In this chapter, the development of the Social Set Visualizer was detailed. It in-troduced the development objectives of application performance, fault tolerance, and maintainability that guide the development of the software tool. Furthermore, it de-scribed the technological foundations, and detailed the visualization framework. For the problem of data storage, a thorough evaluation of NoSQL, relational, distributed, key-value, graph-based, and custom databases was made, which resulted in the se-lection of PostgreSQL, a relational database, for the Social Set Visualizer. Then, software architecture on frontend and backend was outlined giving insight into the client-side and server-side components. After this, the three iterations on the IT artifact of this PhD project were introduced. Set-up and features of the first, sec-ond, and third version of the Social Set Visualizer are described in detail. Lastly, the deployment on a single machine without use of networking or virtualization was outlined.
Chapter 5
Evaluation
This chapter presents an extensive evaluation of the Social Set Visualizer. Multiple case studies have been performed on a wide variety of topics using large-scale Face-bok datasets. In these, the Social Set Visualizer software tool has been utilized by other researchers and practitioners, and as a result contributed to more than seven peer-reviewed publications on the theme of descriptive and predictive analytics.
5.1 Descriptive Case Studies
The Social Set Visualizer software tool was evaluated through four case studies in descriptive analyticsof Big Social Data. These case studies concern the Bangladesh factory disasters in the field of corporate social responsiblity, sports broadcasting for the two television stations TV2 Sport and NRK Sporten, Facebook activity across three music festivals Roskilde, Glastonbury, and Burningman, and lastly the emission scandal of Volkswagen.
5.1.1 Bangladesh Factory Disasters (2015)
The Social Set Visualizer case study on the topic of visualizing social media reactions upon specific events in the garment industry factory disasters in Bangladesh has been published in [Flesch et al. 2015b]. The study follows the Social Set Analysis approach and combines set-theoretical analysis of actors in Big Social Data with an Event Study Methodology in order to provide insights into large-scale events and their various facets.
In this case study, several research questions regarding the social media reactions to the Bangladesh factory disasters expressed on the Facebook walls of international clothing retailers were investigated.
The first version of the Social Set Visualizer was used for this case study. The user interface is illustrated in Figure 5.1. It allows selection of a time period either by using the date range fields in the top left, or by using the timeline selection tool which is marked as [2] in the graphic. Furthermore, selection of Facebook walls for analysis is possible through the color-coded input field at the very top of the screenshot.
Overall, the user interface contains six different visualizations which are annotated with their respective numbers:
(1) Facebook activity chart. It gives a detailed overview about the Facebook activity and real-world events in the selected timeframe. Based on the detail on demand
68 Chapter 5. Evaluation
Figure 5.1: SoSeVi 1 used in case study on the Bangladesh factory disasters [Flesch et al. 2015b]
principle, different zoom levels show additional pieces of information. First, the aggregate number of data points for each Facebook wall is shown, then, each Facebook post is shown with the number of comments, and lastly, every individual post and comment is visualized.
(2) Timeline selection tool. It visualizes the timeline of real-world events on top of a line chart which displays the aggregate volume of Facebook activity. Its two brushes on the left and right side allow the user to select a timeframe by using drag and drop. Scrolling with the mouse wheel while hovering this visualization zooms in and out respectively.
(3) Word Cloud. An alphabetically sorted word cloud visualizes the top 200 most frequently occuring words from Facebook posts and comments. It is dynamically recalculated as the user interacts with the timeline selection tool (2) or changes the selection of Facebook walls.
(4) Set intersections between Facebook walls. This set-based visualization illus-trates the mobility of social media actors between different Facebook walls. A set of all actors on each Facebook wall is calculated, then set intersections are visualized.
(5) Set intersections between before, during, and after periods. This set-based visualization illustrates the mobility of social media actors across the time di-mensions. Sets and intersections for before, during, and after time periods are visualized via an “exploded” Venn diagram.
5.1. Descriptive Case Studies 69 (6) Language Distribution. Based on natural language processing, a language clas-sification of each post and comment within the selected time period is visualized.
