group 4.5

Group Members:

Anete E.

Emma G.

Jil M.

Jonathan M.

Lara R.

Week 1

“Blue Feed/Red Feed” (06.03.2019)

On May 18, 2016, Jon Keegan published a presentation on The Wall Street Journal called “Blue Feed/ Red Feed”. Two feeds are displayed on one page, side-by-side and are updated hourly. The blue feed on the left presents ‘very liberal’ posts from sources published on Facebook, whereas ‘very conservative’ is displayed on the right, covering topics such as “President Trump”, “Health Care”, “Guns”, “Abortion”, “Immigration”, “ISIS”, “Executive order” and “Budget”. The project is based on a study called “Exposure to ideologically diverse news and opinion on Facebook” that was previously conducted by the scientists Bakshy, Eyta and Messing, Solomon in 2015 in which they categorized numerous posts and analyzed the exposure of users to ideologically determined Facebook news.

Screenshot (3)

The researchers tracked and analyzed the top 500 shared sources as well as the content of 10.1 million Facebook user’s feeds, who themselves indicated their political views on their profile. From there, a political ‘alignment score’ is calculated for each article, in order to determine its political nature as well as the larger category that it will be put in, ranging from ‘very liberal’ to ‘very conservative’.

Hence, the data used for this project are these numerous private Facebook profiles as well as the findings of the 2015 Facebook research paper. In short, the project was conducted in order to expose how reality, in this case, a user’s Facebook feed, might appear drastically different for distinct users depending on their political views, due to potential filter bubbles or echo chambers. Consequently, this can result in users not keeping an open mind, which is essential for qualitative political debates.

The researchers used the software development tool Graph API, which is the primary tool to pragmatically extract information from Facebook. This also ensures that the content relevant to the research is being pulled from Facebook automatically. However, the software developer tools and the ‘alignment score’ employed by the researchers in order to determine which posts finally appear on the two feeds are not explained in detail and need to be further researched by the reader.

Despite the positive sides of this project, one can detect several possible limitations as well as potential improvements that could make it even more beneficial. For instance, the researchers set the terms that sources must have at least 100.000 followers and the included posts must have been shared at least 100 times by Facebook users, which excludes a lot of sources that still might have a potential influence on user’s political opinion. In addition, the fact that they used Facebook users’ self-described political orientation as a base for the source’s political value might lead to distortions and impreciseness. Furthermore, the project excludes the sources and websites that have been shared by users from in a broader political spectrum, including The Wall Street Journal, but also social media platforms like Twitter and Youtube, which might lead to a rather extreme simulation of a Facebook feed. Finally, one could argue that the five categories from ‘very liberal’ to ‘very conservative’ lead to a limited and oversimplified classification, especially considering that in the final project these are further narrowed down to two feeds. A possible proposition for the project would be to include Facebook users’ opinions in the form of surveys or opinion polls in order to examine whether they are aware that they scroll within a filter bubble or not.

 

To conclude, the project demonstrates that by indicating personal political views on the Facebook profile, it will influence what kind of posts and sources will appear on one’s News Feed. Therefore, the delivery of the project resulted in a successful visualization of both feeds side-by-side. Nevertheless, there are other aspects of the project and the research that are questionable and/or could have been done differently. For instance, the technique of collecting and categorizing data appeared to be limiting. Namely, certain posts and sources, who still could have a strong influence on one’s political opinion, were excluded from the research, because they did not fit within the established data collection framework. Additionally, to collect the data from Facebook, the software development tool Graph API was used; however, the execution and application of the tool were not thoroughly explained to the reader. Consequently, it was challenging to learn about the process of the project’s realization. Nevertheless, the project still succeeded at bringing attention and informing Facebook users of the threat of echo chambers and filter bubbles.

References

Bakshy,Eytan; Messing, Solomon, and Adamic, Lada A. 2015. “Exposure to ideologically diverse news and opinion on Facebook”. Science. https://education.biu.ac.il/files/education/shared/science-2015-bakshy-1130-2.pdf

Keegan, Jon. 2016. “Blue Feed, Red Feed”. WSJ. http://graphics.wsj.com/blue-feed-red-feed/#methodology.

 

Week 2

Facebook tracking exposed: Investigating algorithms (13.03.2019)

What did we do?

For this week’s experiment, we created a fake Facebook profile with the name Diego de Jong, who is a middle-aged man that has a Colombian mother and a Dutch father, who was raised in Bogota and currently lives in Amsterdam. We created his persona with the ongoing migrant crisis in Colombia in mind, in which thousands of people migrate from Venezuela to Colombia due to horrible inflation rates and other issues within the country.

