest between the
user and friends” and “larger circle represents user has less
interest on that friend”) and the least referred (6.25 %) one is
number of posts and friends’ interest (“small circle means
actually posted and big circle means less posted” and “bigger
circle outside the bubble represents less interest of friend on that
category”).
There are two perception categories that emerged by the thematic
analysis for drag action: common interest and relationship.
Common interest is the most referred desirable perception
category (57.14%) and there are no references for relationship on
desirable perception. The excerpt from the transcript for common
interest perceptions follows: “dragging a category inside means
to share more on that category with the friend”, “dragging in
may represent my future interest”, “drag out because I don’t
want to have common interest”, “drag out means lost interest in
that category from that friend”, and “drag all the friends outside
the bubble means I want to ignore all the news from them”. The
most and least referred undesirable perception category for the
drag action are, relationship (28.57%) and common interest
(14.29%) respectively. Excerpt related to relationship is
“dragging outside a friend/category means unfriend” and for the
common interest is “drag inside represents forcing the friend to
take interest on that category”
4.1.4 Discussion
The results of the qualitative data suggests that the subjects had
both perceptions which are desirable and undesirable. Desirable
perceptions (62.96%) regarding the position of circle had more
references than undesirable perceptions (37.04%). This shows that
most of the time the visualization users were aware of and had a
good understanding about the filtering. In particular, the emergent
codes such as common interests, friends’ interest, friends’ sharing,
interaction with newsfeed, user’s interest and relationship from
the thematic analysis clearly show that the users have some
understanding about the filtering. On the other hand the users also
had undesirable perceptions. This could be due to poor graphical
language of the visualization and interface as a whole. For
example, the reason that friends’ interest was perceived as the
most undesirable perception category, could be the poor label
texts of the dropdown menus which are used to view the filter
bubble in different dimensions. During the experiment, the labels
for the first two dropdown menu were as follows, Friend(s) and
Category(s). This creates a false perception when “Charlie” was
selected and the category was selected as “All” i.e. Charlie’s
interests were shown inside the bubble and what lies outside the
bubble were not the interests of Charlie. As a result the labels
were later changed into “From Friend(s)” and “On Category(s)”
(shown in Figure 1 depicting the updated version) before the
quantitative study.
The size of the circle is another indicator for creating awareness
about the filtering, i.e. having bigger size of the circle outside the
filter bubble would let the users know that there are more of posts
that have been filtered out by the system on that category from
that friend. Having 75% of desirable perceptions for size of the
circle shows that it is intuitive enough to create an awareness
about the filtering. The 25% of undesirable perceptions regarding
the size of the circle shows that the graphical language needs
improvement. For example, it would be clearer if there is a
number shown with the varying size. Moreover, the false
perceptions of common interest for the size of the circle showed
that users may have wrong perceptions about the meaning of the
size of circles. For example, size of the circles represent the
interests of the friends i.e. smaller circle means that the friend has
less interest on that category.
The drag action has 57.14% of desirable perception and 42.86%
of undesirable percentages. Despite the small difference,
considering the number of users who referred the perception gives
some clear indication that the majority of the participants (60%)
were able to understand the control functionality of the filter
bubble visualization. Though the perceptions were classified as
desirable and undesirable, both of them helped to get more insight
about the users perceptions about the visualization, improve the
visualization and helped to prepare the questions and answers for
the questionnaire of the quantitative study, presented in the next
section.
4.2 Quantitative User Study
A quantitative study was carried out to evaluate the
understandability of the visualization and whether the users
understand that the visualization provides awareness,
understanding and control of filtering and the filter bubble. The
study was conducted as an online survey and 163 participants
from different parts of the world participated in the study.
4.2.1 Hypotheses
The goal of this user study was to find out if the visualization is
understandable, if it creates awareness and understanding of the
personalized stream filtering mechanism and ability to control it
to alleviate the filter bubble. So the evaluation aims at testing the
following hypotheses.
1. Users understand that the visualization provides
awareness of the filtering and the filter bubble.
2. Users understand that visualization provides
understanding of the filtering and the filter bubble.
3. Users understand that visualization provides control of
the filtering and the filter bubble.
4. Users understand the visualization and its functions.
4.2.2 Experimental Setup
The study was carried out as an online survey. Unlike the
conventional online surveys, this survey had the interactive
visualization embedded into the survey so that users could explore
it and get some hands-on experience with it before answering the
survey. First, the participants were given some introduction about
the MADMICA social network and the filter bubble problem in
general. Then a sample newsfeed homepage was displayed in the
survey so that users could actually browse through the newsfeed
without leaving the survey page. The sample newsfeed contained
around 15 newsfeed items on 5 different categories such as
Health, News, Movies, Music and Sports from five different
friends named Alice, Bob, Charlie, Dave and Frank. The
participants were given instructions to assume that the
aforementioned people are their friends in MADMICA social
network and to browse through the newsfeed homepage as they
would do in Facebook. In addition to this, the newsfeed did not
show around 7 posts out of those five categories from different
friends i.e. the system filtered out some of the posts. Then the
users were presented with the interactive visualization exactly as
in the MADMICA system and were instructed to explore the
visualization. Then they were directed to the questionnaire to
answer the questions. The link to the online survey is given in the
appendix section of this paper.
