Does the Bubble Go Beyond? An Exploration of the Urban Filter Bubble Annelien Smets Eladio Montero Pieter Ballon imec-SMIT, Vrije Universiteit Brussel imec-SMIT, Vrije Universiteit Brussel imec-SMIT, Vrije Universiteit Brussel Belgium Belgium Belgium ABSTRACT context. To the best of our knowledge, this study is the first attempt The increasing prevalence of algorithms in our everyday lives has to discuss the concept of an urban filter bubble. The aim of this raised concerns about their societal effects. Algorithmic person- paper is therefore two-fold: (1) demonstrate the need to consider alization is said to create filter bubbles that threaten democracy the consequences of algorithmic curation in the urban context, and, by curating the content users are exposed to. However, with our more specifically, (2) discuss a methodological framework to assess urban environments becoming increasingly digitally layered, they the potential reach of the urban filter bubble. become scope of algorithmic curation as well. We therefore argue It is important to note that this study does not aim to assess the that the urban context should also be scope of algorithmic impact existence of an urban filter bubble, rather the objective is to get assessments to avoid the creation of urban filter bubbles; people insights on our proposed methodology and defined measures. This only being exposed to a specific part of the city, which differs from paper will therefore mainly discuss results that are contributing to what others see because of algorithmic personalization. In this pa- this objective. In the next section, we elaborate on the concept of an per, we present a methodology to assess the urban filter bubble urban filter bubble and its relation to existing work. Subsequently, hypothesis and perform a preliminary study to verify our approach. we propose our methodology to study the urban filter bubble, and discuss our pilot experiment. Based on the latter, we conclude this CCS CONCEPTS paper by presenting our lessons learned and avenues for future work. • Information systems → Personalization. KEYWORDS 2 THE URBAN FILTER BUBBLE algorithmic curation, diversity, location, urban filter bubble In this section we elaborate the filter bubble concept by extending it to the urban sphere and discuss how it relates to previous work. 1 INTRODUCTION There has been a fundamental shift in the way we consume infor- 2.1 Conceptualization mation. Not only is there a change in the kind of media through One of the main arguments in the filter bubble debate is algorithmic which we acquire information, these new intermediaries are also curation, indicating that content exposure is no longer the result of characterized by evoking information overload. This, in turn, trig- a human selection process [9]. With the urban environment being gers innovations that automatically filter selections of information pervaded by IoT technologies, it becomes susceptible to algorithmic resulting in "more and more of the information we receive in the curation as well. However, the nature of these technologies and world [being] curated by algorithms" [17]. This algorithmic cu- their inherent connection with the physical environment suggest ration of information is said to be particularly present in online that we are not talking about content exposure but rather context search engines, personalizing their search results based on an indi- exposure. vidual’s characteristics [9]. However, this kind of personalization The way we are exposed to the urban space is no longer a mere also characterizes recommender systems for music, movies, prod- result of decisions by urban planners and architects. The fact that ucts to buy, job offers, and even our potential partners. Yet, despite our cities are becoming digitally layered urban environments [22] its omnipresence, it has raised concerns about filter bubbles; the sets the scene for applications such as Waze or Google Maps to idea that information diversity is diminishing and users are only be an additional context curator by guiding us through the urban exposed to information they agree with [19]. Despite the fact that space. These applications have not only become our primary guide empirical evidence for the filter bubble is sparse [9, 16, 17], we ac- to determine how we navigate through the urban environment, knowledge the relevance of the debate. Algorithms are increasingly they are becoming a dominant source to decide where we are going becoming part of our everyday lives and cause scholarship to ask as well. Research on travel information search shows that we are critical questions about their societal impact [7, 25]. increasingly turning to online information for destination decision- With our urban environments becoming increasingly digitally making [11] and the rise of social media has even amplified this layered, the urban context is no longer excluded from algorithms behavior [26]. The objective