What did we intend to show?

Our intended goal was to have the Facebook algorithm, that is responsible for showing recommended content and tailored advertisements, pick up on the anti-feminist and conservative political views of our profile. Those political views were reinforced due to the fact that he ‘liked’ several conservative newspapers, the current president, who has a more right-oriented political agenda, as well as sharing and commenting on numerous articles from those aforementioned sources.

How did we do it?

We conducted this experiment for seven days, on which we shared, liked and commented on multiple posts at least five times a day and kept a diary of each day, summarizing the interactions. In addition, we tried to make the profile look as authentic and real as possible, by continuously developing Diego’s character through posting pictures and updates, reacting to posts and engaging with articles and friends, as well as fill out every detail about him in his profile, to which Facebook (or rather the Browser extension facebook. tracking. exposed we used for our experiment) sent us this notification;

This notification is interesting as it admits to tracking our personal information and uses it to create tailored content right away! Furthermore, for the sake of the experiment, we accepted every single friend request sent to us and tried to engage with our new Facebook friends as much as possible, for instance by reacting to their posts and answering to a few messages that we received.

What did we observe?

After seven days of active engagement on the site every day, we noticed some peculiarities in means of what was appearing on the profile’s News Feed. The top posts and sources that appeared were those with whom we had engaged before and ones that were similar in content compared to the previously ‘liked’ or ‘shared’ posts.

For instance, on the second and third day, we engaged with some right-wing sources and posts by ‘liking’ and/or ‘sharing’ them. Consequently, when scrolling through the feed on the fourth day, we encountered a lot of sources that covered several political topics from a right-oriented viewpoint. Additionally, since we indicated that our character is from Colombia on the profile and engaged with content that covered the countries news stories, almost all of the posts that appear on the News Feed are in Spanish and about the internal politics of the country. This accentuates how one’s country of origin plays an important role in generating a filter bubble, which is an algorithm used on Facebook that makes it easier for users to engage with like-minded people. To compare, our News Feed also differentiates even though we all live in the Netherlands, because of our diverse origins and languages we speak.

Furthermore, in line with our character’s explicit conservative and hostile views on the migrant crisis in Colombia, we observed a gradually evolving pattern in the recommended articles on the Facebook news feed. Hence, the latter often showed news stories about international conflicts between people with a migration background and locals, even from news outlets which are normally considered as propagating rather neutral political ideas. However, most of the news stories on the feed with this kind of topic are shared by newspapers we liked that have a more or less biased view on certain topics such as migration in this case.

To sum up the experiment; we observed that the comments and reactions of other users to those type of posts often coincide with our character’s views and political opinions. We can also see that the news posts we shared or liked are being fed back to us, which might indicate that a filter bubble has already built up in the short duration of this experiment. At the end of the experiment, we accumulated 342 friends in total and were messaged by some of them in different languages.

 

Week 3

Facebook’s Filter Bubble Exposed

by Lara Rittmeier, Emma Gasparin, Anete Ezera, Jonathan Matalon, Jil Meyers

 

An overview of the fake Facebook profile we created

 

AMSTERDAM, March 19 2019. In the contemporary digital landscape, ‘filter bubble’ or ‘echo chamber’ are well-known terms that represent a scenario in which a user is enclosed in a digital ‘bubble’ with like-minded users and posts echoing their beliefs and values rather than challenging them. These terms are mostly associated with Facebook or other social media platforms or news outlets feeding the content you have been engaging with back to you. To examine the workings of a ‘filter bubble’ and unfold the influence of algorithms, we set up a fake Facebook profile in the name of Diego de Jong. We indicated his origin, his political interests, his profession and other details about his personal life in order to observe if and how Facebook’s News Feed differs from our own News Feed. In addition, we intended to keep track of the kind of content that would be shown and prioritized in the course of the experiment. We conducted this project in strong relation with our persona’s views on the ongoing migration crisis in Colombia, which is therefore also the imagined country of birth in our created profile.

After one week, we collected the data and gathered our observations of the peculiarities we noticed about the content on the user’s News Feed. Namely, similar posts and political viewpoints which were ‘liked’ or ‘shared’ kept appearing on the News Feed, alongside with posts of the same site of one particular newspaper, called El Nuevo Siglo. Also, the user’s origin was a noteworthy factor, which determined the focus in terms of the location of political news and used language. However, we discovered other significant patterns after gathering and analyzing the data from Facebook Data Extractor. Further, we created a visual presentation of the data:

 

A visual representation (circle packing) of the data

 

This visualization displays the difference in the type of content that was distributed on the News Feed of the fake profile. As it shows, posts (50,92%) are the greatest portion of content that appears on the News Feed. Photos follow with 27,20% and videos with 21,88%. However, there were no events displayed on the News Feed, which makes sense as we did not engage with any events during the time of the experiment. Moreover, as we mostly engaged with posts, it corresponds to the high percentage of posts appearing on the News Feed.

