Online Information Review

Visual Twitter Analytics (Vista): Temporally changing sentiment and the discovery
of emergent themes within sport event tweets
Orland Hoeber Larena Hoeber Maha El Meseery Kenneth Odoh Radhika Gopi
Article information:
To cite this document:
Orland Hoeber Larena Hoeber Maha El Meseery Kenneth Odoh Radhika Gopi , (2016),"Visual Twitter
Analytics (Vista)", Online Information Review, Vol. 40 Iss 1 pp. 25 - 41
Permanent link to this document:
[Link]
Downloaded by UNIVERSITY OF AARHUS At 07:38 10 August 2016 (PT)

Downloaded on: 10 August 2016, At: 07:38 (PT)

References: this document contains references to 33 other documents.
To copy this document: permissions@[Link]
The fulltext of this document has been downloaded 292 times since 2016*
Users who downloaded this article also downloaded:
(2016),"Informing brand messaging strategies via social media analytics", Online Information Review,
Vol. 40 Iss 1 pp. 6-24 [Link]
(2016),"Competitive intelligence in social media Twitter: iPhone 6 vs. Galaxy S5", Online Information
Review, Vol. 40 Iss 1 pp. 42-61 [Link]
(2016),"Discovering shilling groups in a real e-commerce platform", Online Information Review, Vol.
40 Iss 1 pp. 62-78 [Link]

Access to this document was granted through an Emerald subscription provided by emerald-
srm:132724 []
For Authors
If you would like to write for this, or any other Emerald publication, then please use our Emerald
for Authors service information about how to choose which publication to write for and submission
guidelines are available for all. Please visit [Link]/authors for more information.
About Emerald [Link]
Emerald is a global publisher linking research and practice to the benefit of society. The company
manages a portfolio of more than 290 journals and over 2,350 books and book series volumes, as
well as providing an extensive range of online products and additional customer resources and
services.
Emerald is both COUNTER 4 and TRANSFER compliant. The organization is a partner of the
Committee on Publication Ethics (COPE) and also works with Portico and the LOCKSS initiative for
digital archive preservation.

*Related content and download information correct at time of download.

 The current issue and full text archive of this journal is available on Emerald Insight at:
[Link]/[Link]

Visual Twitter Analytics (Vista) Visual Twitter

Analytics
Temporally changing sentiment and the (Vista)
discovery of emergent themes within sport
event tweets 25
Orland Hoeber
Received 28 February 2015
Department of Computer Science, University of Regina, Regina, Canada Revised 17 June 2015
Larena Hoeber 17 July 2015
Accepted 18 September 2015
Faculty of Kinesiology and Health Studies, University of Regina, Regina,
Canada, and
Downloaded by UNIVERSITY OF AARHUS At 07:38 10 August 2016 (PT)

Maha El Meseery, Kenneth Odoh and Radhika Gopi

Department of Computer Science, University of Regina, Regina, Canada

Abstract
Purpose – Due to the size and velocity at which user generated content is created on social media
services such as Twitter, analysts are often limited by the need to pre-determine the specific topics and
themes they wish to follow. Visual analytics software may be used to support the interactive discovery
of emergent themes. The paper aims to discuss these issues.
Design/methodology/approach – Tweets collected from the live Twitter stream matching a user’s
query are stored in a database, and classified based on their sentiment. The temporally changing
sentiment is visualized, along with sparklines showing the distribution of the top terms, hashtags,
user mentions, and authors in each of the positive, neutral, and negative classes. Interactive tools are
provided to support sub-querying and the examination of emergent themes.
Findings – A case study of using Vista to analyze sport fan engagement within a mega-sport event
(2013 Le Tour de France) is provided. The authors illustrate how emergent themes can be identified
and isolated from the large collection of data, without the need to identify these a priori.
Originality/value – Vista provides mechanisms that support the interactive exploration among
Twitter data. By combining automatic data processing and machine learning methods with interactive
visualization software, researchers are relieved of tedious data processing tasks, and can focus on the
analysis of high-level features of the data. In particular, patterns of Twitter use can be identified,
emergent themes can be isolated, and purposeful samples of the data can be selected by the researcher
for further analysis.
Keywords Twitter, Visual analytics, Exploratory data analysis, Sport analytics
Paper type Research paper

Introduction
Twitter is a popular micro-blogging platform that allows individuals and organizations
to post and share short messages in a public, open, and unfiltered medium. The widespread
adoption and the willingness of users to comment on a broad range of topics have resulted
in Twitter becoming a valuable source of information regarding public opinion, citizen
reporting, and social interaction. While these features make Twitter attractive for
researchers, there are a number of significant problems for conducting such research.
Given the short and cryptic nature of the textual element of the tweets, it is often
necessary to analyze them manually using traditional content analysis approaches. Online Information Review
In addition, retrieving a large number of tweets via the standard Twitter interface is not Vol. 40 No. 1, 2016
pp. 25-41
straightforward, and there is an important temporal aspect to the tweets that must be © Emerald Group Publishing Limited
1468-4527
considered during the analysis. To complicate matters further, for a given topic of DOI 10.1108/OIR-02-2015-0067
 OIR interest, it is possible for an extremely large number of tweets to be generated in a very
40,1 short period of time. In many cases, analyzing Twitter data is a big data problem,
exhibiting high volume, velocity, and variety (Russom, 2011).
Within the sport context, Twitter has been embraced and promoted as a mechanism
to enhance communication among sport organizations, athletes, fans, and the media
(Hambrick et al., 2010; Pegoraro, 2010). A common approach for analyzing Twitter data
26 within sport management and sport communication research has been to sample the
tweets, using methods such as stratified sampling (Kassing and Sanderson, 2010;
Pegoraro, 2010), stratified random sampling (Hambrick et al., 2010), and systematic
sampling (Blaszka et al., 2012). While such sampling methods allow the large number of
tweets available to be reduced to a collection that can be analyzed manually, new
problems are introduced. These include the possibility of missing important or
meaningful tweets, the interaction between stakeholders via Twitter activity, or
Downloaded by UNIVERSITY OF AARHUS At 07:38 10 August 2016 (PT)

