New tab page recommendations cause a strong suppression of exploratory web browsing behaviors
NNew tab page recommendations cause a strong suppressionof exploratory web browsing behaviors
HOMANGA BHARADHWAJ ∗ , IIT Kanpur, India
NISHEETH SRIVASTAVA,
IIT Kanpur, IndiaThrough a combination of experimental and simulation results, we illustrate that passive recommendationsencoded in typical computer user-interfaces (UIs) can subdue users’ natural proclivity to access diverseinformation sources. Inspired by traditional demonstrations of a part-set cueing effect in the cognitive scienceliterature, we performed an online experiment manipulating the operation of the ‘New Tab’ page for consentingvolunteers over a two month period. Examination of their browsing behavior reveals that typical frequencyand recency-based methods for displaying websites in these displays subdues users’ propensity to accessinfrequently visited pages compared to a situation wherein no web page icons are displayed on the new tabpage. Using a carefully designed simulation study, representing user behavior as a random walk on a graph,we inferred quantitative predictions about the extent to which discovery of new sources of information maybe hampered by personalized ‘New Tab’ recommendations in typical computer UIs. We show that our resultsare significant at the individual level and explain the potential consequences of the observed suppression inweb-exploration.CCS Concepts: •
Human-centered computing → Empirical studies in HCI ; User studies ; Web-basedinteraction ;Additional Key Words and Phrases: filter bubble; personalization; exploration; habitual browsing; informationdiet; human memory; cognitive science
Reference Format:
Homanga Bharadhwaj and Nisheeth Srivastava. 2018. New tab page recommendations cause a strong suppres-sion of exploratory web browsing behaviors.
Journal
1, 1 (December 2018), 25 pages. https://doi.org/10.1145/nnnnnnn.nnnnnnn
The medium is the message [19]. McLuhan’s gnomic aphorism has, in the half-century since itsutterance, launched countless humanistic appraisals of the manner in which the form of moderncommunication technologies have shaped, and continue to shape the way people think. Given theextent to which web-based technologies now saturate our lived environment, substantiating thesecritiques quantitatively is very important.For instance, it seems intuitively plausible, even likely, that the hyper-linked nature of webinterfaces, in contrast with the linear nature of books, leads to greater distraction and shallowerprocessing of information [5]. A survey of recent empirical research addressing this question,however, notes that even though the basic hypothesis is borne out - students reading material onscreens do worse on tests than students reading material on paper, this performance differentialappears to reduce quite rapidly with practice [20]. While it is still too early to conclude definitively ∗ This is the corresponding authorAuthors’ addresses: Homanga Bharadhwaj, IIT Kanpur, Kanpur, U.P., India, [email protected]; Nisheeth Srivastava,IIT Kanpur, Kanpur, U.P., India, [email protected].© 2018 ./2018/12https://doi.org/10.1145/nnnnnnn.nnnnnnn Journal, Vol. 1, No. 1, Article . Publication date: December 2018. a r X i v : . [ c s . H C ] D ec either way about the validity of this particular critique, a general point remains: access to empiricaldata permits broad critiques of technology’s influence on cognition to themselves be substantiated.A similar story can be told about yet another intuitively appealing hypothesis - the filter bub-ble [24]. Here, the claim is that web-based recommendation services, by virtue of their algorithmiclogic, recommend digital objects to people that they are already known to prefer, thus reducing thepossibility for them to experience new content that they may or may not like. Again, the claim iseasily believable, but has proved difficult to substantiate empirically. An analysis of the effect ofaccepting movie recommendations from the Movielens system showed that while the set of moviesrecommended became gradually less diverse with time, users who chose to follow the system’srecommendations received more diverse recommendations than users who ignored them [22].Recent research seeking to evaluate the existence of filter bubbles in Google News’ personalizationalgorithm reported extremely minor differences in news articles presented to different user accountsexplicitly set up to provoke differentiation [13]. Mixed evidence for filter bubbles was obtainedfrom an analysis conducted by Microsoft on the news reading habits of 50000 Internet Explorerusers, which found that while political articles accessed via social media or search engines are morepolitically extreme than those accessed directly from websites, the use of these channels is alsomore associated with exposure to opposite political views than direct access [10].There is an interesting dichotomy between theoretical and empirical evaluations of the filterbubble. Whereas theoretical critiques focus heavily on how recommendations could potentiallynarrow the diversity of users’ demand patterns for digital media [24], empirical research focuseson identifying whether recommendation systems are narrowing the diversity of content providers’ supply of digital objects [13, 22]. Work that explicitly tracks users’ consumption patterns doesappear to detect a small bubble effect [10], but the restriction of this study to studying only politicaldiversity reduces the generalizability of its results.In this paper, we report results from an experimental study of an extremely general versionof the filter bubble hypothesis, evaluated strictly on consumption patterns rather than supply,and without restricting the nature of content consumed. Specifically, we investigate whether thepresence of frequently visited tabs on ‘new tab’ pages in modern browsers leads to a concentrationof browsing behavior into a more restricted set of websites than when such recommendation-baseddisplays are absent.This hypothesis originated from a simple analogy: users’ interaction with the ‘new tab page’display in modern browsers almost precisely reproduces the environmental conditions under whichpsychological test subjects experience memory inhibition in part-set cueing experiments [31]. Inthese classic experiments, when asked to learn a set of words for future recall, people performedworse during memory testing when they were actually shown some subset of the words than whenthey were simply asked to remember as many words as they could.Analogously, in principle an Internet user browsing the web can go to any URL they remember(or can type the first few letters of). This maps on to the free recall memory task. The user’sattention is drawn to a set of page icons on the new tab page during the recall task, reproducingthe partial set presentation condition of part-set cueing experiments. If the lab finding translates tothis natural setting, we’d expect that users’ propensity to access non-displayed pages would beimpaired by the mere presence of web page icons in the new tab page display.We conducted a controlled, longitudinal, within-subject experiment wherein we manipulated thebehavior of consenting volunteers’ new tab page displays without their knowledge and analyzedtheir browsing history at the end of the observation period to measure the degree to which our covertmanipulations influenced users’ behavior. We found strong statistical support for our hypothesis.Usage of new tab pages with personalized recommendations considerably reduced the numberof unique websites visited by users, suggesting concentration of website browsing repertoires. Journal, Vol. 1, No. 1, Article . Publication date: December 2018. ew tab page recommendations cause a strong suppression of exploratory web browsingbehaviors :3
Additional analyses revealed that this concentration was not uniform, and was seen most clearlyprecisely during browsing events most strongly associated with free memory recall - users typingURLs into address bars in the new tab page display, with recommendations visible underneath.To assess the extent to which this measured deterioration in exploratory browsing could affectthe diversity of users’ information foraging on the web, we also ran a simulation study, using ourdata to set parameters for a random walk model of a web surfer seeking information from diversesources. We showed that, for empirically grounded values of exploratory browsing, the size ofsuppression effects seen in our empirical sample would correspond to reductions in the frequencywith which diverse information sources are accessed, of nearly 50% for the median user.Thus, in this paper, we present empirical and simulation evidence to show that the mere presenceof a set of web page icons on browser new tab pages reduces users’ propensity to access infrequentlyvisited web pages by nearly half. We also discuss the implications of this striking finding on the filterbubble literature, the principles of designing communication interfaces, and ethical considerationsof web mediated communication.