4.2.3 Method
The online survey was conducted using Amazon Mechanical Turk
(MTurk) which is a popular crowd-sourced participant pool. We
ensured the data quality by placing attention check questions
(ACQs) and restricting participation to MTurk workers with
certain qualifications [20].
The suggested qualification among researchers to ensure data
quality was to allow participants who have the HIT Approval Rate
(%) for all Requesters' HITs greater than or equal to 95 [20]. But
we set even higher qualification to ensure the high data quality as
follows: HIT Approval Rate (%) for all Requesters' HITs greater
than or equal to 98% AND Number of HITs Approved greater
than or equal to 5000. The data collection continued for 1 week
and reached our target sample of 230. Then we analyzed the data
and checked the ACQ for validity and as a result, 163 valid
responses were collected. For each participant with a valid
response, we paid a compensation of 1$, which is a good rate for
an approximately 30-45 min. long study on MTurk.
The questionnaire contained 25 questions. The questions were
grouped according to the metrics which they intend to measure.
The metrics for understandability of the visualization are adapted
based on the International Standards for Software Quality
Evaluation [21]. Table 2 summarizes the metrics chosen for
measuring the understandability of the visualization [21]. There
are 3 independent variables: awareness, understanding and control
to assess the understandability of the visualization. Each of the
independent variables was evaluated using the metrics given in
Table 2 i.e. understandability of each independent variable was
calculated. In addition to that, the overall understandability
(referred as understandability hereafter) was also calculated using
the understandability metrics. Six (6) questions (2 Yes/No and 4
Multiple Choice Questions) were used to evaluate each of the
independent variables. Altogether, there were 18 questions that
were used to evaluate the overall understandability with 6
questions for each metrics. Our original hypotheses mentioned in
section 4.2.1 were converted into the statistical form with the
corresponding null hypothesis (see Table 3).
Table 2. Understandability Metrics
Metric Name Purpose Formula
Interpretation
of measured
value
Evident Functions
What proportions of functions
users were able to identify by
exploring the visualization
X = A / B
A = Number of functions identified by
the user
B = Total number of actual functions
0<=X<= 1
The closer to
1.0 is the better.
Function understand-ability What proportions of functions X= A / B 0<=X<= 1
users were able to understand
correctly by exploring the
visualization
A= Number of functions whose purpose
is correctly described by the user
B= Number of functions available
The closer to
1.0 is the better.
Understandable input and
output
Can users understand what is
required as input data and what is
provided as output by the
visualization?
X= A / B
A= Number of input and output data
items which user successfully
understands
B= Number of input and output data
items available from the visualization
0<=X<= 1
The closer to
1.0 is the better
Table 3. Statistical Hypotheses
Test H0 (null) H1 (alternative)
1 μ
Awareness
0.5 μ
Awareness
> 0.5
2 μ
Understanding
0.5 μ
Understanding
> 0.5
3 μ
Control
0.5 μ
Control
> 0.5
4 μ
Understandability
0.5 μ
Understandability
> 0.5
As shown in Table 3, we considered the mean value of
understandability for our hypothesis testing. The mean value is
0.5 according to the scale of metrics used to measure the
understandability. We set the null hypothesis as the mean value of
understandability is less than or equal to 0.5 i.e. users do not have
a clear understanding about the visualization. Our research
hypothesis is the mean value is greater than 0.5 i.e. users do have
clear understanding about the visualization. As mentioned in the
metrics Table 2, the closer this mean value to 1.0 is, the better the
understanding.
4.2.4 Results
4.2.4.1 Reliability Test
The internal consistency (reliability) of question items was
measured using the Cronbach’s alpha. Higher value of a reliability
coefficient (Cronbach’s alpha) is associated with lower random
error and greater measurement of the true score of the
understandability. The acceptable value of Cronbach’s Alpha
should be the range of 0.70 to 0.95 [22]. The rules of thumb when
considering Cronbach’s Alpha value explanation are as follows:
greater than 0.9 means excellent, greater than 0.8 means good,
greater than 0.7 means acceptable, greater than 0.6 means
questionable, greater than 0.5 means poor, and less than 0.5 is
unacceptable [23]. The measured value for the Cronbach’s alpha
is 0.7 for our questionnaire. This value is in the acceptable range.
4.2.4.2 Normality Test
The assessment of the normality of the data is a prerequisite and
essential to t-tests. The Normal Q-Q plot for understandability
was generated using SPSS (see Figure 3). If the data are normally
distributed, the data points will be close to the diagonal line. If the
data points move away from the line in a non-linear way then the
data are not normally distributed [24]. As we can see from the
Normality Q-Q Plot shown in Figure 3, the data is normally
distributed because the data points stay close to the diagonal line.