the temporal relationships between the tweets and micro-events associated with the
phenomenon under investigation.
Some have called for better methods for studying such data, arguing that small-
scale approaches are inappropriate for large-scale data sets (Mahrt and Scharkow,
2013; Tinati et al., 2014). In particular, Tinati et al. (2014, p. 6) advocate for the use of
“technical capabilities with in-depth qualitative research methods” to enhance the rigor
and theoretical development of research involving big data. We propose that such
technical capabilities should simultaneously leverage the capabilities of automatic
information processing and human judgment (Shneiderman and Plaisant, 2010).
Our particular work takes a visual analytics approach to supporting the analysis of
Twitter data. Information visualization takes advantage of the powerful processing
capabilities of human visual perception (Ward et al., 2010; Ware, 2013), enabling people
to see the information and visually identify patterns and relationships. Visual analytics
extends the reach of information visualization, focussing on the application of visual
techniques to support human-centric problem solving and data analysis tasks. Data
processing and machine learning approaches are combined with information
visualization and human-computer interaction methods, with the goal of supporting
data exploration, analytic reasoning, information synthesis, hypothesis development
and testing, and human decision making (Keim et al., 2008; Thomas and Cook, 2006).
The ultimate aim is to take advantage of the powerful computation and storage
capabilities of the computer to extract and infer relevant information, present this
information to the analyst in a visual manner, and support their cognitive and analytic
processes through interactive exploration.
Visual Twitter Analytics (Vista[1]) was developed as a tool to support the
exploration of the temporally changing sentiment within sport-related tweets (Hoeber
et al., 2013), guided by the fundamental principles of visual analytics. The software
extracts data from Twitter based on user-specified queries, performs automatic
sentiment analysis on the tweets (Feldman, 2013), provides visual timeline
representations that allow comparisons of the positive, neutral, and negative
sentiments among the tweets, and shows the locations of the tweets on a map. The
system automatically extracts the most commonly used hashtags, terms, user
mentions, and authors within the collection of tweets, providing sparkline
representations of the distributions of each of these over time. Vista is highly
interactive, supporting temporal zooming, adjustment of the temporal aggregation,
tweet inspection, sub-querying, and timeline comparisons. A screenshot of the core
interface is provided in Figure 1.
 Visual Twitter
Analytics
(Vista)

27
Downloaded by UNIVERSITY OF AARHUS At 07:38 10 August 2016 (PT)

Figure 1.
A screenshot of vista
showing the timeline
and geolocation of
positive, neutral, and
negative tweets
during one week of
the 2013 Le Tour de
France cycling race

While there are a large number of open source, commercial, and research prototypes
that use sentiment classification to support the analysis of Twitter data, the novelty
of Vista is in the interactive features developed within a visual analytics framework
to support the exploration and analytical reasoning about the data. Our goal is not to
use sentiment to detect emerging themes directly, but instead to use it as a starting
point for a data analyst to discover these themes. The human decision maker is
empowered to make discoveries within the data, aided by machine learning and
interactive visual interfaces. The remainder of this paper outlines related work in this
domain, presents the core features of Vista, provides examples of exploration of
Twitter data and the creation of purposeful samples using Vista, and concludes with
a summary of the primary contributions, the generalizability of the approach, and
future work.

Literature review
A review of Twitter-based sport communication research indicates that many
researchers manually analyze and classify the tweets under investigation. In order to
enable this process, researchers generally select a small subset of the total sample of
relevant tweets. For example, Blaszka et al. (2012) undertook a content analysis of
 OIR tweets using the #WorldSeries hashtag to determine who was using it and in what
40,1 context during the 2011 World Series. Using a data collection tool called DiscoverText,
the researchers collected 17,404 tweets. The tool was not able to collect re-tweets
and the researchers only collected tweets immediately before and after the seven game
series. They randomly selected 1/12th of the total tweets as a way of reducing the data
set to a manageable size for manual analysis. Data were then analyzed based on
28 the type of person tweeting (e.g. fan, player, coach, team representative, league
representative, media, celebrity), and the substantive nature of the tweets were
organized based on categories that included interactivity, information sharing, fanship,
promotion, and diversion. Their use of systematic sampling allowed the researchers to
work with a manageable size of data; however, their understanding of the context of the
tweets was compromised.
In a study of fan-athlete interaction during a multi-day event, Kassing and
Downloaded by UNIVERSITY OF AARHUS At 07:38 10 August 2016 (PT)