of online applications to become con- and should be scope of critical assessments as well. In this work, text curators is also exemplified by Google Maps’ recent feature we therefore argue that the filter bubble debate should go beyond, which recommends places to go to [3]. Obviously, the mere avail- and also take into account the potential consequences in a physical ability and use of information about the urban environment does not necessarily imply that our choices are being algorithmically ImpactRS ’19, September, 2019, Copenhagen, Denmark Copyright ©2019 for this paper by its authors. Use permitted under Creative Commons steered by these applications. Except that there is more: Pan et al. License Attribution 4.0 International (CC BY 4.0). [18] showed that users are biased towards higher ranked results ImpactRS ’19, September, 2019, Copenhagen, Denmark Smets, Montero and Ballon even when there are doubts about the relevance of the search result. specifically select Google Maps’ search results as they are guaran- This implies that when someone uses a search engine to look for teed to have a corresponding geolocation. a particular thing to do in the city, it is likely that they will only The queries in Google Maps are performed manually from a consider the first results. Although this argument can be contested blank state computer with no cookies stored, using Google Chrome. by appealing to the autonomy of the human decision-maker, it To vary the search location (see 3.3) the GPS location sensors on becomes essential when thinking of algorithmic decision-makers. Google Chrome are overridden. This option was chosen instead It is no longer science-fiction to think of a scenario where we ask of changing the IP address location with a VPN, because "Google our digital personal assistant to make a dinner reservation [14] and Search personalizes search results largely based on the provided if we do not specify the venue ourselves, it will be algorithmically GPS coordinates rather than the IP address" [13]. picked. What would this algorithmic decision-making mean for our The construction of the search terms is based on a set of travel- explorative right to the city? related terms that are found to be frequently used when exploring This example demonstrates the need to consider the conse- a new destination [26]. To account for differences that might be quences of algorithmic curation in the physical sphere and even- related to proximity [13], we combine these search terms with 3 tually the curation of our experiences as well. Especially in urban cities (Brussels, Ghent and Berlin) that have a different proximity environments, cities, where diversity and the right to the city are to our search location. This results in the following 9 queries: fundamental aspects we should take into account the contextual consequences of algorithmic curation. After all, the consequences • Restaurant in go beyond the mere experience of citizens, and there are also politi- • Hotel in cal and economic interests at stake. If an algorithm is deciding how • What to do in people are navigating through a city, which parts of the city they After performing the queries, we scrape the results using the are exposed to; it is deciding where they spend their time, where lmxl package in Python [15]. This way, each element can be mapped they spend their money. using an XPath (XML Path Language) route [24]. In this study, we therefore explore the filter bubble hypothesis in an urban context. We start from the assumption that people are increasingly using online search engines to get recommendations 3.2 Diversity measures on what to do in the city or where to go. The urban filter bubble Spending time in a physical environment implies that we are do- then articulates the idea that by using the results of online search ing something somewhere. To account for these two concepts, we engines people are likely to only be exposed to a specific part of expand the diversity measurement from similar studies [10] to a the city, which differs from what others see because of algorithmic two-dimensional concept. We use two measures of diversity based personalization. on respectively content and context. The former is similar to other diversity measures in related studies [10], while the latter is based on the actual location coordinates of the search results and ex- 2.2 Related work presses how they are distributed among the urban space. Moreover, The filter bubble hypothesis has been mainly studied within the these diversity measures are calculated for two perspectives: the domain of public policy discourse, where it is linked to concepts intra-diversity perspective accounts for the diversity among one’s such as echo chambers [6, 23] and viewpoint diversity [4]. While individual search results (how diverse are my search results?), while those studies focus on the impact of personalization, there is a vast the inter-diversity perspective