For this case study, the Facebook walls of Bennetton, Calvin Klein, Carrefour, El Corte Ingles, H&M, JC Penny, Mango, Primark, PvH, Walmart, and Zara have been analyzed. Selected findings of this case study are:
(a) Global supply chain concerns with regard to Bangladesh garment factories have been expressed by Facebook users from as far back as 2009.
(b) There are many instances of authentic displays of support and expressions of empathy from Facebook users as well as robotic incidents of ‘slacktivism’.
(c) Many of the uses of the word “please” were in relation to opening requests for new stores in the case of H&M.
(d) Protestors and activists employed different social media strategies on the different Facebook walls of companies but with little evidence for social influence (in terms of the number of likes and comments on their posts).
(e) Similarly, companies followed not only different corporate social responsibility strategies but also different social media strategies before, during and after the Bangladesh garment factory accidents.
(f) For almost all of the accidents, a majority of the users posting during the news-cycle do not return to the Facebook walls again. That is, social media engagement during factory accidents is detected as episodic and ‘bursty’ with little overlap to the “business-as-usual” period before or after the accident.
In this case study, the full archive of the social data from the Facebook walls of the 11 clothing retail companies was extracted using the Social Data Analytics Tool [Hus-sain & Vatrapu 2014b]. Given this data basis, I designed, developed and evaluated the Social Set Visualizer dashboard on this corpus of Big Social Data consisting of approximately 90 million data points. This case study analyzed social media data and presentated significant findings on the topic of corporate social responsibility.
The set-based approach of the Social Set Visualizer software tool enabled a com-parative, holistic study based on social media data from relevant companies in the clothing retail industry. Furthermore, set-based visualizations were utilized for visual analytics of various real-world events and social media reactions of an international audience.
5.1.2 Sports Broadcasting by TV2 and NRK (2016)
A comparative case study between two major Scandinavian sports broadcasters was carried out using the Social Set Visualizer. In this case study the Facebook pages of TV2 Sporten, a Danish TV station, and NRK Sport, a Norwegian TV station, were analyzed to give insight into differences in terms of audience reactions to important
70 Chapter 5. Evaluation
Figure 5.2: Temporal distribution of total Facebook activities for NRK Sport and TV2 Sporten (SoSeVi 1) [Henniget al.2016]
sports events. The case study has been published as a student project in the Big Data Analytics course, to which the Social Set Visualizer tool was provided [Hennig et al. 2016].
Figure 5.2 showcases the temporal distribution of Facebook activity for both TV stations in the Social Set Visualizer dashboard. The visualization enables researchers to identify event-related spikes, seasonal variations, and demographic patterns in the data at hand. Furthermore, it provides proof for large-scale temporal patterns, as data volume is steadily increasing over the two-year observation window of the study.
Subsequently, the comparison of activity numbers during live broadcasting and re-run programs showed significantly more Facebook activity during live broadcasting.
Figure 5.3showcases the utilization of Social Set Visualizer to illustrate overlaps between the Facebook pages of TV2 Sporten and NRK Sport. Additional findings of this case study include the quantification of sentiment distribution between “good”
and “bad” sports events. Furthermore, the researchers were able to discover that the distribution of emotional patterns expressed through Facebook reactions changed over time. The development of positive, negative and neutral sentiment activity on the Facebook pages during the olympic games was analyzed using the Social Set Visualizer. It was found that TV ratings are significantly affected by sports and gender, as “single events showing male biathlon athletes have more viewers, but female biathlon drew more viewers than male biathlon in total”.
In this case study, the first version of the Social Set Visualizer software tool was used for Visual Analytics of Big Social Data on a dataset of 7,532,000 pieces of Facebook activity. Its set-based approach to Visual Analytics enabled the student
5.1. Descriptive Case Studies 71
Figure 5.3: Unique Facebook actors during complete event window on TV2 Sporten (SoSeVi 1) [Henniget al.2016]
researchers to quantify overlaps in social media activity. This was done by calculating and visualizing set intersections of the before, during and after time periods of real-world, TV-broadcasted sports events.