Sanderson (2010) examined the tweets posted by eight cyclists in the 2009 Giro d’Italia.
The authors tracked these tweets beginning three days before the race commenced and
ending one day following the race. The race ran for three weeks. In order to collect the
tweets sent by these riders, the researchers accessed the cyclists’ individual Twitter
pages and downloaded all of the tweets sent during the race. The content of the tweets
were analyzed inductively and themes were generated; the athletes were categorized
based on the frequency of tweeting (modest users, moderate users, or prolific/heavy
users). A significant limitation of this study was that fans’ tweets were not
simultaneously collected and analyzed to assess the degree of engagement or
interaction with the athletes’ tweets.
While Twitter is a useful source of data in the sport world, the means by which
researchers examine and analyze its use remains relatively limited (Hardin, 2014;
Hutchins, 2014; Wenner, 2014). In particular, the manual analysis process that is
commonplace in such work leads to the overuse of random sampling methods. This
results in the potential for missing important tweets that are relevant to the focus of the
study, as well as the interactions among stakeholder groups. Furthermore, the temporal
relationships to micro-events occurring during the sporting event may be lost
(Hutchins, 2014).
Methods to automate data analysis sometimes seek to apply text analytics to the
research process. One such approach is sentiment analysis, which has become a
popular research topic in computer science (Feldman, 2013). Most work in this domain
takes a supervised learning approach, using existing collections of text tagged with a
specific sentiment or emotion to generate classification schemes for new text. The
accuracy and validity of these approaches is highly dependent on the quality of the
training data, as well as the match between the general topics of the training data and
the text to be classified. Unsupervised approaches have also been explored, analyzing
the linguistic characteristics of the text in relation to language-specific words that have
been identified as bearing a specific sentiment.
Given the short and cryptic nature of the text used in Twitter, both supervised and
unsupervised approaches to sentiment analysis have limitations. However, recent
advances in the supervised approaches have been made by using tweets as the source
of the training data set (Agarwal et al., 2011; Go et al., 2009). For example,
Sentiment140 uses a maximum entropy classifier to analyze emoticons and infer the
sentiment (positive, neutral, and negative) of the associated text. After training on a
collection of 1.6 million tweets, this approach is able to achieve a classification
accuracy of over 80 percent (Go et al., 2009; Sentiment140, 2013). A similar approach
 has been employed with Chinese tweets, using four categories of sentiment ( joyful, Visual Twitter
sad, angry, disgusting) associated with emoticons, and over 3.5 million tweets to train Analytics
a naïve Bayes classifier (Zhao et al., 2012). Even amid such improvements, one of the
main drawbacks of simply performing sentiment analysis on a large collection
(Vista)
of text is that little explanation for the assignment of a particular sentiment to a given
tweet are provided.
To address this particular issue, a number of researchers have explored methods 29
for visually conveying the sentiment within a collection of tweets. A pixel-based
approach allows for the visualization of the distribution of tweets and their sentiment
over both time and geographic space (Hao et al., 2011). This allows the user to readily
interpret the temporal and geospatial distribution of the positive, neutral, and negative
tweets. A timeline may be used to illustrate the number of tweets related to a specific
set of keywords, using text processing methods to label the peaks in Twitter activity
Downloaded by UNIVERSITY OF AARHUS At 07:38 10 August 2016 (PT)

as events (Marcus et al., 2011). In this work, sentiment is encoded within coordinated
visual representations that include a geovisualization, a list of tweets, and a pie chart
of the overall sentiment. Interactive filtering allows a user to explore specific events
extracted from the Twitter feed in detail. Others have used timelines to illustrate the
changing frequency of a set of topics (Uren and Dadzie, 2015). Dashboards have
been developed to provide decision makers with a view of how their companies are
being discussed in comparison to their competitors based on industry-specific topics
(He et al., 2015).
Other Twitter visualization approaches have also been studied to enhance
researchers’ abilities to explore and analyze the temporal, relational, and other aspects
of such data. These include approaches for illustrating the social network aspect of
Twitter that emerges via re-tweets (Lotan, 2011), clustering the user and message
contents of the tweets using a self-organizing map (Cheong and Lee, 2010), and
visualizing summaries of tweets using tag clouds, organized hierarchically based
on a clustering of Twitter users (Archambault et al., 2011). Each of these approaches
seeks to provide a visual representation of the complex features of the textual data
within Twitter.
Moving beyond simply visualizing data, there has been recent movement toward
integrating automatic machine learning methods with interactive visualization
approaches, producing visual analytics systems (Keim et al., 2008). Such work takes
advantage of the powerful information processing capabilities of automatic
algorithms to address information overload and big data problems, using
visualization to convey this information to the user, and interaction to support
filtering, exploration, reasoning, and sense making. In this context, various different
machine learning approaches can be used to extract interesting features from user-
generated content such as that posted on social media, and visualization approaches
can be used to show these features to the user in a way that allows them to apply their
knowledge and analytical reasoning skills to explore and understand the data
(Schreck and Keim, 2013). A more general survey of visual text analytics approaches
has recently been conducted, highlighting the breadth of approaches that have been
explored in the literature (Alencar et al., 2012).