Examples of posts on Diego’s News Feed

 

To continue, most of the posts we engaged with were from a particular news outlet El Nuevo Siglo which correlates with Diego’s strong political stance on anti-immigration matters. The latter is known for spreading rather conservative political news, which is why we chose it as the main news outlet for our persona to engage with on a regular basis. With that in mind, we noticed that other recommended sources and posts appearing on the News Feed had a very similar narrative. When browsing the News Feed, numerous posts are by the news outlet El Nuevo Siglo which was to expect regarding our constant engagement with the source. Moreover, as we have engaged with quite a few posts about the Colombian football team Millonarios F.C., as well as with the Dutch club AFC Ajax which the algorithm picked up on and subsequently continued displaying stories about football on the Feed. Furthermore, it is worth mentioning that we explicitly stated Diego’s passion for football and the Millonarios in the ‘Intro’ section of his profile.

 All in all, we argue that a vast number of posts appearing on the Feed are closely related to the information that we were feeding the algorithm with. In other words, Facebook provided mostly stories about political incidents in Colombia, the two above mentioned football clubs, as well as miscellaneous posts about either Diego’s home country or the Netherlands, which paints a more or less accurate picture of the persona we created. However, we also observe that the data and the general outcome of this project are not as explicit and clear in order to draw definitive conclusions, but that might change in the future development of this experiment.

 Final report:

Facebook’s Filter Bubble Exposed

 

Image 1

 

by Lara Rittmeier, Emma Gasparin, Anete Ezera, Jonathan Matalon, and Jil Meyers

 

Word count: 2258

 

AMSTERDAM, April 1st 2019. In the contemporary digital landscape, algorithmic personalization, ‘filter bubbles’ and ‘echo chambers’ have acquired a central prominence due to their influence on both the content and the user of social media platforms. More explicitly, those terms represent a scenario in which a user is enclosed in a personalized digital ‘bubble’ with apparent like-minded users and posts which tend to echo their beliefs and values rather than challenging them. As contemporary research shows, this increase can have a far-reaching influence on the formation of political views, undermine the diversity of opinions and hence contribute to an overall distortion of reality. (https://bit.ly/2TQ1TPH)

Thus, our aim is to examine the potential formation and working of a ‘filter bubble’ and unfold the influence of algorithms on the average Facebook user, depending on the kind of content that is engaged with.  In other words, we explore to what extent Facebook’s algorithms exclusively lead to more tailored content and how the digital presence of our created persona will develop during the course of the experiment.

We conducted this experiment for two weeks, using the open source web browser extension fbTREX (Facebook Tracking Exposed), designed in 2016 with the growing influence of Facebook as a source for news stories. Moreover, the tool is intended for Facebook users, researchers and journalists who are interested in the collection of data as well as the analysis of filter bubbles and algorithms on Facebook. By collecting and analyzing the content on user’s News Feeds, the developers of this add-on claim that they intend to ‘increase transparency behind personalization algorithms’ and open up Facebook’s black box in order to increase the awareness of the average Facebook user. (https://facebook.tracking.exposed/)

In order to initiate the experiment, we set up a fake Facebook profile of a middle-aged Colombian, born in Bogota and living in Amsterdam who goes by the name Diego de Jong. First of all, we indicated his origin, his political interests, and his profession, followed by other details about his personal life in order to make him appear authentic; he works as a gardener, is strongly Catholic, married, enjoys fishing and is a big fan of the Colombian football team Millonarios and Amsterdam football team AFC AJAX , as well as of the current Colombian president Ivan Duque.

 