4.2.4.3 Hypothesis Test
One-sample t-test was used to determine whether the mean of a
particular data set is different from the particular value. Before
doing the t-tests, the following 4 assumptions were met:
understandability is measured at the ratio level,
the collected data
are independent which means that there is no relationship between
the observations, there are no significant outliers in the data, and
the understandability is approximately normally distributed [25].
Then the t-tests were conducted for the 4 hypothesis tests and the
results are summarized in the Table 4.
The first t-test was conducted for the hypothesis 1 defined in
section 4.2.1. The Mean understandability of awareness (M =
0.7117, SD = 0.2379) was higher than the tested understandability
value of 0.5, a statistically significant mean difference of 0.21,
95% CI [0.18 to 0.25], t (162) = 11.358, p < .001. Similarly, the t-
tests for hypothesis 2, 3, 4 were conducted and the results follow
respectively, the Mean understandability of understanding the
filtering (M = 0.6176, SD = 0.2159) was higher than the tested
understandability value of 0.5, a statistically significant mean
difference of 0.12, 95% CI [0.08 to 0.15], t (162) = 6.953, p <
.001, the Mean understandability of control (M = 0.7607, SD =
0.2246) was higher than the tested understandability value of 0.5,
a statistically significant mean difference of 0.26, 95% CI [0.23 to
0.30], t (162) = 14.824, p < .001 and the Mean understandability
of visualization (M = 0.6967, SD = 0.1808) was higher than the
tested understandability value of 0.5, a statistically significant
mean difference of 0.20, 95% CI [0.17 to 0.23], t (162) = 13.884,
p < .001. In all four tests, there were a statistically significant
difference between means (p < .001) and, therefore, we can reject
the null hypotheses defined in Table 3, and accept the alternative
hypotheses.
Figure 3. Normality Q-Q Plot of Understandability
Table 4. Hypothesis Analysis
Test Variable Mean 2-
tailed t
Degree of
freedom
(df)
1-tailed
Critical t
1-tailed
t <
2-tailed
t
Means
are in
correct
order
Alternative
Hypothesis
Accepted
1 Awareness .7117
11.358
162
1.6543
YES YES YES
2 Understanding .6176
6.953
162
1.6543
YES YES YES
3 Control .7607
14.824
162
1.6543
YES YES YES
4 Understandability .6967
13.884
162
1.6543
YES YES YES
4.2.4.4 Additional Test on Graphical Language
The key graphical language constructs of this visualization are,
1. The relative position of user's circles to the bubble (inside /
outside)
2. The size of the users' circles (larger - more posts)
3. Dragging user circles in and out (showing / filtering away)
In addition to the above 3 constructs, we identified another
potential construct from the qualitative study as follows: the
position of circles inside the bubble (closer to the center or to the
periphery). All the 3 other constructs were as part of each function
of the visualization (providing awareness, providing
understanding, and providing control) and were tested for
statistical significance. In order to test whether users interpret this
fourth construct or not, we included the answers based on this
construct for two of the questions in the survey. During the
analysis, we created a score for users based on how many out of
the 2 questions they did not select this construct as an answer.
Then the hypotheses were formed as follows: H0: μ
Score
0.5,
H1: μ
Score
> 0.5. One sample t-test was conducted and the results
are as follows: the Mean score for not selecting the graphical
construct (M = 0.9571, SD = 0.1405) was much higher than the
test score value of 0.5, a statistically significant mean difference
of 0.46, 95% CI [0.44 to 0.49], t (162) = 41.523, p < .001.There
were a statistically significant difference between means (p <
.001) and, therefore, we can reject the null hypothesis, and accept
the alternative hypothesis.
4.2.5 Discussion
The results of the quantitative study suggest that overall the new
users had better understanding about the visualization. By
comparing the means of variables Awareness, Understanding, and
Control, we can see that users have a better understanding
(0.7607) about the control of filtering and the filter bubble
provided by the visualization. This can be linked with the drag
and drop feature of the visualization, which is very popular and
commonly used action in many user interfaces and it is a very user
friendly user interface construct. On the other side, the users’
understanding about the visualization providing understanding to
the filtering and the filter bubble has the lower value (0.6176).
Though it is higher than 0.5, it clearly shows that the visualization
has to be improved on this aspect. A possible improvement could
be to provide some context sensitive help to the visual cues in the
visualization. The overall understandability value of the
visualization (0.6967) shows that the users had a better
understanding about the visualization after exploring it for the
first time and it could be considered as an intuitive visualization.
But it can be envisioned that the users will better understand if
there is a context sensitive help provided with the visualization.