Vista
While visualizing the raw data extracted from Twitter can result in significant
information overload problems, performing fully automatic analyses of this same
data may isolate the analyst from the underlying meaning and small-scale features
 OIR of the data. Our goal in the design of Vista was to avoid these two extremes, and
40,1 instead take advantage of both by following a visual analytics methodology.
Combining the power of automatic methods with interactive visual representations of
the extracted features, researchers are empowered to navigate and explore among the
data, inspect the raw data and perform purposeful sampling, discover interesting
features and relationships, and apply their knowledge to make sense of what has
30 been found.
Vista is a software system created to support the interactive and visual analysis of
the temporally changing sentiment expressed in Twitter (reference withheld to
preserve blind peer review). The first step in the process is to obtain a collection of
tweets that match a user-specified query. These queries can take a number of different
forms, including general text (word, phrase), hashtags (#) representing topics, or at
tags (@) representing users. Two processes are available for retrieving the data from
Downloaded by UNIVERSITY OF AARHUS At 07:38 10 August 2016 (PT)

Twitter. If a live query is desired, the Twitter API is used to obtain the most recent
1,000 tweets (note that Twitter currently limits the number of tweets per query to 100,
and allows 450 queries per 15 minute interval). If instead the user wishes to collect a
large amount of data over an extended period of time, a long-term query may be
initiated that will monitor the live Twitter stream. In both cases, the tweets are stored in
a local database for processing.
After a collection of tweets is retrieved from Twitter, sentiment analysis is
automatically performed on the message contents of the individual tweets.
Sentiment140 was chosen for our case study due to its tweet classification accuracy
(over 80 percent) and the availability of a convenient API (Sentiment140, 2013). The
output of this process is a classification of the sentiment of each tweet in the collection
as being positive, neutral, or negative. This process of performing sentiment analysis
on the tweets works the same regardless of whether the tweets are retrieved as part of a
live search, or as part of a long-term query. In the current implementation of Vista, we
excluded non-English language tweets since Sentiment140 provides classification for
English terms and phrases only. Should another classifier be identified that performs
better than Sentiment140, or is able to classify text written in other languages, the
modular architecture of the software can allow the classifier to easily be replaced and
the non-English tweets to be retained for analysis.
Given the importance of the temporal nature of the tweets in relation to sport events,
a timeline is provided as the core visual representation. This timeline was implemented
with D3 (Bostock et al., 2011) and a number of third-party libraries. The tweets
associated with each of the three different sentiment classifications form three data sets
in the timeline (see Figure 1). Color encoding is used to differentiate between the
different sentiment classes (green represents positive sentiment; gray represents
neutral sentiment; red represents negative sentiment). Data are aggregated within
user-controlled temporal ranges (e.g. six hours, one hour, 15 minutes, or one minute). In
order to illustrate the divergent nature of the positive and negative sentiments, the
negative sentiment data are inverted in the graph.
If the Twitter users provided their locations, dots are provided on a geovisual map
for each tweet, allowing for the observation of the geospatial distribution of those who
have expressed an interest in the given topic. These dots are color-encoded based
on the sentiment of the tweets, allowing the analysts to observe the geographic
distribution of the tweets of each sentiment class.
An important use case in the design of Vista was the need to compare the data for
multiple queries. Since merging the sentiment data for multiple timelines would result
 in ambiguity and visual clutter, a parallel timeline approach is used. For each query or Visual Twitter
sub-query, a separate timeline is produced and rendered underneath the previous one, Analytics
allowing the user to scroll up and down to compare the timelines, or zoom out in their
browser to view them together (see Figure 2). The temporal scale between the multiple
(Vista)
timelines is synchronized, allowing the analyst to readily view the correlations and
patterns between the collections of tweets.
In order to enhance the understanding of the sentiment within the tweets, the 31
system automatically extracts the most commonly used hashtags, terms, user
mentions, and authors. This is done independently for each of the sentiment
Downloaded by UNIVERSITY OF AARHUS At 07:38 10 August 2016 (PT)

Figure 2.
Viewing multiple
parallel timelines
allows for the
synchronized
comparison of the
temporally changing
pattern of tweets
 OIR classifications. For each of these, a sparkline is shown, on the right hand side of the
40,1 screen, illustrating the distribution of the given item over the extent of the timeline
(see Figure 1). A visual inspection of this extracted information can provide a quick
overview of the topics related to the query, as well as a sense of how these have
changed over time.
Each of the parallel timeline visualizations can be manipulated to allow the
32 analyst to focus on an aspect of interest within the data. The simplest of these is a
sentiment filter. By clicking on the type of sentiment within the legend of the timeline
visualization, the particular sentiment data will be hidden. Doing so results in a
re-calculation of the vertical axis and a re-rendering of the remaining data. This type of
filter is useful in situations where the user is only interested in a subset of the sentiment
(e.g. comparing positive to negative, but not interested in neutral sentiments), or when
the scale of one particular type of sentiment makes it difficult to observe the patterns
Downloaded by UNIVERSITY OF AARHUS At 07:38 10 August 2016 (PT)