considers the diversity among the amount of research discussing the actual techniques for algorith- search results of multiple users (how much do my search results mic search personalization [5] and some studies specifically focus differ from yours?). This 2x2 framework results in a set of diversity on geolocation [1, 27]. Most of the work described in this paper measures and their corresponding calculations as shown in Table 1. builds upon the research of Kliman-Silver et al. [13], showing that The purity accounts for the diversity among one’s individual differences in search results grow as physical distance increases. search results by calculating the extent to which the result set (i.e. However, in this work, we take the online search results back to cluster) contains elements of a single class. their physical location in the urban space and question what per- The Jaccard-index defines the overlap in search results between sonalization might imply for the places we actually visit in the two sets: 0 when there is no overlap, and 1 when they have exactly city. the same results. This index does not include the order of the results and therefore the edit distance is taken into account as well: it 3 METHODS shows how many transformations (i.e. insert, delete, substitute) are In this section, we describe our search methodology, followed by needed to make the lists identical. outlining the different diversity measures that are used to analyze The total sum of squares represents the sum of all pairwise the results. Finally, we discuss the research design of the pilot study square distances between the locations of the search results. This to validate our approach. is an indication of how disperse the results are. The cluster centroid distance points at the distance between the centroids of two result sets and can be used to assess if differ- 3.1 Search methodology ent users are indeed seeing different parts of the city. The cluster Considering both Google’s and Google Maps’ high market share centroid is calculated as the average of the location coordinates of [21] this study focuses on search results from Google Maps. We the elements in the cluster (i.e. the set of search results). Does the Bubble Go Beyond? ImpactRS ’19, September, 2019, Copenhagen, Denmark Table 1: Diversity measures we would expect based on search proximity, Figure 1 indicates that the overlap does not increase when the physical distance increases. Intra-diversity Inter-diversity The average edit distance is 23 and taking into account that there are approximately 20 results in each list and there is only 30% simi- Jaccard-index, Content Purity larity, this edit distance indicates that the order of the shared results edit distance is quite similar. Cluster total Cluster centroid Context The Jaccard-indexes for the categories of the what to do queries sum of squares distance are also represented in Figure 1. The similarity of the categories is on average 23.7%, which means that the different search results 3.3 Study design not only represent different locations but also different kinds of activities. Nevertheless, the average purity1 of the Dutch and French We set up a small pilot study to validate our methodology and assess what to do results is respectively 0.64 and 0.66. Hence, within one the value of our defined metrics. To this end, we use an agent-based language, on average 65% of the results are of the same category testing approach [9] and define 3 features that will be used for and neither language shows more diversity in these categories. personalization: language, search location and user profile. The language of the queries varies between Dutch and French, two of the official languages in Belgium. We chose to vary the language of the search queries because it is a realistic reflection of the Belgian population. The next variable is the search location, i.e. the location from which the searches are being performed. We query from two dis- tinct locations in the Brussels area, based on the idea that "users’ geolocation can be used as a proxy for other demographic traits" [13]. The search locations are Sint-Jans-Molenbeek (SJM) and Sint- Pieters-Woluwe (SPW), two municipalities in Brussels that show significant differences in terms of unemployment rate and educa- tion levels, which are socio-demographic traits that are found to Figure 1: Average Jaccard-index: Dutch vs French. have an impact on the kind of activities people do in a city [2]. Finally, we construct two different user profiles that are used to perform the queries. These profiles vary mainly in terms of ed- On the other hand, the diversity among the results of queries ucational level, employment status and family situation as these from two search locations (SJM and SPW) is significantly smaller. have been indicated as influencing the kind of cultural activities The average overlap between those search results is 77.2%, more one undertakes. Similar