Analyzing the t-test values gives us more insight into the
understandability measures. As mentioned earlier, the
understandability of visualization is calculated using the three
variables awareness, understanding and control. These three
variables are understandability variables and are measured using
the metrics presented in Table 2. The variables awareness,
understanding and control obtained a high 2-tailed value
respectively 11.358, 6.953, and 14.824. These values are
comparatively very high when compared with their relevant one-
tailed t-test value, which is 1.65. This indicates that these three
variables are a very good measure for the understandability of this
visualization.
The additional test on graphical language results suggest that the
users very rarely interpreted the position of circles inside the
bubble (closer to the center or to the periphery) i.e. very few users
selected it. A possible reason for this might be the nature of the
question; the users might have only focused on the first 3
graphical constructs which are intuitive and obvious. But it seems
a useful construct and could be added as an improvement to the
visualization in future.
5. CONCLUSION AND FUTURE WORKS
This paper presented the results of a qualitative and a quantitative
evaluation of an interactive visualization which metaphorically
visualizes the filter bubble in a P2P Social Network. The
qualitative study reveals several user perceptions which provide
desirable explanation for the awareness, understanding and
control of the filter bubble provided by the interactive
visualization. The quantitative study with 163 participants
demonstrates that the visualization leads to increased users’
awareness of the filter bubble, understandability of the filtering
mechanism and to a feeling of control over the data stream they
are seeing. Future work directions include conducting a study of
evaluating the intuitiveness of the visualization by comparing it to
the same interactive visualization provided with guided help.
6. REFERENCES
[1] Resnick, P., Varian, H.R.: Recommender systems.
Communications of the ACM. 40, 3, 56–58 (1997)
[2] Kywe, S.M. et al.: A Survey of Recommender Systems in
Twitter. Proc. 4th International Conference, SocInfo 2012,
Lausanne, Switzerland, December 5-7, 2012. pp. 420–433
(2012)
[3] Manish, A. et al.: An Online News Recommender System for
Social Networks. SIGIR-SSM. (2009)
[4] Kincaid, J.: EdgeRank: The Secret Sauce That Makes
Facebook’s News Feed Tick | TechCrunch, Available online
at: http://techcrunch.com/2010/04/22/facebook-edgerank/.
(accessed 6 July 2013)
[5] Tandukar, U., Vassileva, J.: Selective Propagation of Social
Data in Decentralized Online Social Network. Proc. UMAP
2011 Workshops, LNCS 7138. pp. 213–224 Springer-Verlag
Berlin Heidelberg (2012)
[6] Konstan, J.A., Riedl, J.: Recommender systems: from
algorithms to user experience. User Modeling and User-
Adapted Interaction. 22, 1-2, 101–123 (2012)
[7] Webster, A., Vassileva, J.: Visualizing personal relations in
online communities. Proceedings of the adaptive hypermedia
and adaptive web-based systems (AH’2006), June 21–23, pp.
223–233, Springer LNCS 4018, Dublin (2006)
[8] Indratmo, Vassileva, J., Gutwin, C.: Exploring blog archives
with interactive visualization. In: International conference on
Advanced Visual Interfaces (2008)
[9] Webster, A., Vassileva, J.: The KeepUP Recommender
System. Proc. 2007 ACM Conference on Recommender
Systems RecSys ’07. pp. 173–176 ACM, Minneapolis,
Minnesota, USA. (2007)
[10] Shi, S.: Keeping Up with the Information Glut by
Visualizing Patterns of Posting by Friends on Facebook,
http://hdl.handle.net/10388/ETD-2011-09-139, (accessed 11
Feb 2013)
[11] Pariser, E.: The Filter Bubble: What the Internet Is Hiding
from You. Penguin Press HC (2011)
[12] Resnick, P., A. Munson, S., Garrett, R.K., Stroud, N.J.,
Kriplean, T.: Bursting Your (Filter) Bubble: Strategies for
Promoting Diverse Exposure. Proc. Conference on Computer
supported Cooperative Work CSCW 2013 Companion proc.
pp. 95–100 ACM (2013)
[13] Tandukar U., Vassileva J.: Ensuring Relevant and
Serendipitous Information Flow in Decentralized Online
Social Network. Proc. AIMSA’2012, 15th biennial
conference on AI Methods Systems, Applications, Springer
Verlag, LNAI 7557, pp. 79-88. (2012)
[14] Munson, S.A., Resnick, P.: Presenting diverse political
opinions. Proc. 28th international conference on Human
factors in computing systems - CHI ’10. p. 1457 ACM Press,
New York, New York, USA (2010)
[15] Nagulendra, S., Vassileva, J.: Minimizing Social Data
Overload through Interest- Based Stream Filtering in a P2P
Social Network, Proc. IEEE International Conference on
Social Computing, SocialCom’2013 (2013)
[16] Macgirvin, M.: DFRN – the Distributed Friends & Relations
Network, Available online at:
https://macgirvin.com/spec/dfrn2.pdf. (accessed 2 Aug 2012)
[17] Nagulendra, S., Vassileva, J.: Providing Awareness,
Understanding and Control of Personalized Stream Filtering
in a P2P Social Network. 19th International Conference,
CRIWG 2013. pp. 61–76 Springer Berlin Heidelberg (2013).