in the others.
Below each timeline is a compact representation of the same data, providing support
for temporal zooming. Using left and right control bars, the temporal extent of interest
can be interactively manipulated, updating the data that is shown in the timeline. This
feature allows the analyst to start with a wide temporal range of data, observe an
interesting feature or phenomena, and temporally zoom into this region for further
investigation. In cases where the user has initiated multiple queries and sub-queries,
temporal zooming in any timeline affects all others in the same way.
Within the geovisualizaiton of the tweets, standard pan and zoom operations are
possible. Zooming into a region of interest enlarges the map, reducing the overlap of
nearby dots and allowing the user to see more clearly the geodistribution of the tweets.
In addition, the geovisualization of the tweets can be used as a filtering mechanism.
Rather than linking this to the zoom operation, a rectangular region may be drawn on
the map and used as a filter, resulting in an update of the timeline to show only those
tweets from users in the selected region.
At any time after submitting an initial query, the user may choose to issue
a supplemental query that will load a parallel timeline underneath the previous
one. This feature allows for the comparison of the sentiment-based timelines for
multiple independent topics. Vista also provides the ability to generate a sub-query
within a given query. A query box is included within each timeline, providing a
mechanism to specify the sub-query. This query box can be populated automatically
using hashtags, terms, user mentions, and authors associated with the sparklines.
The results are produced in a new timeline, added below the current one. This
feature allows the analyst to explore a given aspect of the data in comparison to
the whole data set, and will feature prominently in the case study provided in the
following section.
During the exploration and evaluation of Twitter data, the researcher may identify
that the level of temporal aggregation is not sufficient for the desired analysis activity.
Selecting a different temporal aggregation can allow the observation of either a finer
level of detail (small temporal aggregation), or a coarser level of detail (large temporal
aggregation). Such a change re-produces all of the previously generated timelines at the
new level of granularity.
An important aspect of any visual analytics interface is the ability for the analyst to
drill down to the raw data; Vista is no exception. Accessing the raw data (i.e. the
individual tweets) is required in order to make sense of what has been discovered. This
is supported both within the timeline and the sparklines. Clicking on any data node in
 the timeline will bring up a modal window with a list of the tweets associated with the Visual Twitter
selected sentiment and within the selected timeframe. A similar interaction of clicking Analytics
on any sparkline will bring up a list of the tweets that make use of the term in the
context of the associated sentiment.
(Vista)
Within the tweet inspection mechanism, green, gray, and red icons are provided to
allow the analyst to override the automatically assigned sentiment. If the researcher
identifies a particular tweet for which the automatic sentiment analysis algorithm 33
makes an incorrect classification, this can be corrected with a simple click. While this
correction will only be stored for the current search session for data from a live Twitter
search, it will be saved and can be used again for future analyses when the data are
from a long-term query.

Case study: from exploration to purposeful sampling

Downloaded by UNIVERSITY OF AARHUS At 07:38 10 August 2016 (PT)

To illustrate the value and benefits of Vista for analyzing Twitter data, a case study is
provided based on the 2013 Le Tour de France. This mega-sport event was held from
June 29-July 21, 2013, with cyclists racing every day except for two rest days. The event
is broadcast globally, with a large and dedicated following. The event organizers
actively promote the use of Twitter, publicizing their own Twitter account (@letour) as
well as a specific hashtag to use when discussing the event (#tdf).
Using this official hashtag for the event, a data set of over 409,000 tweets was
collected during the three-week period of the event. In this case study, we show how
Vista can provide a visual overview of this large collection of tweets, how high-level
patterns within these tweets can be identified with respect to the positive, neutral, or
negative sentiment, and how analysts can zoom into a smaller temporal range in order
to study the patterns in greater detail. We explain how the data can be filtered with
sub-queries, either selected from the top hashtags, terms, user mentions, and authors,
or entered by the analyst based on specific interest, supporting interactive discovery
and exploration of the topics embedded within the data. This filtering and exploration
process provides a mechanism for selecting a meaningful subset of the data, which can
be inspected and exported for further analysis using other qualitative research tools
such as NVivo.
An initial visual inspection of the data (see Figure 3) shows there are many more
neutral tweets than positive or negative tweets. This is an indication of the large
number of informational tweets that are posted during a live sporting event (e.g. score
or position updates, overall event status updates, and links to images and videos of
the live action). This visual overview also shows daily spikes of tweets that are
occurring at regular points in the race, highlighting the bursty nature of the data.
Some of the spikes are higher than others indicating more tweet activities related to
specific micro-events during the race. The top hashtag per sentiment is #tdf (the
official event hashtag). After that we can see hashtags related to particular athletes
(#vavafroome for Chris Froome), the sport (#cycling), the significance of the event
(#tdf100 refers to the 100th anniversary of the event in 2013).
From Figure 3, one can readily identify that the two highest bursts occurred on the
first ( June 29) and last days ( July 21) of the event. It is reasonable to expect a significant
amount of Twitter activity on the final stage of the event, as the overall winner would
have been decided. Additionally, there was likely significant discussion of the
conclusion of the 100th running of Le Tour de France. The heightened activity on stage
one ( June 29), however, warranted further examination, since the Twitter activity
exceeded what one would expect for the start of the race.
 OIR
40,1

34

Figure 3.
Viewing all of the
tweets for the event
(including two days
before and two
Downloaded by UNIVERSITY OF AARHUS At 07:38 10 August 2016 (PT)

days after) illustrates

the temporal
distribution of the
Twitter postings
during each day
of the race

Using the zoom feature allowed us to examine the Twitter activity, in more detail,
during the first day of the race (see Figure 4). Here we can see a considerable amount of
tweeting happening around 3 p.m. (near the end of the stage). There were over 2,000
neutral posts and 422 negative posts at that time.
Zooming further into this data, we isolated a two-hour window for more detailed
inspection (see Figure 5). This was particularly interesting in that there were a
relatively high number of negative posts, suggesting a negative incident had happened

Figure 4.
Zooming into the
first day of the race
at five-minute
intervals shows a
sudden peak in
neutral and negative
posts near 3 p.m.
 Visual Twitter
Analytics
(Vista)