Fig. 1. A typical new tab page with different web page icons shown. Although the layout differs for differentweb-browsers and different versions of the same browser, the basic premise of displaying different web pageicons persists across browsers.
Journal, Vol. 1, No. 1, Article . Publication date: December 2018.
We anticipated large individual differences in browsing behavior across subjects, suggestingthat cohort-level differences may or may not be representative of the median user’s experience.Therefore, we focused on designing a within-subject experiment, with experimental manipulationsthat each participant in the experiment would experience in different time blocks.
We designed a Firefox web extension using JavaScript, HTML and CSS. It broadly consists ofthree parts, a webpage to replace the default new tab display, a browser action page to log thebrowsing data into a local file and a remote script located in the experimenter’s server that is usedto tweak the type of websites displayed in the new tab page, as shown in Figure 2. In this design,the experimenter does not have access to the participant’s data during the course of the experiment.At the end of the experiment, the experimenter explicitly asks for the logs, and participants areadvised to curate their history to remove personal and private content before handing them over, ifthey choose to do so.We programmed four different behavior modes in the replacement new tab page generated whenpeople install our extension, controlled by a parameter in the extension script that we could changeremotely without the user’s knowledge. Each participant’s replacement new tab page shows theweb page icons as per exactly one of the following four modes at a given time.(1) Display the most visited sites in browsing history(2) Display the least visited sites in browsing history (at least one visit, naturally)(3)
Default behavior, following Mozilla’s frecency algorithm - a combination of frequency andrecency [11](4) Display a blank pageOur web-extension logs the webpages the user visits, the time of visits and the transition typefrom one webpage to the next - typed in the address or search bar, linked to from the content of adifferent site, or clicked through the browser UI. It also automatically logs all the websites currentlydisplayed in the new tab page whenever it is opened. The replacement new tab page is designed tovisually resemble the default Firefox new tab page so that after a certain period of time, the user isnot very conscious of the extension running in background. Once the extension is installed on aparticipant’s computer, the experimenter can change the behavior of the replacement new tab pageby simply changing a parameter in a participant-specific file hosted on the experimenter’s server.We used a fixed switching interval of 5 days, with all participants beginning with the default mode ofnew tab behavior, and with further switches assigned pseudorandomly, ensuring that all participantsreceived roughly equal duration of exposure to each of the three new behavior modes. Since we useda rolling recruitment process the 5-day blocks don’t systematically interact with weekends for allparticipants, an important consideration in a behavior measurement experiment. This interaction isapproximately random across subjects in our experiment. We used a pseudo-random allocation ofconditions instead of using a fixed schedule counterbalanced across participants because in our in-the-wild study, we could not control when participants would stop participating in the experiment,and expected large attrition rates. If all participants completed the full term of the experiment,we would have approximately the same block-wise assignment as if we had assigned fixed blocksto begin with. By good fortune, we had a low attrition rate ( 20%), and all eighty participantswho lasted the full two month term complied with our reminders for sending in browsing logsimmediately. As a result, condition viewing time was approximately balanced both across andwithin our sample, averaging 24.46, 25.18, 24.56, 25.78 by percent across our four conditions (Most
Journal, Vol. 1, No. 1, Article . Publication date: December 2018. ew tab page recommendations cause a strong suppression of exploratory web browsingbehaviors :5
Fig. 2. The experiment design visited, Least visited, Default, Blank UI) with standard deviations 1.41, 1.81, 1.49, 2.41 across all80 participants. Further, rolling recruitment randomizes weekday-weekend interactions amongparticipants.
Eighty volunteers from the general population (32 female, age 28 . ± . Journal, Vol. 1, No. 1, Article . Publication date: December 2018. type of device for a participant was imposed. A total of 10 participants reported to have used mobilephones / tablet computers while the remaining participants used either a laptop / desktop computerfor installing the extension. At the end of two months of browsing for each participant, we askedthem to curate the log files and share them with us, retaining the option to refuse. The two-monthlong browsing logs volunteered by these eighty participants formed our primary dataset. 42 of the80 participants continued using the web extension for a longer duration (another two months). Theresults from this additional data are presented in Section 3.1.At the time of submission of the logs, we asked the participants what fraction of the website logshad been curated by them. None of the eighty participants had curated more than 5% of the totalcontent, with most participants not curating at all, which is reasonable and does not negativelyaffect our study. We had not disclosed to these participants that we intended to manipulate theirnew tab page’s behavior remotely. To assess the extent to which they may have become awareof this possibility, at the time of collecting the logs we also asked them whether anything aboutthe websites displayed in the new tab page ‘bothered’ them over the past two months. They wereasked to express their opinion on a 0-10 scale with 0 being the ‘not aware’ state and 10 the ‘veryaware’ state. They were not specifically asked if they thought the type of websites shown werebeing altered remotely to avoid experimenter demand effects [6]. Subjects’ responses are shown inTable 1.