[18] NVivo 10 for Windows, Available online at:
http://www.qsrinternational.com/products_nvivo.aspx.
(accessed 25 Feb 2014)
[19] Boyatzis, R.E.:Transforming qualitative information:
Thematic analysis and code development. Thousand Oaks,
London, & New Delhi: SAGE Publications (1998)
[20] Paolacci, G., Chandler, J.: Running experiments on Amazon
Mechanical Turk. 5, 5, 1–14 (2014).
[21] ISO/IEC TR 9126-2:2003 – Product Quality – External
Metrics, Available online at:
http://www.iso.org/iso/iso_catalogue/catalogue_tc/catalogue
_detail.htm?csnumber=22750. (accessed 2 Feb 2014)
[22] Bland, J. M., Altman, D. G.: "Statistics notes: Cronbach's
alpha," pp. 570-572 (1997)
[23] George, D., Mallery, P., SPSS for Windows step by step: A
simple guide and reference. 11.0 update (4th ed.), Boston:
Allyn & Bacon (2003)
[24] Testing for Normality using SPSS, Available online at:
https://statistics.laerd.com/spss-tutorials/testing-for-
normality-using-spss-statistics.php. (accessed 2 Mar 2014)
[25] One-Sample T-Test using SPSS, Available online at:
https://statistics.laerd.com/spss-tutorials/one-sample-t-test-
using-spss-statistics.php. (accessed 2 Mar 2014)
[26] An Infographic: The Biggest Shift since the Industrial
Revolution | TechnoBuffalo, Available online at:
http://www.technobuffalo.com/2010/06/01/an-infographic-
the-biggest-shift-since-the-industrial-revolution/. (accessed 5
Sept 2013)
[27] Chart of the day: How People Share Content on the Web,
Available online at: http://www.businessinsider.com/chart-
of-the-day-social-networking-sites-dominate-sharing-2009-7.
(accessed 4 Aug 2013)
[28] Facebook Investors, Available online at:
http://investor.fb.com/releasedetail.cfm?ReleaseID=780093.
(accessed 4 Aug 2013)
APPENDIX
The online survey link used for the quantitative user study:
http://nsayooran.usask.ca/limesurvey/index.php/985992/lang-en
... Ways of providing readers with feedback on their past media diet have already been explored by several studies and browser plug-ins (e.g. Blue Feed Red Feed; 6 Munson et al., 2013;Nagulendra and Vassileva, 2014). Research from other domains, such as movie recommendations, can also serve as an inspiration here (e.g. ...
... Generally, these tools provide a comprehensive overview of users' overall media diets compared with those of other users (e.g. Nagulendra and Vassileva, 2014) or with an ideal standard, such as ideological balance (e.g. Munson et al., 2013). ...
... While this approach might appear comparably drastic, there are also more subtle ideas. For example, visual exploration tools have been proposed to explore dissonant viewpoints (Graells-Garrido et al., 2016), new topics (Sullivan et al., 2019) or content beyond one's own filter bubble (Nagulendra and Vassileva, 2014). Relatedly, Bountouridis et al. (2018) developed a tool that lets readers explore content that particular articles omitted. ...
Article
Full-text available
Growing concern about the democratic impact of automatically curated news platforms urges us to reconsider how such platforms should be designed. We propose a theoretical framework for personalised diversity nudges that can stimulate diverse news consumption on the individual level. To examine potential benefits and limitations of existing diversity nudges, we conduct an interdisciplinary literature review that synthesises theoretical work on news selection mechanisms with hands-on tools and implementations from the fields of computer science and recommender systems. Based thereupon, we propose five diversity nudges that researchers and practitioners can build on. We provide a theoretical motivation of why, when and for whom such nudges could be effective, critically reflect on their potential backfire effects and the need for algorithmic transparency, and sketch out a research agenda for diversity-aware news recommender design. Thereby, we develop concrete, theoretically grounded avenues towards facilitating diverse news consumption on algorithmically curated platforms.
... However, from the user's point of view, exploration-oriented recommender systems seem to be essential to fulfill users' needs for variety and novelty seeking (Liang, 2019;McAlister & Pessemier, 1982). Additionally, these recommenders can help to broaden users' taste horizons, encourage users to explore their blind spots (Tintarev et al., 2018), and mitigate potential filter bubble issues (Nagulendra & Vassileva, 2014). ...
... Users may not always be aware of the exploration possibilities, and from the recommender perspective, system-initiated exploration is a difficult task but could be supported by using psychological mechanisms such as visualizations and nudging to create a psychology-informed recommender system (Lex et al., 2021). To encourage users to explore away from their current preferences, previous work has utilized visualizations to help users identify their blind spots (Tintarev et al., 2018) or filter bubbles (Nagulendra & Vassileva, 2014) by comparing their tastes with the global tastes of other users or the tastes of their friends. By exposing users to their under-explored tastes, users were nudged to explore their blind spots (Tintarev et al., 2018). ...