With regard to the research question of this thesis, this case study shows that the utilization of a set-based approach to Big Social Data Analytics generates unique and relevant insights. Furthermore, student researchers were quickly able to use the Social Set Visualizer software tool to produce research findings that were presented to an international audience at the 2016 IEEE Conference on Big Data.
5.1.3 Roskilde, Glastonbury & Burningman Festivals (2016)
A comparative case study using the Social Set Visualizer was conducted on the topic of festival analytics. In this study, the Facebook activity of three world-famous music festivals, namely Roskilde Festival (Denmark), Glastonbury (UK), and Burningman (USA) is analyzed using our set-based approach to Big Social Data Analytics.
This case study is attached as Publication III [Fleschet al. 2016] to this disser-tation. It was created using the second version of the Social Set Visualizer software tool. Methodology and findings were presented to an international audience of set visualization experts at the Set Visualization and Reasoning workshop (SetVR) at Diagrams conference 2016.
The case study showcases that all three festivals have interactions with several hundreds of thousands actors on Facebook. Figure 5.4illustrates the set intersections that were performed with the Social Set Visualizer on this dataset. The visualization shows cardinalities of various large-scale sets with between 200 and 500,000 actors each. For each festival, sets for the user-selected before, during and after periods are calculated. Then, pairwise intersections between all sets are performed, and their cardinality is visualized. The set intersections between the three festivals Roskilde, Glastonbury and Burningman emphasize the fact that Glastonbury and Roskilde, but also Glastonbury and Burningman share a significant amount of users. However,
72 Chapter 5. Evaluation
Figure 5.4: Visualization of set intersections and set intersection cardinality before, during, and after the user-selected time period, illustrating the distribution of social media actors over time and space. Publication III [Fleschet al.2016]
the intersection between Burningman and Roskilde is very small. Furthermore, the number of users who are active on all three festival pages during the examination period is very low, only around one thousand users.
In this evaluational case study, the Social Set Visualizer software tool has used data collected from three different Facebook walls with massive user bases, totaling more than 10M data points. Therefore, this case study demonstrates the feasibility of large-scale set intersections in the interactive visual analytics dashboard based on dynamic user-driven selection of investigation timeframes, and thus presenting an example case for answering the first research question.
5.1.4 Volkswagen Emission Scandal (2016)
The Social Set Visualizer was applied to the Facebook walls of Volkswagen in an unpublished case study on the Volkswagen emission scandal. Figure 5.5 depicts a screenshot of the set-based visualizations in this case study.
The capabilities of Social Set Analysis are showcased through a visualization of set intersections between the different Volkswagen entities and a visualization of migration patterns between each set intersection. The Social Set Analysis method-ology was demonstrated by comparisons of actor movements across dimensions of time and space in order to visualize social media migration patterns.
The findings of this study emphasize the fact that dominant discussion topics can quickly change when a scandal emerges from news reports. Once reports on the emission scandal were published, the user activity on the Facebook walls of four
5.1. Descriptive Case Studies 73 corporate entities of Volkswagen displayed very similar patterns.
In this case study, it was shown that the Social Set Visualizer software tool is able to perform large-scale set-based Big Social Data Analytics in order to compare different country pages of the Volkswagen AG. The dataset consisting of 8M Facebook interactions from four corporate entities in three countries, namely UK, USA and Germany, outlines the utility of the Social Set Visualizer in comparative event studies of Big Social Data.
With regard to the first research question of this thesis, the second version of the Social Set Visualizer as utilized in this case study implements a novel approach to set-based visualizations inspired by UpSet [Lex et al. 2014]. Through this, the visualization quantifies social media migration patterns which happen before, during and after the user-selected time periods in relation to important real-world events.
This unique approach to Big Social Data Analytics allows an efficient comparison of four different social media presences of Volkswagen AG in light of the emission scandal.
Figure 5.5: SoSeVi 2 displaying 8M interactions from the Volkswagen AG Facebook pages in a study on the emission scandal. [Fleschet al.2016]
74 Chapter 5. Evaluation