Table 1. Participants’ level of awareness of our experimental manipulation.
Opinion 0-4 5-7 8-10Number of subjects 70 7 3These survey results suggest that the majority of our participants remained naive to the purposeof the experiment, and continued to behave naturally in the face of our remote manipulations.
The self-curated browsing logs obtained from all 80 participants in the experiment constituted ourprimary analysis material. These logs contained time-stamped instances of different website URLsvisited by the participants. From these logs, we filtered out pop-ups and ads using a white-list, andused a 30-minute gap between page views to delimit browsing sessions. We collapsed web-pagesvisited on the same primary web domain into multiple visits to the same website, and likewisecollapsed page refreshes (immediate repetitions of the same page in the log) as a single visit to thepage. The few occurrences of dynamic web-pages in the browsing log were collapsed to that ofthe parent domain. We explicitly handled pages visited from search engines in the browser. Forall occurrences of Google, Yahoo, Bing, Ask.com, Baidu, Wolframalpha and DuckDuckGo in thebrowsing log, we considered the page clicked immediately afterwards in the search results. If thesearch engine address was typed out in the new tab page, we treated the page clicked from thesearch results as if it was typed out in the address bar.
Since our hypothesis generally concerns the potential reduction of the diversity of websites thatusers visit as a consequence of the format of the new tab page display, our first coarse measurementof this property was the number of unique websites (defined via the primary web-domain ineach URL) visited by our experiment participants during each mode of the new tab page’s displayproperties.
Journal, Vol. 1, No. 1, Article . Publication date: December 2018. ew tab page recommendations cause a strong suppression of exploratory web browsingbehaviors :7
Fig. 3. Number of unique website count for each of the four modes averaged over all 80 individuals. Errorbars represent ± Combining the default and most visited modes’ data into one sample and for the blank and leastvisited modes’ data into another, a two-sample T-test yields strong statistical significance p < − and a very large effect size (Cohen’s d = 1.81) as shown in Figure 3. Thus, there is clearly a largedifference in user behavior across these experimental cohorts, in the predicted direction. Users visita lot more unique websites on average when the new tab page is blank or shows infrequently visitedwebsites, suggesting that the presence of recommendations by default is suppressing the repertoireof webpages they might naturally visit. For completeness, we show the results of bonferronicorrected pair-wise T-test among all the pairs of conditions in our study (Table 2). However, the Table 2. Results of Pairwise T-test for all the pairs of conditions (as shown in Fig. 3). Number of uniquewebsite counts for each of the four modes are averaged over 80 individuals.
Condition 1 Condition 2 p-value Inference
Most visited Least visited 0.000011 p <0.001
Most visited Default UI 0.620102 p>0.05Most visited Blank UI 0.000008 p<0.001
Least visited Default UI 0.000024 p<0.001
Least visited Blank UI 0.522001 p>0.05Default UI Blank UI 0.000018 p<0.001 heterogeneity of web browsing behavior across individuals could conceivably inflate these statistics,in the sense that a few highly prolific web users in our participant pool could skew the cohort-leveldifferences should they happen to conform to our hypothesis. Thus, it is important to also analyzethe data at the individual-level. We plot the difference between behavior during blank UI mode and
Journal, Vol. 1, No. 1, Article . Publication date: December 2018. default mode for all participants in Figure 4. Note that not a single of the eighty participants visitedmore unique websites while the new tab page behaved under its default settings, vis-a-vis when thedefault recommendation display was switched off (Blank UI mode). A one-sample T test of theseabsolute differences with respect to zero sample mean was statistically significant at p < − . Theaverage participant’s unique site visit count increased by 15% (median improvement 12%) whenthe new tab page was set to display a blank page over their corresponding counts when the newtab page worked by default. Fig. 4. Absolute difference between the number of unique websites visited under the blank new tab pagemode and under the default mode for all experiment participants.
It is reasonable to conclude from this observation that changing the behavior of the new tab pagedoes in fact reduce the diversity of websites browsed. But this analysis does not substantiate ourhypothesis about the mechanism responsible for this reduction in diversity - a reduction in users’propensity to explore content, caused by interaction with the new tab page interface.We needed a more granular view of the data to see if this might actually be true. Recall that westore both the timestamped page visit, and the web event that brings the user to that page. Wecategorized the pages according to the visit frequency and transition type to focus especially onpage visits wherewith the user’s exploratory intent to retrieve information from a new informationsource can be discerned. This categorization is illustrated with a pseudo-flowchart in Figure 5,showing visually how we use the two decision points (frequency and transition type) to categorizeweb page visits into non-exclusive categories. To interpret the set-theoretic implications of theflowchart correctly, both UI and pure exploration are subsets of exploration, but exploration and habit are mutually exclusive. A brief description of these terms and their rationale follows. • Habit.
Websites with high visit counts (removing the influence of interstitial URLs) areexpected to correspond to content that users are in the habit of accessing. Since the number
Journal, Vol. 1, No. 1, Article . Publication date: December 2018. ew tab page recommendations cause a strong suppression of exploratory web browsingbehaviors :9
Fig. 5. The analysis flowchart. The two decision points (diamond boxes) represent decisions by frequency(high = 4th quartile, low = 1st quartile) and transition type (link, typed, UI). Numbers inside boxes representthe number of webpage visits categorized under each category. of webpages an individual visits will inform what constitutes high for them, we categorizeas habit -based visits all page visits to webpages whose visit frequency occurs in the fourthquartile ( Q ) of visit frequencies sorted in ascending order per browsing history log. In ameasure of the heavy skew in the distribution of webpage counts, this quartile accounts forabout 70% of all site visits across all participants’ data. The absolute value of average visitcount of websites in the fourth quartile ( Q ) across all 80 participants is 42 ±
8. That is, onaverage, participants visited websites in this category thrice in every four days during our 60day observation period. A distinct sub-category “pure habit" takes into account frequentlyvisited webpages visited by typed transitions from the address bar (or search the web box) ofan existing tab. • Exploration.