Article
Full-text available
Recommender systems are efficient at predicting users’ current preferences, but how users’ preferences develop over time is still under-explored. In this work, we study the development of users’ musical preferences. Exploring musical preference consistency between short-term and long-term preferences in data from earlier studies, we find that users with higher musical expertise have more consistent preferences for their top-listened artists and tags than those with lower musical expertise, while high expertise users also show more diverse listening behavior. Users typically chose to explore genres that were close to their current preferences, and this effect was stronger for expert users. Based on these findings, we conducted a user study on genre exploration to investigate (1) whether it is possible to nudge users to explore more distant genres, (2) how users’ exploration behavior within a genre is influenced by default recommendation settings that balance personalization with genre representativeness in different ways, and (3) how nudging for exploration increases the perceived helpfulness of recommendations for users to explore. Our results show that users were more likely to select the more distant genres if these genres were nudged by presented at the top of the list, however, users with high musical expertise were less likely to do so. We also find that the more representative default slider, which by default recommended more genre-representative tracks, made users set the slider at a less personalized level. The more representative slider position alone did not promote more exploration of users’ current preferences but when combined with the nudge for distant genres, the more representative slider positions effectively nudged users to explore. Nudging to explore does not necessarily lead to an increase in perceived helpfulness. On the one hand, nudging for exploration improves the perceived helpfulness for exploration by making users explore away from their current preferences. On the other hand, it reduces helpfulness due to lower perceived personalization as users move away from their personalized recommendations. To improve perceived helpfulness, it seems necessary to provide a balanced trade-off between exploration and personalization.
... Gobo, a web platform, tries to increase transparency by explaining how algorithms work and by enabling users to regain control over their news feeds, by managing and filtering contents [1]. Another example is a filter bubble visualization design from Nagulendra & Vassileva [9]. This design allows users to see which of their friends' feeds they are currently exposed to, which visually explains how algorithms manage friends' influence on social media. ...
... Source filter ( Figure 4) categorizes a user's feed by source. Unlike Nagulendra & Vassileva's filter bubble visualization design [9], VisualBubble not only shows friends as influencing sources, but also other sources such as social media influencers or organizations the user is following. Outside "Your bubble" are the sources that the user is not exposed to but that are popular in the region. ...
... One approach is preventive -trying to avoid filter bubbles by running algorithms in the background to ensure balance and diversity (e.g., [17]). The other approach is permissive -allowing filter bubbles but giving the user control over them, allowing the user to play the corrective factor (e.g., [4,5,16,25,26]), and enabling a kind of value co-creation between the user and the recommender system. We want to focus on the latter -the manageability of filter bubbles. ...
... Nagulendra and Vassileva [25,26] introduced an interactive visualization approach to illustrate and increase the awareness of the filter bubble in Online Social Networks. They designed their visualization based on a bubble metaphor to improve the comprehensibility of the filtering process in a way that is intuitive for the user. ...
... A filter bubble refers to a social environment that lacks the exposure to diverse beliefs and viewpoints. It can occur in undiversified communities, where people are encapsulated within a stream of media content (e.g., videos or news articles) that is optimized to match their specific preferences [70]. This effect can be created or reinforced by recommender systems, by over-personalizing the media content based on the users' interests, and consequently, trapping them within an unchanging environment [77]. ...
... Instead, they will be surrounded by the viewpoints of like-minded users, and protected from surprising information, or information that challenges their opinions. This may ultimately weaken their creativity, as well as their thinking and learning capabilities [70]. 2 Echo chambers-another potential effect of recommendations-refer to a polarized environment, where only certain viewpoints, information, and beliefs are shared via communication. In such an environment, the users' viewpoints are repeatedly amplified through recurring exposure to similar media content [40]. ...
Article
Full-text available
Reading or viewing recommendations are a common feature on modern media sites. What is shown to consumers as recommendations is nowadays often automatically determined by AI algorithms, typically with the goal of helping consumers discover relevant content more easily. However, the highlighting or filtering of information that comes with such recommendations may lead to undesired effects on consumers or even society, for example, when an algorithm leads to the creation of filter bubbles or amplifies the spread of misinformation. These well-documented phenomena create a need for improved mechanisms for responsible media recommendation, which avoid such negative effects of recommender systems. In this research note, we review the threats and challenges that may result from the use of automated media recommendation technology, and we outline possible steps to mitigate such undesired societal effects in the future.
... On an academic level, more research should be conducted on algorithmic (non-)transparency and how it impacts exposure to news content. Previous research on algorithmic transparency has found that algorithmic transparency can help people increase their algorithmic awareness and reflect on their behaviour (Eslami et al., 2015;Munson et al., 2013;Nagulendra & Vassileva, 2014). However, to the best of our knowledge, little research has been done on how algorithmic transparency can steer people towards more diverse news exposure. ...