Image 2: Overview of the created Facebook profile

Moreover, in order to facilitate the tracking of the content shown and prioritized in the course of the experiment, we conducted this project in strong relation to the ongoing migration crisis in Colombia. The crisis concerns thousands of people who migrate from Venezuela to Colombia due to drastic inflation rates and other issues within the country, with the objective to live in temporary tent villages, of which the largest amount is placed in Bogotá, the capital of Colombia and also his place of birth. More specifically, we depicted our persona as a strongly conservative and right-wing oriented man, in order to increase the difference between the suggested content of the fake profile and our own personal views. In addition, we argue that the focus on and the engagement with one specific narrative makes it more feasible to scrutinize and track the development of the News Feed over time.
During the experiment, we first installed the aforementioned browser extension “Facebook Tracking Exposed”, disabled our Adblock and then shared, liked and commented on multiple posts at least five times a day for seven days in a row, while keeping a diary of each day, summarizing the interactions. The majority of the posts and articles we engaged with were from a few particular news outlets; El Colombiano and  El Nuevo Siglo, which correlate with Diego’s strong political stance on anti-immigration matters. The latter is known for spreading politically conservative news in Colombia (see https://www.bbc.com/news/world-latin-america-19390073), which is why we chose it as the main news outlet to engage with on a regular basis. In addition, we made the profile look as authentic as possible by continuously developing Diego’s character through posting pictures and updates, reacting to posts and engaging with articles and befriended users that send us a friend request. For instance, we deliberately searched for articles who spread anti-immigration or anti-feminist narratives and shared, commented or liked them in order for the algorithm to pick up on Diego’s politically conservative views.

Image 3: Examples of politically conservative articles we engaged with (sources: El Nuevo Siglo, El Tiempo)

 

Furthermore, we indicated his strong passion for the Colombian football team Millonarios F.C., as well as the Dutch club AFC Ajax through the liking of sites and engagement with posts. In addition to that, we made sure to only post and comment in Spanish, so that our profile appears more authentic, which consequently resulted in Spanish being the main language of the suggested content and posts that appeared in our timeline, followed by Dutch, as the Netherlands are his current indicated residence.

After several days, the content displayed on Diego’s News Feed started to take a certain shape and we noticed several interesting patterns. Firstly, as we did accept every friend request for the sake of the experiment, rather unexpectedly we ended up by making about 350 friends in the first week. However, we also observed that a lot of the “friends” appeared to be fake profiles as well, of which almost all had minimum content or information on their profiles and were practically all from an African origin. Consequently, at the end of the week, a lot of content appearing on the News Feed consisted of photos posted by the friends we accepted. Yet, despite the many friends we acquired and their presence on the News Feed, hardly any of them reacted to our posts or extensively engaged with our persona.

As a result, the suspicion that a majority of the friend requests we received were fake profiles, as well as the fact that most of them had an African background, made us think about various reasons behind this observation; potentially, it could mean that Facebook is recommending new user profiles to other new or fake profiles, so that a new profile can either gain a lot of friends fast or connect with similar profiles (e.g. fake profiles). However, this would mean that Facebook deliberately tolerates the presence of fake profiles which is rather questionable. With that in mind, it could be argued that Facebook does have a certain mechanism to identify fake profiles, and then connect them to other ones thus creating a ‘filter bubble’.

At the beginning of the experiment, we expected to have some friends with the same political views; however, the requests we got were rather random. Therefore, it made us think about other reasons why Facebook would recommend to connect to our profile. When looking at the friend’s profiles more closely, we noticed that they also acquired friends really fast and seemed to send requests to anybody (some of them have 5000+ friends). Additionally, the friend requests we got were from mutual friends we already had accepted. With all that in mind, it became clear, that Facebook does not recommend user profiles to other users based on similar political views or interests. Nevertheless, that does imply that Facebook pursues to connect new profiles to users, which have a behavioural pattern to ‘connect to everyone’ and gain a lot of friends fast. However, the reasons behind such recommendation systems could serve as a question for future analysis. To really uncover how Facebook’s algorithm works a more detailed profile, as well as a longer period of time, would be required.

 

Figure 1. Visual representation in percentages of the content on the News Feed (, divided into ‘photos’, ‘videos’ and ‘posts’

 

Subsequently, as the content  posted by our Facebook friends unexpectedly started to take up a major part of the News Feed (see figure 1), it distorted the actual purpose of the research, which is why decided to ‘unfollow’ all of them, in order to see what content was being recommended beyond the posts of our friends.

To visualize the number of videos, posts and photos that were shown in our Facebook feed, we used fbtrex to analyze the content, downloaded the data as an scv. file and added it into Microsoft Excel. Using formulas, we let Excel calculate the number of photos, videos and posts and the percentage of each one. We added the data into Rawgraphs.io to create a pie chart, in order to visualize the proportion of content on our Facebook feed. On the graph, we can see that 55.1% (357) of content in our newsfeed were posts, 23.5% (155) photos and 21.4% (140) videos, out of 652 recorded posts on the timeline. There are no events or groups included since we didn’t join any group or reacted to any kind of event.