As above, it is intuitive that when people are browsing in an exploratorymanner, such browsing will be marked by visiting new sources of content, hence newwebpages. We mark as exploration -based visits all visits to pages that fall in the first quartile( Q ) of visit frequencies sorted in ascending order per browsing history log . This quartileaccounts for about 9% of all site visits across all participants’ data. The absolute value of Not all visits to infrequently visited webpages constitute exploration. Sometimes users may have only need to go to awebsite, like a bank website very infrequently, on purpose. Thus, our definition overestimates the amount of exploration.Since we are ultimately interested, as we will see below, in the change in the amount of exploration under controlledmanipulation, this overestimation of the absolute quantity ends up not being crucial.Journal, Vol. 1, No. 1, Article . Publication date: December 2018.
10 Homanga Bharadhwaj and Nisheeth Srivastava average visit count of websites in the first quartile ( Q ) across all 80 participants is 8 ±
3. Thatis, on average, participants visited websites in this category about once a week during our 60day observation period. • Content / Event -driven.
Since we are interested in the interaction between the user’smind and the browser interface, we want to remove from our analysis all webpage visits thatoriginate from an information source outside of these two. To this end, we mark all hyperlinkbased visits (marked in our logs as having the transition type ‘link’) irrespective of theirfrequency rank as content-driven, and exclude them from our analysis. • UI exploitation.
The presence of clickable page icons obviously makes those webpagesmore accessible to users, over and above their mental propensity to visit these sites. Websitesthat users visit frequently, but use the UI to access, constitute an interesting sub-category ofthe habit category. For these webpages, it is not as clear as for others that repeated visitationis purely a function of a hedonic preference to visit them. The increased accessibility of thepages is clearly, but unquantifiably, also a factor. We refer to this category of pages, selectingall habit pages arrived at either via UI clicks or by typing in the URL bar while the specificpage icon was visible in the new tab page, as
UI exploit pages. • UI exploration.
When users visit rarely visited webpages via interaction with UI elements,we can infer that the UI interaction is driving them away from their typical behavior. Thismakes the subset of exploration pages that users arrive at from the new tab page (whetherwith transition ‘link ’or ‘typed ’), especially interesting. We call this sub-category
UI explore . • Pure exploration.
The final category we define is, expectedly, sparsely populated, represent-ing about 3% of all web page visits in our dataset. We define this as the subset of exploration pages whose icons are not displayed on the new tab page, have a transition type of ‘typed’and have been transitioned to directly from the new tab page. To be a member of this set,the user has to have entered the webpage URL (at least its first few letters, keeping in mindauto-complete capabilities of browsers) with no input from the browser UI. Visiting rarelyvisited sites without the (measured) influence of either other websites’ content or the contentof UI elements means that this subset of webpage visits reflect exploration driven purelyby memory considerations. Privileging the user’s mental content in a semantic sense, welabel this category pure exploration . Within this category, we make two further distinctions: manipulated pure exploration takes into account only typed transitions to rarely visited pagesfrom the new tab page.
Un-manipulated pure exploration takes into account typed transitionsto rarely visited pages from the address bar (or web search box) from an existing tab.Using this categorization of browsing events, we sought to quantify the category-wise change inusers’ web browsing behavior across the three different UI manipulations. Because the differentcategories defined above have vastly different base rates, it is most sensible to measure the changeacross UI modes in terms of the category-wise percentage change,100 × (cid:18) New- category- occurrence- frequencyDefault- category- occurrence -frequency − (cid:19) , where the category occurrence frequency is mode-sensitive and is simply the number of page visitsmarked in a particular category for a user divided by their total page visits while their browser’snew tab page’s behavior was in a particular mode .Figure 6 summarizes the variation in this quantity as a function of mode manipulations, averagedacross all subjects for all categories defined above. It is evident that behavior is not changed, inaggregate, as a consequence of these manipulations. The bulk of page visits in our dataset aretagged as habitual visits, and the ratio of habitual site visits changes insubstantially if at all acrossour users. Likewise, for content-driven behavior, which shows no statistically meaningful change. Journal, Vol. 1, No. 1, Article . Publication date: December 2018. ew tab page recommendations cause a strong suppression of exploratory web browsingbehaviors :11
Fig. 6. The change in each category’s occurrence fraction with respect to the same fraction seen in the defaultmode per subject, averaged across all subjects. Error bars represent ± S.D. Pure Exp UM and Pure Exp Mrespectively denote Pure Exploration Un-manipulated and Manipulated.
These null findings are reassuring. They suggest that our manipulations did not overtly changesubjects’ browsing behavior by introducing experimenter demand effects [6]. This is also congruentwith the survey results of low awareness of our manipulations shown in Table 1.If the presence of web page recommendations on the new tab page affects behavior by reinforcinghabitual browsing patterns, we expected to see that our exploration-linked categories would showpositive percentage changes with respect to the default condition when we remove all page iconsfrom the new tab display (the ‘empty UI’ condition) and when we present users with their leastvisited webpages as icons in the new tab display (the ‘least visited’ condition). As Figure 6 shows,this prediction is borne out clearly by our data. For both these conditions (least visited and emptyUI), exploration and pure exploration categories showed large positive changes ( ∼ p < . manipulated pure exploration category accountedfor almost all the change in the pure exploration category as a whole; there are only marginalincreases in the non-manipulated pure exploration category. The difference between these twocategories was only that the manipulated category page visits started with typing in URLs whilethe new tab page recommendations were visible, whereas the non-manipulated category page visitsstarted from other pages, with new tab recommendations not visible. Since all other aspects of theusers’ experience were constant across these two sub-categories, the stark difference in browsingbehavior emerged only by virtue of the new tab page recommendations being visible or invisible. Journal, Vol. 1, No. 1, Article . Publication date: December 2018.