Article
Scholars are increasingly studying how news recommenders can provide audiences with diverse news offers. However, despite their increased interest, little research has been conducted on how news audiences engage with news recommenders. This article assesses how a news recommender’s design affects people’s news exposure behaviours and perceptions. To do so, we conducted an online experiment (n = 337) in a virtual news environment with four different news recommenders, including three news recommenders that aimed to stimulate diverse news exposure. The results show that all diversity-based news recommenders steered users towards more diverse exposure behaviour, with a personalised diversity-based news recommender being the most effective. We also found that people using a diversity-based news recommender did not think they read more diversely, pointing towards a so-called “diversity paradox.” We discuss what this paradox means and how it relates to the secretive nature of news algorithms. We also make a call to policymakers, news organisations and scholars to give transparency and diversity-based news recommenders a more pronounced place in the news environment and in future news recommendation research.
... For this reason they may constitute not only a great opportunity for the diffusion of knowledge, but also a great threat for the stability of social fabric and the function-ing of democracy. A filter bubble occurs when a user is selectively exposed predominantly to content that tends to reinforce his/her current opinion/belief/state, while suppressing other alternatives [1][2][3] . In OSM this typically happens because of personalized recommender systems, which leverage information on past user activity to provide suggestions which, aiming at maximizing user satisfaction, tend to be very similar to what the user has already shown to prefer [4][5][6] . ...
Preprint
Full-text available
Social media influence online activity by recommending to users content strongly correlated with what they have preferred in the past. In this way they constrain users within filter bubbles that strongly limit their exposure to new or alternative content. We investigate this type of dynamics by considering a multistate voter model where, with a given probability $\lambda$, a user interacts with a "personalized information" suggesting the opinion most frequently held in the past. By means of theoretical arguments and numerical simulations, we show the existence of a nontrivial transition between a region (for small $\lambda$) where consensus is reached and a region (above a threshold $\lambda_c$) where the system gets polarized and clusters of users with different opinions persist indefinitely. The threshold always vanishes for large system size $N$, showing that consensus becomes impossible for a large number of users. This finding opens new questions about the side effects of the widespread use of personalized recommendation algorithms.
Article
Filter Bubbles, exacerbated by use of digital platforms, have accelerated opinion polarization. This research builds on calls for interventions aimed at preventing or mitigating polarization. This research assesses the extent that an online digital platform, intentionally displaying two sides of an argument with methodology designed to “open minds” and aid readers willingness to consider an opposing view. This “open mindedness” can potentially penetrate online filter bubbles, alleviate polarization and promote social change in an era of exponential growth of discourse via digital platforms. Utilizing “The Perspective” digital platform, 400 respondents were divided into five distinct groups varying in number of articles reading material related to “Black Lives Matter” (BLM). Results indicate that those reading five articles, either related or unrelated to race, were significantly more open-minded towards BLM than the control group. Those who read five race-related articles also showed significantly reduced levels of holding a hardliner opinion towards BLM than control.
Article
Social media influence online activity by recommending to users content strongly correlated with what they have preferred in the past. In this way, they constrain users within filter bubbles strongly limiting their exposure to new or alternative content. We investigate this type of dynamics by considering a multistate voter model where, with a given probability [Formula: see text], a user interacts with “personalized information,” suggesting the opinion most frequently held in the past. By means of theoretical arguments and numerical simulations, we show the existence of a nontrivial transition between a region (for small [Formula: see text]) where a consensus is reached and a region (above a threshold [Formula: see text]) where the system gets polarized and clusters of users with different opinions persist indefinitely. The threshold always vanishes for large system size [Formula: see text], showing that a consensus becomes impossible for a large number of users. This finding opens new questions about the side effects of the widespread use of personalized recommendation algorithms.
Article
Full-text available
Broadcast media are declining in their power to decide which issues and viewpoints will reach large audiences. But new information filters are appearing, in the guise of recommender systems, aggregators, search engines, feed ranking algorithms, and the sites we bookmark and the people and organizations we choose to follow on Twitter. Sometimes we explicitly choose our filters; some we hardly even notice. Critics worry that, collectively, these filters will isolate people in information bubbles only partly of their own choosing, and that the inaccurate beliefs they form as a result may be difficult to correct. But should we really be worried, and, if so, what can we do about it? Our panelists will review what scholars know about selectivity of exposure preferences and actual exposure and what we in the CSCW field can do to develop and test ways of promoting diverse exposure, openness to the diversity we actually encounter, and deliberative discussion.