to previous studies [8–10], we build the than twice the overlap of the French and Dutch search results. user profiles by means of two phases. In the training phase, we feed Hence, in this case, the personalization due to language is much the profile by (1) searching Google for five agent-specific terms, (2) stronger compared to the one based on search location. In line with adding a related product in an online shopping cart and (3) brows- previous research [13], the Jaccard-indexes in Figure 2 illustrate that ing through 5 articles in a specified media outlet. We execute this queries related to Brussels are more diverse than queries related training phase for five consecutive days and alternate the training to Ghent and Berlin. In this case, the search location (SJM or SPW) phase with the testing phase during which the search queries are thus stronger influences the personalization when the proximity performed. This is in contrast to the above-mentioned studies, who increases. did not alternate these phases. Our decision to alternate the phases is motivated by the goal to observe the evolution of the personal- ization over time. Using this approach, we build two user profiles which we refer to as Agent A and Agent B. 4 RESULTS Based on the aforementioned methodology we collected 4,988 search results that consist of 422 unique locations. 4.1 Content diversity The Jaccard-index and edit distance have been used in similar re- Figure 2: Average Jaccard-index: search locations. search to assess content diversity [10, 13], and prove to be valuable in our pilot study as well. For example, Figure 1 shows the average Jaccard-indexes for the Dutch and French search results (e.g. Hotel in Brussel and Hôtel à Bruxelles). Over the 9 queries, the average 1 Purity was only calculated for the what to do results, since we only considered the most Jaccard-index is 0.31, indicating that on average 31% of results ap- general categories. The categories of the restaurant and hotel results were consequently pear in both French and Dutch search results. In contrast to what the same due to the nature of the query. ImpactRS ’19, September, 2019, Copenhagen, Denmark Smets, Montero and Ballon 4.2 Context diversity While the content diversity measures provide valuable and straight- forward insights, the context measurements need some additional attention during interpretation. For example, a visual inspection of the data in Figure 3(a) shows that generally most locations are clustered in the city center, with a few outliers for the Dutch results. However, our metric to indicate the dispersity of results (total sum of squares) does not account for outliers and consequently indicates that French results are 87% denser than Dutch results (see Figure 4(a) restaurants in Ghent). The mere interpretation of the measurement without the visual inspection could lead to a conclusion that is inconsistent with the actual data. Figure 4: (a) Relative density between Dutch and French search results. (b) Centroid distance between Dutch and French search results. Figure 3: Restaurant results in Ghent and Brussels. Colors represent the query language in which the result appears: purple (only Dutch), green (only French) and yellow (both). Although the same consideration in terms of outliers applies to the cluster centroid distance, this metric holds some potential as well. For example, as illustrated in Figure 4(b) the cluster centroid distance for restaurants in Ghent is 257, and 450 for Brussels. Indeed, Figure 3 shows that the French and Dutch search results are located closer to each other in Ghent (a) compared to Brussels (b). Finally, our preliminary results also demonstrate the need to account for both content and context diversity since one does not necessarily imply the other. For example, Figure 5 shows that French and Dutch results are practically covering the same parts of the Brussels’ area (low context diversity), while the Jaccard-index of Figure 5: What to do results in Brussels. Colors represent the this result set is only 32% (high content diversity). query language in which the result appears (see Figure 3). 4.3 A note on user profiles The reader might have noticed that we did not discuss any of the interesting pattern for the Jaccard-index of Berlin, a closer look at results related to the personalization based on user profiles. As the data did not provide any additional path to continue on. Addi- explained before, we alternated the training phase with the testing tionally, the visual inspection of the data did not show significant phase to study the influence of the profile on the search results. differences in terms of locations of the search results, even not for Our assumption is that the more Google learns about the agent, those related to Berlin. Therefore, we will not continue the discus- the more personalization of search results would occur. Our main sion of the diversity of these