12 Homanga Bharadhwaj and Nisheeth Srivastava
This performance dissociation allows us to confidently pinpoint the source of the difference inbrowsing behavior - the recommendations on the new tab page - with precision.Secondary observations include an interesting pattern seen in the difference between the defaultcondition, which uses Mozilla’s frecency algorithm [11], and a simple approximation of it that usesonly frequency information, ignoring recency - our ‘most visited’ condition. We find a substantialnegative impact on exploration for this condition with respect to the frecency condition, statisticallysignificant ( p < . of their own volition . However, thisis a statistically marginal effect when observed across the entire span of behavior captured in ourbrowsing logs. No more than 1% of all page visits are categorized as pure exploration, so even largerelative changes in this category’s representation are small shifts in user behavior in the aggregate.Thus, a more convincing analysis of the data would try to look for evidence beyond the purelystatistical to show that the effect we are postulating is real.One way to do this is to see if the pattern of behavior we see at the cohort level also showsup in individual subjects. Our hypothesis is a hypothesis about individuals, not populations. So ifthe results we are seeing are genuine and not the result of statistical fluctuations in the aggregatedatabase, the same pattern of increased exploration for the least visited and empty UI conditionsand reduced exploration for the most visited UI condition should hold across most of our subjects’individual data. Figures 7 and 8 show the percentage change in exploration and pure explorationfor each of our 80 individual subjects. It is immediately evidence that the within-individual changesacross the three conditions closely match the averaged changes observed across the sample. Thislends further credence to the view that, while subtle in aggregate, the influence of the UI manipula-tions affects user psychology at a personal level precisely in the directions we have predicted ontheoretical grounds.An interesting observation arises from the participants who reported to using their mobilephones (smartphones) / tablet computers for the experiment. These were the participants numbered10, 20, 26, 33, 34, 44, 68, 77, 78 and 79 . Rest of the participants used either a laptop / desktopcomputer for installing the extension we provided. As evidenced by Figures 7 and 8, their % changein exploration and pure exploration from the default mode to the ‘least visited’ and ‘blank UI’modes is quite high. In particular, the average change over these six individuals is around 10%higher for the ‘least visited’ mode and around 14% higher for the ‘blank UI’ mode for explorationover the aggregate for the entire sample shown in Figure 6. One of the reasons for a more acutesuppression of exploration in the default mode for mobile device users could be due to the smallerscreen sizes that incentive clicking on the web-page icons of the new tab page over typing in theaddress bar. However, since the number of individuals in our sample who used a mobile device isvery low, additional experimentation is needed to accurately quantify this distinction. Journal, Vol. 1, No. 1, Article . Publication date: December 2018. ew tab page recommendations cause a strong suppression of exploratory web browsingbehaviors :13
Exploration % C h a n g e f r o m d e f a u l t b e h a v i o r Fig. 7. Pattern of percentage change for overall exploration behavior for all the eighty individual subjects.Legends in top panel propagate through to all panels.
Journal, Vol. 1, No. 1, Article . Publication date: December 2018.
14 Homanga Bharadhwaj and Nisheeth Srivastava
Pure Exploration % C h a n g e f r o m d e f a u l t b e h a v i o r Fig. 8. Pattern of percentage change for overall pure exploration manipulated (Pure Exp M) behavior for allthe eighty individual subjects. Legends in top panel propagate through to all panels.
Journal, Vol. 1, No. 1, Article . Publication date: December 2018. ew tab page recommendations cause a strong suppression of exploratory web browsingbehaviors :15
Fig. 9. Temporal variation of exploration for one participant. The y-axis shows the % change in explorationfrom the averaged (across sessions) value in Default UI mode. The x-axis shows the flow of time in number ofsessions since the beginning of experiment. ( Blue - Default, Red - Least Visited, Green - Blank UI, Yellow - MostVisited ) DefaultSites LeastVisited MostVisited Blank UIDefault sites
X 41.08 ± ± ± Least Visited -36.68 ± ± ± Most Visited ± ± ± Blank UI -36.47 ± ± ± Table 3. Average ( ± SD) percentage change in exploration between last session of old mode and first sessionof new mode for all participants. The ( i , j ) th entry represents transition from mode i to mode j . While our experiment design is not meant to differentiate the possible influences of memoryinterference and motivational shifts based on changed presentations, some insight can be gleanedfrom the speed with which exploratory behavior changes with respect to UI changes. The part setcuing explanation for behavior change involves no learning or adaptation of preferences, whereasany sort of motivational effect on preferences would likely require learning over a somewhatextended time-scale. Table 3 documents the average percentage change across these mode tran-sitions, and shows that transitions from the default mode to either the least visited or blank UImode caused about 80% of the eventual change in exploration propensity to be manifest within thefirst session post-change. Such rapid transitions between exploratory behaviors contraindicatesa learning-based explanation for the phenomenon, and support priming or interference basedexplanations. Figure 9 shows the pattern of temporal variation in exploration of one individualover the course of the experiment.
Journal, Vol. 1, No. 1, Article . Publication date: December 2018.