Conference Paper
Full-text available
Research in recommender systems has traditionally focused on improving the predictive accuracy of recommendations by developing new algorithms or by incorporating new sources of data. However, several studies have shown that accuracy does not always correlate with a better user experience, leading to recent research that puts emphasis on Human-Computer Interaction in order to investigate aspects of the interface and user characteristics that influence the user experience on recommender systems. Following this new research this paper presents SetFusion, a visual user-controllable interface for hybrid recommender system. Our approach enables users to manually fuse and control the importance of recommender strategies and to inspect the fusion results using an interactive Venn diagram visualization. We analyze the results of two field studies in the context of a conference talk recommendation system, performed to investigate the effect of user controllability in a hybrid recommender. Behavioral analysis and subjective evaluation indicate that the proposed controllable interface had a positive effect on the user experience.
Conference Paper
In Online Social Networks (OSNs) users are often overwhelmed with the huge amount of social data, most of which are irrelevant to their interest. Filtering of the social data stream is the way to deal with this problem, and it has already been applied by OSNs, such as Facebook. Unfortunately, personalized filtering leads to “the filter bubble” problem where the user is trapped inside a world within the limited boundaries of her interests and cannot be exposed to any surprising, desirable information. Moreover, these OSNs are black boxes, providing no transparency of how the filtering mechanism decides what is to be shown in the social data stream. As a result, the user trust in the system can decline. This paper proposes an interactive method to visualize the personalized stream filtering in OSNs. The proposed visualization helps to create awareness, understanding, and control of personalized stream filtering to alleviate “the filter bubble” problem and increase the users’ trust in the system. The visualization is implemented in MADMICA – a privacy aware decentralized OSN, based on the Friendica P2P protocol. We present the results of a small-scale study to evaluate the user experience with the proposed visualization in MADMICA.
Article
Source: Democracy Now! JUAN GONZALEZ: When you follow your friends on Facebook or run a search on Google, what information comes up, and what gets left out? That's the subject of a new book by Eli Pariser called The Filter Bubble: What the Internet Is Hiding from You. According to Pariser, the internet is increasingly becoming an echo chamber in which websites tailor information according to the preferences they detect in each viewer. Yahoo! News tracks which articles we read. Zappos registers the type of shoes we wear, we prefer. And Netflix stores data on each movie we select. AMY GOODMAN: The top 50 websites collect an average of 64 bits of personal information each time we visit and then custom-designs their sites to conform to our perceived preferences. While these websites profit from tailoring their advertisements to specific visitors, users pay a big price for living in an information bubble outside of their control. Instead of gaining wide exposure to diverse information, we're subjected to narrow online filters. Eli Pariser is the author of The Filter Bubble: What the Internet Is Hiding from You. He is also the board president and former executive director of the group MoveOn.org. Eli joins us in the New York studio right now after a whirlwind tour through the United States.
Conference Paper
This paper presents a novel peer-to-peer architecture for decentralized Online Social Network and a mechanism that allows each node to filter out irrelevant social data, while ensuring a level of serendipity, by letting important information pass, even if it does not fall in the areas of interest of the user. The evaluation of the approach, using an Erlang simulation with 2318 nodes shows that it works as it was designed to: with the increasing number of social data passing through the network, the nodes learn to filter out irrelevant data, while serendipitous important data is able to pass through the network. Future work will implement the mechanism in a decentralized OSN and evaluate it in a real community of users, in an enterprise context.
Conference Paper
Twitter is a social information network where short messages or tweets are shared among a large number of users through a very simple messaging mechanism. With a population of more than 100M users generating more than 300M tweets each day, Twitter users can be easily overwhelmed by the massive amount of information available and the huge number of people they can interact with. To overcome the above information overload problem, recommender systems can be introduced to help users make the appropriate selection. Researchers have began to study recommendation problems in Twitter but their works usually address individual recommendation tasks. There is so far no comprehensive survey for the realm of recommendation in Twitter to categorize the existing works as well as to identify areas that need to be further studied. The paper therefore aims to fill this gap by introducing a taxonomy of recommendation tasks in Twitter, and to use the taxonomy to describe the relevant works in recent years. The paper further presents the datasets and techniques used in these works. Finally, it proposes a few research directions for recommendation tasks in Twitter.
Conference Paper
In Online Social Networks (OSNs) users are overwhelmed with the huge amount of social data, most of which are irrelevant to their interest. Filtering of the social data stream is the way to deal with this problem, and it has already been applied by centralized OSNs, such as Facebook. However, it is much harder to filter the social data stream in decentralizedOSNs. Decentralized OSNs, mostly based on P2P architectures, such as Diaspora or Friendica, have been proposed as an alternative to the currently dominant centralized OSNs, where people are forced to share their data with the site, and thus lose their control and rights over it. This paper presents an implementation of an interest based stream filtering mechanism using Mad mica - a decentralized OSN, based on the FriendicaP2P protocol. The mechanism uses the interaction between users to construct a model of user interests overlaid on the relationships of users with their friends, which acts as a filter later while propagating social data. So Mad mica provides a solution to two problems simultaneously - the problem of user privacy and control over their data (through its decentralized architecture) and the problem of social data overload (through its filtering mechanism). We present the results of a pilot study to evaluate the user experience with Madmica.