results, because it would lead beyond interest then lies in the geographic location of these search results. the scope of this paper. However, as Figure 6 shows, the overlap in the search results is not significantly decreasing, and, in some cases, even increases 5 CONCLUSION over time. The figures for other queries and agents are similar and This paper aims to contribute to the debate on algorithmic cura- we omit them for brevity. Despite the fact that Figure 6 shows an tion by addressing this topic in the urban context. To this end, we Does the Bubble Go Beyond? ImpactRS ’19, September, 2019, Copenhagen, Denmark many cities are known for having a fashion district, or areas that are populated by specific subgroups. Hence, there is a thin line between these inherent clusters and urban filter bubbles induced by algorithmic curation. These considerations stress the importance to critically assess the algorithmic curation of our contexts. In our understanding, this also comes with a thorough reflection on its societal impact including questions such as what is the best way to get people engaged with the city? Why is it beneficial to have an unfiltered discovery of the city? If there is an urban filter bubble, who is responsible to burst it? Figure 6: Jaccard-index: Agent A vs. Agent B (restaurants). In our future work, we will address some of these questions by exploring the value and meaning of serendipity in urban recom- mender systems. We hope that this work contributes to the debate proposed a methodology that conceptualized diversity in a two of algorithmic curation and inspires others to continue on this topic. dimensional way: content-context, and individual-group. Making these distinctions allows to (1) not only discuss the content diversity ACKNOWLEDGMENTS but also consider the physical diversity and (2) account for both We gratefully acknowledge the help of Ms. Azadasl in collecting individual diversity as well as diversity among multiple users. We the data, and thank the anonymous reviewers for their constructive argue that this multidimensional concept is required to account comments on an earlier version of this paper. for the complexity of establishing diversity in an urban context and allows to foresee different configurations depending on the REFERENCES situation. [1] Leonardo Andrade and Mário J Silva. 2006. Relevance Ranking for Geographic The results of our pilot study demonstrate the need to account IR.. In GIR. [2] Julie Badisco, Ignace Glorieux, Lindsay Jacobs, and Katrien Lauwerysen. 2008. for both content and context diversity, however, there is still room De weg naar cultuur. Ruimtelijke aspecten van culturele interesse en deelname for improvement to our approach. Although content diversity can aan het cultuuraanbod. (2008). be assessed in a meaningful and straightforward way, the context [3] Muhammad Ahmad Bashir, Sajjad Arshad, and Christo Wilson. 2016. "Recom- mended For You": A First Look at Content Recommendation Networks. In Proceed- measurements appear to require some adjustments to deal with the ings of the 2016 ACM on Internet Measurement Conference - IMC ’16. ACM Press, skewing impact of outliers in the physical dimension. Moreover, due Santa Monica, California, USA, 17–24. https://doi.org/10.1145/2987443.2987469 [4] Engin Bozdag. 2015. Bursting the filter bubble: Democracy, design, and ethics. Ph.D. to our manual approach, our study remains small scale while for an Dissertation. in-depth analysis of the urban filter bubble hypothesis it would be [5] Zhicheng Dou, Ruihua Song, and Ji-Rong Wen. 2007. A large-scale evaluation and beneficial to collect data on a large scale. This would allow applying analysis of personalized search strategies. In Proceedings of the 16th international conference on World Wide Web. ACM, 581–590. statistical methods to verify the hypothesis and implementing more [6] Elizabeth Dubois and Grant Blank. 2018. The echo chamber is overstated: the variation, for example in terms of search locations or languages. moderating effect of political interest and diverse media. Information, Communi- Another limitation of our current study is that we have mainly cation & Society 21, 5 (May 2018), 729–745. https://doi.org/10.1080/1369118X. 2018.1428656 focused on the context: where are the suggestions physically located. [7] Tarleton Gillespie. 2017. Algorithmically recognizable: Santorum’s Google prob- This focus arose due to our plea for the inclusion of context diversity. lem, and Google’s Santorum problem. Information, Communication & Society 20, 1 (Jan. 2017), 63–80. https://doi.org/10.1080/1369118X.2016.1199721 We did look at the content to a certain extent by taking into account [8] Mario Haim, Florian Arendt, and Sebastian Scherr. 2017. Abyss or Shelter? On the categories of the activities, however, future work could focus on the Relevance of Web Search Engines’ Search Results When People Google for studying this in more depth (e.g. price range). Finally, there is still Suicide. Health Communication 32, 2 (Feb. 2017), 253–258. https://doi.org/10. 