16 Homanga Bharadhwaj and Nisheeth Srivastava
Forty two of our participants continued using our add-on for an extended period beyond the 60 dayexperiment window. We conducted some follow-up analyses on additional data we obtained fromthem to obtain a more complete picture of comparative behavior in Blank UI mode versus the Defaultmode. These participants were subjected to the default mode UI for three weeks continuously andthen to the Blank UI mode for seven weeks continuously.Using the data for default mode, we compute the base rate of exploration at different blocks oftime in the day per user. Now, we plot the percent change in exploration in the Blank UI mode overdefault mode for each time block averaged over a rolling window of 7 days for the total duration of7 weeks. Figure 10a shows this analysis averaged over all participants. The error bars representStandard Deviation upon averaging the means over all participants.We observe that the percentage-wise increase in exploration for the Blank UI mode over theDefault mode increases with time for all time blocks. This is indicative of the conditioning dueto tweaking of the new tab page UI from default mode to blank UI mode taking effect. This is aninteresting observation, and contrasts with the rapid onset of change seen above in Fig. 9. Combined,these observations appear to support a partial psychological role for both priming and conditioningin driving exploration suppression for web users, especially ones accustomed to using the defaultinterface for long periods of time. We similarly plot the percent change in exploration in the BlankUI mode over default mode for weekdays and weekends averaged over a rolling window of 5 daysand 2 days respectively for the total duration of 7 weeks. Figure 10b shows this analysis averagedover all participants. The error bars represent Standard Error of the Mean upon averaging themeans per week over all participants. Figure 10c shows another picture of the "Weekday/Weekend"analysis averaged over all participants. For each participant in each week, the exploration per dayis computed. The SD of exploration is computed for a 5 day block (weekday) and a 2 day block(weekend). This plot shows how this SD varies over time averaged across all individuals. The errorbars represent SD of the SD upon averaging across all participants.From Figures 10b and 10c we observe that the absolute value of increase in exploration for theBlank UI mode over the Default mode increases with time. This is similar to the observation madein Figure 10a. (a) Exploration with time of the day (b) Mean of exploration (c) SD of explorationFig. 10. Plot of % increase in exploration for Blank UI mode over Default mode averaged over 40 participants.(a) Variation of Mean of exploration over different time chunks of the day. (b) Variation of Mean explorationover weekdays/weekends (c) Variation of SD of exploration over weekdays/weekends
Our experiment shows that users’ behavior adapts to manipulations of the browser new tabdisplay along lines predicted by classic theories of memory interference, and that this adaptation
Journal, Vol. 1, No. 1, Article . Publication date: December 2018. ew tab page recommendations cause a strong suppression of exploratory web browsingbehaviors :17
Fig. 11. (Top row) Sample graphs with different ϵ values (Bottom row) Corresponding shortest paths betweenall node pairs in these graphs. The key point to note is that the connectivity of the graphs rises super-linearlyas a function of ϵ . is statistically significant both across and within individuals. We have not, however, establishedthat this adaptation is also practically significant. Prima facie , the changes in behavior are onlyapparent at the margins, not in the large aggregate of browsing patterns in our subjects, whichcomprises overwhemingly of habitual access to highly visited websites. Viewed from this aggregateperspective, the deviations we have managed to introduce are of the order of about 0 .
5% of all pagevisits. Why should this miniscule change in behavior be consequential? This is the question wehave sought to answer using the in silico experiments we report below.Our primary hypothesis in this phase of the project was that minor changes in explorationpropensity would affect the facility with which users can access crucial transition nodes in theirinformation network that connect them to fresh sources of information that they do not habituallyencounter during routine browsing. This hypothesis was inspired by related observations aboutthe role of social preference feedback in the polarization of internet communities - a phenomenonrecently popularized by the term ‘filter bubble’ [24].In classic filter bubble settings, the source of feedback and recommendations is external, buttuned to personal preferences. It is this personalization that, perversely, reduces the diversity ofinformation sources that the user acquires their social media-mediated knowledge from [24]. Inour case, unlike in conventional filter bubble settings, there are no external sources of informationand/or feedback. The browser UI is simply passively reflecting the habitual behavior of the user,modulo some modelling of the inter-temporal preferences for various forms of content of users.Nonetheless, we think the filtering effect of reinforcement of existing preferences works bythe exact same mechanism in our case, suppressing the user’s propensity to sample content fromdiverse sources. Since the user’s response to their own previous history is much more observablethan social influences seen in social media recommender systems, we can try to quantify the extentto which suppression effects of the size seen in our empirical data might influence the diversity ofinformation sourcing for model agents.To do this, we have modelled website browsing behavior for an individual agent as a randomwalk on a graph. Within the graph for each subject, each node represents a distinct website. Withthis abstract representation in place, the next step is to realize that the structure of the graph itselfmust reflect the Zipfian nature of webpage-website counts - users visits a small number of sites alarge number of times, and a large number of sites a small number of times [14].
Journal, Vol. 1, No. 1, Article . Publication date: December 2018.