 

Moreover, besides the content of Diego’s friends on the News Feed, there were additional observations that we made in terms of the content appearing in his  News Feed. Namely, posts with a politically conservative viewpoint, which were ‘liked’ or ‘shared’ kept appearing on the News Feed, alongside with posts of the same site of the above-mentioned news outlet El Nuevo Siglo, with which we engaged with a lot, as well as a lot of football-related content.

Now, after more than two weeks and hence the end of the experiment, Diego’s Facebook feed still looks very similar to our initial observations, which was to be expected as we did not alter his likes or changed his viewpoints.

Image 4: Screenshot of selected posts on the feed of the fake profile after the end of the experiment

 

During the duration of this experiment, we experienced two main challenges; firstly we have to question the ethics of this experiment, as we had to write openly anti-feminist and nationalistic posts and comments under public accounts, which did not feel right as we do not share those viewpoints in the slightest. Consequently, we will delete those comments and the entire profile right after the end of this experiment to ensure that our comments cannot hurt anyone that might potentially read them. However, this issue represents an important dimension of research projects and experiments, which also implies limits. For instance, since our online persona was created for the sake of research purposes, its digital footprint cannot be compared to the one of a real Facebook user, simply due to ethical reasons.

Furthermore, the biggest challenge we had to face was the extraordinary short time frame, in which we could conduct this experiment, as it left us with vague results and issues left unsolved. Accordingly, it is evident that Facebook’s algorithm did not have enough time to truly pick up on our account’s online behaviour. In addition to that, it is noteworthy that the open source web browser extension “Facebook tracking exposed” is currently only available in the beta form, and therefore it does not run entirely correct yet, as well as only collects data of public posts. The question that emerges from this is whether the extension would offer more accurate insights of the tracking algorithm of Facebook if it would also take (private) groups into account, as it is very possible that Facebook tracks every single movement and engagement an account has in order to suggest content and advertisements to them. However, if we were to conduct a similar type of experiment compared to this one, we would surely wait until the extension is fully developed. Lastly, it is important to note that this is our first attempt at data journalism and we did not have any kind of pre-existing knowledge of data scraping, using excel or creating graphs of any form with the help of websites. This resulted in various issues that arose while using the scraped data in excel, as well as while trying to create the chart, caused by our lack of knowledge and skill. Subsequently, in a hypothetical new experiment, it would make sense to have other group members with those aforementioned skills in the group so that we could present our data in much more detail and in an aesthetic that we thought of but could not realize.

Ultimately, we argue that although we have been able to observe a certain pattern in Diego’s News Feed, the results we are presented with are rather inconsistent and unclear. Firstly, the fact that the befriended users appear to be fake users is interesting; however, it also complicated the process of observing the potential formation of a filter bubble. In other words, the users we befriended did not substantially engage with the content we posted or our digital persona in general. However, when taking into account the content that the Facebook algorithm displayed on Diego’s News Feed, it becomes obvious that the algorithm does indeed take into account the preferences and interests of a user for the sake of future recommendations. A recurring topic of the suggested content was about the already mentioned migrant crisis and only from sources we had actively engaged with, therefore they were portraying a conservative view of it. It is therefore indicated that a filter bubble exists and is active right after having created a profile, but to be certain, whether Facebook genuinely shows us recommended content based on a filter bubble, or simply very similar content to the one we have most recently been engaged with, we would need more time to investigate the issue.

References:

Facebook Tracking Exposed. https://facebook.tracking.exposed/. Accessed 2 Apr. 2019.

ColombiaProfile.20Feb.2018, https://www.bbc.com/news/world-latin-america-19390073. Accessed 1 Apr. 2019.

Image 1:Raus aus der Blase – der Filter-Bubble-Effekt | XOVI. 2 June 2016, https://www.xovi.de/raus-aus-der-blase-der-filter-bubble-effekt/. Accessed 6 Mar. 2019.

Image 2: Screenshot of the profile of the fake profile

Image 3: Screenshot of the engagement of the fake profile

Image 4: Screenshot of selected posts on the  feed of the fake profile after the end of the experiment

Figure 1: our created graph

 

Footnote:

As for the division of the workload between the group members, we tried to split it as evenly as possible. For instance, everyone did have their turn in using the fake profile for the requested time, one group member a day and wrote a diary entry about their observations afterwards. However, there were certain tasks that we divided between the group members. For instance, while Jonathan was responsible for the visualizations and the data, Anete and Lara were ensuring the upload of a blogpost on datajlab.com. Emma and Anete were more engaged with the observations of the Facebook News Feed and the implications. Finally, Jil and Lara transformed those observations into a coherent whole while trying to make sure that the text follows the journalistic story we had in mind. All in all, we worked on this project as a group rather than individually, which led to a good team spirit.