1080/10410236.2015.1113484 the question of validity. One may indeed question if an agent-based [9] Mario Haim, Andreas Graefe, and Hans-Bernd Brosius. 2018. Burst of the Filter testing approach could be said to be representative for an actual Bubble?: Effects of personalization on the diversity of Google News. Digital user. Nevertheless, taking into account the limitations of this pilot Journalism 6, 3 (March 2018), 330–343. https://doi.org/10.1080/21670811.2017. 1338145 study and our recommendations, future studies could apply our [10] Aniko Hannak, Piotr Sapiezynski, Arash Molavi Kakhki, Balachander Krish- approach to large scale data collected from real-world users. namurthy, David Lazer, Alan Mislove, and Christo Wilson. 2013. Measuring personalization of web search. In Proceedings of the 22nd international conference The question of how to research the (societal) impact of algo- on World Wide Web - WWW ’13. ACM Press, Rio de Janeiro, Brazil, 527–538. rithms is still part of an ongoing debate. However, we agree with https://doi.org/10.1145/2488388.2488435 authors like Kitchin [12] and Seaver [20] arguing that algorithms [11] Jens Kr Steen Jacobsen and Ana María Munar. 2012. Tourist information search and destination choice in a digital age. Tourism management perspectives 1 (2012), should be considered as algorithmic systems. Consequently, re- 39–47. searching their impact is not limited to one methodology or assess- [12] Rob Kitchin. 2017. Thinking critically about and researching algorithms. Infor- ment, but should encompass multiple research acts of which the mation, Communication & Society 20, 1 (2017), 14–29. [13] Chloe Kliman-Silver, Aniko Hannak, David Lazer, Christo Wilson, and Alan work in this paper could be one. In line with this reasoning, we Mislove. 2015. Location, Location, Location: The Impact of Geolocation on Web conclude this paper by asking questions that go beyond the mere Search Personalization. In Proceedings of the 2015 ACM Conference on Internet Measurement Conference - IMC ’15. ACM Press, Tokyo, Japan, 121–127. https: issue of measuring the impact of algorithmic curation in the urban //doi.org/10.1145/2815675.2815714 sphere, as they provide directions for future work. After all, we [14] Yaniv Leviathan and Yossi Matias. 2018. Google Duplex: An AI System for acknowledge that avoiding urban filter bubbles is not a straight- Accomplishing Real-World Tasks Over the Phone. http://ai.googleblog.com/ 2018/05/duplex-ai-system-for-natural-conversation.html forward exercise since the urban environment itself is already to [15] lxml. 2017. lxml - Processing XML and HTML with Python. Retrieved April 18, a great extent characterized by districts or clusters. For example, 2019 from https://lxml.de/ ImpactRS ’19, September, 2019, Copenhagen, Denmark Smets, Montero and Ballon [16] Judith Möller, Damian Trilling, Natali Helberger, and Bram van Es. 2018. Do not [22] Joe Shaw and Mark Graham. 2017. An informational right to the city? Code, blame it on the algorithm: an empirical assessment of multiple recommender content, control, and the urbanization of information. Antipode 49, 4 (2017), systems and their impact on content diversity. Information, Communication & 907–927. Society 21, 7 (2018), 959–977. [23] Cass R Sustein. 2009. Republic. com 2.0. Nova Jersey: Princeton University Press [17] Jacob Ørmen. 2018. A discussion of research designs for assessing algorithmic (2009). curation. The Routledge Handbook of Developments in Digital Journalism Studies [24] W3C. 2017. XML Path Language (XPath) 3.1. Retrieved April 18, 2019 from (2018). https://www.w3.org/TR/2017/REC-xpath-31-20170321/ [18] Bing Pan, Helene Hembrooke, Thorsten Joachims, Lori Lorigo, Geri Gay, and [25] Michele Willson. 2017. Algorithms (and the) everyday. Information, Communica- Laura Granka. 2007. In google we trust: Users’ decisions on rank, position, and tion & Society 20, 1 (Jan. 2017), 137–150. https://doi.org/10.1080/1369118X.2016. relevance. Journal of computer-mediated communication 12, 3 (2007), 801–823. 1200645 [19] Eli Pariser. 2011. The Filter Bubble: What The Internet Is Hiding From You. Penguin [26] Zheng Xiang and Ulrike Gretzel. 2010. Role of social media in online travel UK. information search. Tourism Management 31, 2 (April 2010), 179–188. https: [20] Nick Seaver. 2014. Knowing algorithms. //doi.org/10.1016/j.tourman.2009.02.016 [21] Dennis Sellers. 2018. Google Maps has 67% of the naviga- [27] Bo Yu and Guoray Cai. 2007. A query-aware document ranking method for tion app market compared to Apple Map’s 11%. Retrieved geographic information retrieval. In Proceedings of the 4th ACM workshop on April 19, 2019 from https://www.appleworld.today/blog/2018/7/10/ Geographical information retrieval. ACM, 49–54. google-maps-has-67-of-the-navigation-app-market-compared-to-apple-maps-11