18 Homanga Bharadhwaj and Nisheeth Srivastava
To capture these particular dynamics while maintaining analytic simplicity, we specialized thenetwork stucture we consider to be the family of all graphs with two pseudo-cliques whereinthe cross-links between these pseudo-cliques are controlled by an ϵ parameter such that with aprobability of ϵ , a link gets set up between two cross-pseudoclique nodes . When we simulatea random walk on any one such graph, the reasons for this construction become clear. Nodeswithin pseudo-cliques have several short paths connecting them, and so are likelier to be visitedby a random walker known to be situated at another node within the clique. Nodes in differentpseudo-cliques will have longer shortest paths on average, so depending on the distribution oflength of random walks for the agent, visits across pseudo-cliques will be rarer. Thus, we inducedifferences in accessibility of different nodes in the graph by assuming its structure to take thisparticular shape.Unlike more intricate agent-based models of browsing behavior, such as in [14], which tryto capture both within and across website browsing behavior, the model we use above makespredictions only for the quantity we are interested in - the across website browsing behavior of theuser. Dwell times and page depths within websites are not modelled in our approach. We offload asmuch of the preference information about browsing behavior as possible onto the graph structure,resulting in the simplest possible observer model for the user - a random walk on the graph -that retrieves the Zipfian cross-site browsing behavior we are interested in simulating. The keydifference between our simplistic model and the actual model of the web in terms of structureis that there are multiple pseudo-cliques in the web corresponding to different localizations ofinformation [8]. The results we demonstrate for two pseudo-cliques can be shown for any pairs ofpseudo-cliques without loss of generality, thus generalizing to more complicated web models asin [14].The core of our succeeding analysis is identifying the slope of the relationship between the ϵ parameter, which controls the number of cross-links between the two pseudo-cliques, and theaverage first hitting time between nodes in different pseudo-cliques. Such an analysis does notrequire that the graphs contain pseudo-cliques of only equal size, or that they contain only twopseudo-cliques. We consider the simplest subset that instantiates the distance asymmetry werequire. As a consequence of this distance asymmetry, we are able to map user behavior to networkstructure. Movements within a clique correspond to habitual behavior while a leap from one cliqueto another corresponds to exploration. Embedding the usually habitual, occasionally exploratorybehavior of users in this way in our graph model, we can say formally that there is on averageexploratory behavior ϵ of the time and habitual behavior 1 − ϵ of the time, assuming randominitial node initialization in our specific graph structure. The same general principle is expected tohold for a much more general family of graphs - graphs with k pseudo-cliques, interconnected bycross-links generated via a stochastic process mediated by the parameter matrix ϵ ij , i , j ∈ { , · · · k } . In this setup we are interested, fundamentally, in quantitatively characterizing the relationshipbetween changes in ϵ and changes in the average first hitting time between cross-pseudocliquenodes. The first hitting time between two nodes is simply the expected number of moves a randomwalker originating at node i will take to first reach node j across multiple random walks originatingat i . Since the UI manipulation affects ϵ by a large relative amount but a small absolute amount, weare interested in measuring the extent to which this small parametric change affects the visibilityof cross-pseudoclique nodes in the graph, which corresponds in our browsing user model to theaverage first hitting time. Users will access websites that lie within their radius of experience, as A pseudo-clique is defined as the graph structure obtained by removing some edges from a clique. We generate pseudo-cliques by first assuming an exponential degree distribution within a pseudo-clique of nodes, then assigning edges to nodesrandomly in decreasing order of degree, permitting ± ew tab page recommendations cause a strong suppression of exploratory web browsingbehaviors :19 measured by hitting time. First hitting times larger than the typical radius of experience - measuredin terms of walk length - will correspond to websites typically inaccessible to the user.Although theoretical properties of first hitting times on graphs have been investigated forrandom graphs [15, 32, 36], no direct results are specifically relevant for our analysis. Knowntheory suggests, on average, that adding extra nodes or edge to a graph increases the first hittingtime of any two nodes in the graph but such analyses are true only on average across randomgraphs [2]. We, on the other hand, are interested in graphs with a very specific direction of variationin structure, changing from barbell style graphs for low values of ϵ into well-connected graphsas ϵ rises. Figure 11 shows various graphs generated corresponding to different ϵ values. Theimportant thing to note is the rapidity with which adding cross-links changes sample graphs drawngeneratively using different values of ϵ from barbell style graphs (known to have worst case shortestpath distances) to well-connected graphs, with much smaller shortest path distances. Even thoughthe graphs with higher ϵ have more edges, we intuitively expect that the non-random manner inwhich the new edges are introduced by our graph generative model counteracts the mere fact thatthey are being added, and reduces the first hitting time. Fig. 12. Histogram showing the visit count of unique sites for the participants
The generative model for our graph has two free parameters, the size of the graph N and thecross-linking parameter ϵ . We fit this generative model to our data by fixing the size of the graphand the value of the ϵ parameter using the typical size of our users’ unique website repertoireand their measured exploration/habit ratios in our dataset respectively. The modal value of N (thenumber of unique sites visited by a user in a month) as inferred from the histogram in Figure 12 is 25.To approximate this, each of our simulated graphs used 20 nodes divided into two pseudo-cliquesof equal sizes. We generated graphs using ϵ values varied between 0 and 0.2 in steps of 0.005.Once an ( N , ϵ ) graph for a particular value of epsilon is ready, we empirically measured thehitting time of every cross-pseudoclique node pair by simulating random walks between all cross-pseudoclique node pairs. We calculated the average first hitting time for 100 such random walksacross all cross-pseudoclique node pairs for each of 100 ( N , ϵ ) generated graphs for each ϵ value. Journal, Vol. 1, No. 1, Article . Publication date: December 2018.
20 Homanga Bharadhwaj and Nisheeth Srivastava
Figure 13 shows the result of this simulation, plotting average first hitting times obtained from thisexperiment for different values of ϵ .We found empirically that the range of ϵ , measured as the ratio of exploration to habit pagesfor each subject, is 0 . − .
07 across our subjects. In this range, as illustrated in Figure 13 weobserve an approximately exponential decline in hitting time with increase in ϵ . This is because, asanticipated, the effect of the ‘cliqueish’ structure of the graph dominates over increase in number ofedges with ϵ . Gradually, as ϵ increases, the effect of increase in number of edges between pseudo-cliques starts dominating and hence the average hitting time asymptotes beyond ϵ ≈ .