35

Figure 5.
Looking at a two-
hour span shows
more clearly the
Downloaded by UNIVERSITY OF AARHUS At 07:38 10 August 2016 (PT)

abnormal spike in
twitter traffic;
inspecting the tweets
shows that people
were tweeting about
a bus crash

during the event. From here we could inspect the tweets to determine the nature or
context of what people were discussing. By clicking on one of the points in the timeline,
we could see that people were tweeting about a bus (also referred to as a coach)
crashing into the finish line. Team buses travel along the route of the race, and park
near each stage’s finish line. On this day, one team’s bus hit a banner above the finish
line and got stuck with just minutes left before the first cyclist was to finish.
A sub-query of the data allowed us to explore how the official organizers (@letour)
were handling this situation (see Figure 6). This highlights the value of searching the

Figure 6.
Viewing the official
@letour tweets at
this same time
period shows the
organization’s
response to the crash
 OIR data to understand how organizations use Twitter for crisis communication (Brown
40,1 et al., 2015; Procter et al., 2013). A manual inspection of the individual tweets posted by
the official organizers showed that they provided basic, descriptive information about
the bus crash (“Bus crash changes finish”, “Finish: 3 km shorter than planned”). There
were few details related to any injuries to spectators or team staff (there were none) or
how the crash may impact the cyclists’ strategies. Further qualitative analysis could be
36 done to determine if other stakeholders (e.g. cycling teams, fans) provided different
details in their description of the incident or how they responded to the posts from the
event organizers.
Returning to the original data set, we can also use Vista to help us view how
particular topics of interest are being discussed throughout the event. For example, we
may be interested if and how issues of gender are discussed over Twitter in relation to
the race. To start, we conducted a sub-query for “women” (see Figure 7).
Downloaded by UNIVERSITY OF AARHUS At 07:38 10 August 2016 (PT)

Overall, we can see that the word “women” was not discussed regularly or in much
volume during the event. There was little discussion of women early on in the race.
Nonetheless, we did see a couple of spikes near the end of the event. We can also
examine the top hashtags associated with “women” to provide us with some context for
the tweets. For example, the hashtag #girorosa refers to a women’s cycling race held in
Italy (the women’s counterpart to the men’s Giro d’Italia). Our interest was in the large
spike of positive tweets near the end of the race.
By clicking on the timeline, we can inspect the tweets associated with the large
positive spike (see Figure 8). From this we were able to determine that many of the
tweets (and re-tweets) were a response to a call, early in the race, from The Guardian
newspaper (#Guardian) to establish a women’s race at future Le Tour de France
events. We could further inspect the data to determine the key contributors to this
discussion to understand how campaigns like this function.
We could also examine the nature of the negative tweets associated with the campaign.
Perhaps this could reveal resistance to the inclusion of women in Le Tour de France.

Figure 7.
Viewing the entire
collection of tweets,
but sub-querying for
the use of “women”
in the tweet shows
that this topic had a
number of spikes
of interest
 Visual Twitter
Analytics
(Vista)

37

Figure 8.
Inspecting the tweets
in the large positive
Downloaded by UNIVERSITY OF AARHUS At 07:38 10 August 2016 (PT)

spike shows that this

was a result of a
campaign to include
women racing in the
next Le Tour de
France

However, as shown in Figure 9, much of the negative sentiment was a call to emphasize
the Giro Donne (the name was changed to Giro Rosa in 2013). This event is scheduled
prior to Le Tour de France and thus would not compete for attention with
the men’s event.
Within this case study, Vista supported our exploration and study of the use of
Twitter among supporters, followers, and fans of the event as well as the organizers, in
a number of different ways. It provided an overview of the use of Twitter during the

Figure 9.
An inspection of the
most frequent terms
in the negative
tweets highlights a
particular opinion
and the large
number of re-tweets
of this opinion
 OIR entire event, and allowed us to identify a temporal range during which deeper analysis
40,1 could be done. Within a smaller temporal range, more detailed analyzes and
comparisons could be made.
Rather than randomly sampling the data, Vista supports exploration that can lead
to purposeful sampling. In the case study of the 2013 Le Tour de France, over 409,000
tweets were extracted, tagged based on their sentiment, aggregated, and visually
38 encoded on a timeline. While well-known features of the race are readily apparent
using Vista (e.g. the increase of tweets during each race day’s finish), micro-events
can also be studied by zooming into a smaller temporal range and/or adjusting the
level of temporal aggregation. Exploration of the data based on the sentiment
timeline, the sparklines, and the tweets themselves supported the discovery of
interesting and unexpected features within the data. Adding sub-queries allowed for
further filtering of the data, arriving at a much smaller and purposeful collection of
Downloaded by UNIVERSITY OF AARHUS At 07:38 10 August 2016 (PT)

tweets that could be examined in detail to understand some underlying phenomena

regarding the tweets.
Without the use of Vista, it would have been difficult to gain a sense of the level of
Twitter activity related to the event (e.g. number of tweets), the sentiment associated
with the event, the key terms and concepts used in the tweets, and the level of
interaction and involvement of key stakeholders. Our understanding of these
phenomena emerged as a result of the interactive exploration and analysis of the data.
While this research was conducted in the context of a mega-sport event, the approach
could be applied to other domains in which the public is willing to engage on social
media platforms, such as politics and entertainment. Furthermore, the Vista could be
modified to support other textual data sets that have a meaningful temporal
component, such as correspondence, historical references, and traditional news media.
The benefits of Vista, and other visual analytics approaches, is that they allow
researchers to access and manage big data sets, provide alternative ways to visualize
and understand textual data, and discover new insights within the data.

Conclusion
This work provides an example of the power of visual analytics methods for
supporting the study of public opinion posted on Twitter. Vista provides multiple
parallel timelines with synchronized temporal scales, and showing the sentiment of
tweets matching different analyst-specified queries, supports the comparison of many
different aspects of the Twitter data. The simple visual encoding allows for easy
separation of positive, neutral, and negative sentiment, and for immediate
interpretation of how this is changing over time. Interactive filtering, zooming,
aggregation, and inspection operations support the analyst in observing both large-
scale and small-scale patterns among the data. The data can be explored through a
dynamic and fluid interface supporting the study of a priori and emergent topics of
interest to the researcher.
The primary value of this approach is that it addresses the shortcomings of prior
work in analyzing Twitter data. In particular, rather than choosing a sampling method
based on convenience or a priori assumptions about the data or the users of Twitter,
Vista supports an exploratory analysis and dynamic filtering of a large collection of
tweets, ultimately leading to purposeful sampling that allows specific data of interest to
be analyzed further.
While the case study highlighted how Vista could be used to explore emergent
themes within a sport event, the methods are not specific to the study of fan and
 stakeholder behavior. The approach can be used more generally to explore any topic of Visual Twitter
interest that the public is engaging with on Twitter. For example, such social media Analytics
analytics could be used by a company to monitor how their brand is being discussed in
the context of known and emerging events that are relevant to the brand (e.g.
(Vista)
sponsoring events, following crises, release of new products). Such knowledge could
then be used to allow the organization to intervene in negative situations, and measure
the benefits of positive social media campaigns, for example. 39
In ongoing and future work, we are simultaneously refining and improving the
software, using it to explore interesting features of the use of Twitter in the context of
sport events, and studying other domains where knowledge of the temporally changing
public sentiment can provide a competitive advantage. We are also exploring the use of
a general classifier that can allow the user to separate the tweets based on other aspects
of interest (e.g. more complex emotional characteristics, patterns of language use).
Downloaded by UNIVERSITY OF AARHUS At 07:38 10 August 2016 (PT)

Field trial evaluations of the software to measure the usefulness of the various features
are in-progress.

Note
1. [Link]

References
Agarwal, A., Xie, B., Vovsha, I., Rambow, O. and Passonneau, R. (2011), “Sentiment analysis of
twitter data”, Proceedings of the Workshop on Languages in Social Media, pp. 30-38.
Alencar, A.B., de Oliveira, M.C.F. and Paulovich, F.V. (2012), “Seeing beyond reading: a survey on
visual text analytics”, Wiley Interdisciplinary Reviews: Data Mining and Knowledge
Discovery, Vol. 2 No. 6, pp. 476-492.
Archambault, D., Greene, D., Cunningham, P. and Hurley, N. (2011), “ThemeCrowds: multiresolution
summaries of twitter usage”, Proceedings of the International Workshop on Search and
Mining User-Generated Contents, pp. 77-84.
Blaszka, M., Burch, L.M., Frederick, E.L., Clavio, G. and Walsh, P. (2012), “#WorldSeries: an
empirical examination of a twitter hashtag during a major sporting event”, International
Journal of Sport Communication, Vol. 5 No. 4, pp. 435-453.
Bostock, M., Ogievetsky, V. and Heer, J. (2011), “D3: data driven documents”, IEEE Transactions
on Visualization and Computer Graphics, Vol. 17 No. 12, pp. 2301-2309.
Brown, N.A., Brown, K.A. and Billings, A.C. (2015), “May no act of ours bring shame: fan-enacted
crisis communication surrounding the penn state sex abuse scandal”, Communication and
Sport, Vol. 3 No. 3, pp. 288-311.
Cheong, M. and Lee, V. (2010), “A study on detecting patterns in Twitter intra-topic user and
message clustering”, Proceedings of the International Conference on Pattern Recognition,
pp. 3125-3128.
Feldman, R. (2013), “Techniques and applications for sentiment analysis”, Communications of the
ACM, Vol. 56 No. 4, pp. 82-89.
Go, A., Bhayani, R. and Huang, L. (2009), “Twitter sentiment classification using distant
supervision”, technical report, Stanford University, available at: [Link]
people/alecmgo/papers/[Link]
Hambrick, M.E., Simmons, J.M., Greenhalgh, G.P. and Greenwell, T.C. (2010), “Understanding
professional athletes’ use of Twitter: a content analysis of athlete tweets”, International
Journal of Sport Communication, Vol. 3 No. 4, pp. 454-471.
 OIR Hao, M., Rohrdantz, C., Janetzko, H., Dayal, U., Keim, D.A., Haug, L.-E. and Hsu, M.-C. (2011),
“Visual sentiment analysis on twitter data streams”, Proceedings of the IEEE Conference on
40,1 Visual Analytics Science and Technology, pp. 277-278.
Hardin, M. (2014), “Moving beyond description: putting twitter in (theoretical) context”,
Communication & Sport, Vol. 2 No. 2, pp. 113-116.
He, W., Wu, H., Yan, G., Akula, V. and Shen, J. (2015), “A novel social media competitive
40 analytics framework with sentiment benchmarks”, Information & Management, Vol. 52
No. 7, pp. 801-812.
Hoeber, O., Hoeber, L., Wood, L., Snelgrove, R., Hugel, I. and Wagner, D. (2013), “Visual Twitter
analytics: exploring fan and organizer sentiment during Le Tour de France”, Proceedings
of the VIS 2013 Workshop on Sports Data Visualization, pp. 1-7.
Hutchins, B. (2014), “Twitter: follow the money and look beyond sports”, Communication & Sport,
Downloaded by UNIVERSITY OF AARHUS At 07:38 10 August 2016 (PT)

Vol. 2 No. 2, pp. 122-126.

Kassing, J.W. and Sanderson, J. (2010), “Tweeting through the Giro: a case study of fan-athlete
interaction on Twitter”, International Journal of Sport Communication, Vol. 3 No. 1, pp. 113-128.
Keim, D.A., Andrienko, G., Fekete, J.-D., Görg, C., Kohlhammer, J. and Melançon, G. (2008), “Visual
analytics: definition, process, and challenges”, in Kerren, A., Stasko, J.T., Fekete, J.-D. and
North, C. (Eds), Information Visualization: Human-Centered Issues and Perspectives, LNCS
4950, Springer, pp. 154-175.
Lotan, G. (2011), “Mapping information flows on Twitter”, Proceedings of the ICWSM Workshop
on the Future of the Social Web, pp. 23-27.
Mahrt, M. and Scharkow, M. (2013), “The value of big data in digital media research”, Journal of
Broadcasting & Electronic Media, Vol. 57 No. 1, pp. 20-33.
Marcus, A., Bernstein, M.S., Badar, O., Karger, D.R., Madden, S. and Miller, R.C. (2011), “TwitInfo:
aggregating and visualizing microblogs for event exploration”, Proceedings of the SIGCHI
Conference on Human Factors in Computing Systems, pp. 227-236.
Pegoraro, A. (2010), “Look who’s talking - athletes on twitter: a case study”, International Journal
of Sport Communication, Vol. 3 No. 4, pp. 501-514.
Procter, R., Vis, F. and Voss, A. (2013), “Reading the riots on twitter: methodological innovation
for the analysis of big data”, International Journal of Social Research Methodology, Vol. 16
No. 3, pp. 197-214.
Russom, P. (2011), “Big data analytics”, TDWI research, Renton, WA.
Schreck, T. and Keim, D. (2013), “Visual analysis of social media data”, IEEE Computer, Vol. 46
No. 5, pp. 68-75.
Sentiment140 (2013), “Sentiment140: a {T}witter sentiment analysis tool”, July, available at:
[Link]
Shneiderman, B. and Plaisant, C. (2010), Designing the User Interface, Designing the User
Interface, 5th ed., Addison Wesley, Boston, MA.
Thomas, J.J. and Cook, K.A. (2006), “A visual analytics agenda”, IEEE Computer Graphics and
Applications, Vol. 26 No. 1, pp. 10-13.
Tinati, R., Halford, S., Carr, L. and Pope, C. (2014), “Big data: methodological challenges and
approaches for sociological analysis”, Sociology, Vol. 48 No. 4, pp. 663-681.
Uren, V. and Dadzie, A.-S. (2015), “Article information”, Aslib Journal of Information
Management, Vol. 67 No. 3, pp. 337-355.
Ward, M.O., Grinstein, G. and Keim, D. (2010), Interactive Data Visualization: Foundations,
Techniques, and Applications, Interactive Data Visualization, A K Peters, Ltd, Natick, MA.
 Ware, C. (2013), Information Visualization: Perception for Design, Information Visualization, Visual Twitter
3rd ed., Morgan Kaufmann, Waltham, MA.
Analytics
Wenner, L.A. (2014), “Much ado (or not) about Twitter? Assessing an emergent communication
and sport research agenda”, Communication & Sport, Vol. 2 No. 2, pp. 103-106.
(Vista)
Zhao, J., Dong, L., Wu, J. and Xu, K. (2012), “MoodLens: an emoticon-based sentiment analysis
system for Chinese tweets in Weibo”, Proceedings of the ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, pp. 1528-1531. 41
About the authors
Dr Orland Hoeber is an Associate Professor in the Department of Computer Science at the
University of Regina (Canada) and an Adjunct Professor at the Memorial University of
Newfoundland (Canada). His primary research interests include information visualization, visual
and geovisual analytics, web search interfaces, mobile computing, human-computer interaction,
and sport analytics. Dr Orland Hoeber is the corresponding author and can be contacted at:
Downloaded by UNIVERSITY OF AARHUS At 07:38 10 August 2016 (PT)

[Link]@[Link]
Dr Larena Hoeber is an Associate Professor in the Faculty of Kinesiology and Health Studies
at the University of Regina (Canada). Her primary research interests are in sport management,
volunteer management, gender and sport, and qualitative research methods.
Maha El Meseery is a MSc Student in the Department of Computer Science at the University
of Regina. Working under the supervision of Dr Orland Hoeber, her primary research interests
are in geovisual analytics and exploratory data analysis.
Kenneth Odoh is a MSc Student in the Department of Computer Science at the University of
Regina. Working under the co-supervision of Dr Orland Hoeber, his primary research interests
are in visual text analytics, information dashboard design, and automatic trend analysis.
Radhika Gopi is a MSc Student in the Department of Computer Science at the University of
Regina. Working under the supervision of Dr Orland Hoeber, her primary research interests are
in social media analytics and mobile computing.

For instructions on how to order reprints of this article, please visit our website:
[Link]/licensing/[Link]
Or contact us for further details: permissions@[Link]