1, and willlikely increase for still higher values. The key finding, also as anticipated, is that within the rangeof parametric values seen in our experiment, even small changes in ϵ cause large changes in thefirst hitting time. As we describe above, first hitting times represent website accessibility in ourmodel, so the result obtained here implies that small changes in ϵ cause exponential declines in theaccessibility of infrequently visited (cross-pseudoclique) websites for users. Fig. 13. Plot of average and % confidence bound of first hitting time for all cross-pseudoclique nodesempirically measured across 100 simulated random walks in each of 100 graphs generated for each value of ϵ .The best fit exponential for the data is also shown. Finally, the relationship between a finite random walk and the probability that it will pass throughany given node is straighforward to estimate. It is simply the probability that a random walk aslong as N might occur in the empirical distribution of first hitting times for that specific node.Hence, the p-value of a one-sided probability hypothesis test will give us exactly the probabilisticquantity of most interest to us - the probability with which a user with exploration propensity ϵ will access a diverse information source in any given browsing session of any particular estimatedlength.We identified individual sessions in our participants’ browsing logs using a 30 minute interval todetermine the boundaries of individual sessions (as mentioned in Section 2.3), and counting totalpage visits as one step in the random walk. Across all participants, the average session length cameout to be around 20 for our dataset. Next, we measured the kurtosis of the histogram of hittingtimes for each node pair in multiple random walks on a particular graph. We found that the kurtosisvalues themselves were well-represented by a normal distribution with mean 3.1 and SD 0.2. Hence,the first hitting time distribution for each node pair could be considered approximately Gaussian. Journal, Vol. 1, No. 1, Article . Publication date: December 2018. ew tab page recommendations cause a strong suppression of exploratory web browsingbehaviors :21
So, we ran one-sided z-tests testing whether a session length of k =
20 or lower might occurnaturally in the hitting time distribution of cross-clique node pairs. We ran this calculation for100 randomly sampled cross-pseudoclique node pairs across 100 graphs generated for each valueof ϵ and plotted the p-values obtained against ϵ . This relationship is visualized in Figure 14, andis approximately linear. This observation quantifies concretely the importance of large relativechanges in exploration propensity, even if exploration constitutes a small relative share of overallbrowsing behavior. While the change in behavior is on the margin in terms of aggregate browsingbehavior, it has non-marginal consequences on which information sources users will be able toaccess. Fig. 14. Variation of p-value of the average session length with exploration. p-value is averaged across multiplecross-pseudoclique node-pairs and graphs.
Since the beginning of internet browsing, designers have always tried to design browsers in waysthat reduce the user’s typing effort. In early browsers, this was done using history auto-completesuggestions. More recently, both OS and browser designers have sought to reduce typing effort stillfurther by adding apps to home screens and frequent (and recent) web pages to ‘new tab’ pages.The inarguable logic of such design is that, because website visit counts are approximately powerlaw distributed, being able to simply click on the most visited pages optimizes the number of typingresponses needed over the course of users’ use of the browser.But website visits are a means to an end, not an end in themselves. Ultimately, they are expressionsof our information preferences [25]. This paper makes the case that, by showing people the sitesthey visit most frequently over and over again in new tab displays, current practice in browser andOS UI design traps people into a solipsistic feedback loop, reinforcing their strongest preferencesto the detriment of weaker ones that potentially offer more scope for diversity of experience andlearning [29, 34]. We quantified the size of this feedback effect at its source - natural browsingbehavior - by collecting browsing history from subjects who consented to having the behavior of
Journal, Vol. 1, No. 1, Article . Publication date: December 2018.
22 Homanga Bharadhwaj and Nisheeth Srivastava their new tab page manipulated remotely via a plug-in they installed in their browsers. With insilico experiments, we further demonstrated the impact of the change in browsing behavior onpeoples’ proclivity for acquiring information from diverse sources.
Our work follows a rich vein of empirical research in human-computer interaction that is increas-ingly discovering subtle but powerful ways in which website interaction, access and even simpleUI decisions can influence human decisions. For example, posting to Facebook has been shownto increase users’ activity on the website immediately before and immediately after the posting,on the timescale of days [12]. More substantively, time of day of Twitter activity has been shown,strikingly, to correlate strongly with the probability of users being clinically depressed [7]. Evenmore strikingly, Epstein and colleagues have shown that presenting potential voters with searchresults for potential candidates in an actual national election in a UI with artificially manipulatedsearch result rankings can shift the voting pool’s vote fraction by up to 2% in the experimenter’schosen direction, a shift large enough to sway a reasonably close election [9]. There have also beenrecent works studying memory recall in simple tasks like clicking of pictures through differentcapture modalities [23]. We identify a large and consequential psychological effect: conventional UIdesign choices suppress peoples’ propensity to access diverse information sources while accessingthe web. This finding is not as specific as the ones documented previously in the literature, but alsoas a consequence, is likely affecting a lot more people at any given point in time by virtue of itsgenerality.Our results also relate to research efforts ongoing to characterize and surmount the recommendersystems (RS) filter bubble [4, 18, 28]. There is mixed evidence regarding the effects of recommendersystems on users’ consumption patterns. Whereas work focused on identifying changes in con-tent supply as a function of personalization tools has found no effects [13], other work studying consumption patterns have reported both negative [4, 10] and positive effects on consumptiondiversity [22]. Our work finds unambiguous and large negative effects on information diversitydriven by the use of recommendations. By our simplest measure of diversity - count of uniquewebsites visited - users operating browsers with blank new tab pages visited on an average 15%more unique websites over a two month period.One possible reason for previously reported null or positive results in the literature might bethat the active use of recommender systems like Movielens involves users actively looking forinformation they don’t already possess, an explicit cognitive task where they are aware that theymust evaluate the incoming information quasi-critically [30]. In contrast, constantly stoppingby the new tab page during transitions between websites is a much subtler phenomenon; therecommendations here are passive , in the sense that the ostensible function of these bookmarks isto facilitate access to frequently visited pages, not to shift preferences. People can reason through,or actively ignore, redundant information while they are actively deciding what to do. It is harderfor them to guard against the subtle impact of memory inhibition through visual presentation ofweb page icons time after time in the normal flow of their interaction with the browser [21].While it might seem superficially strange that such a small UI design component could potentiallyhave such a large impact on browsing behavior, behavioral economics offers striking examplesof similar subconscious nudges significantly affecting behavior for better and for worse [16].Similarly, Thaler and Sunstein offer striking examples of how small and often insignificant thingscan influence (nudge) behavior [17]. As McLuhan foresaw, but many modern empirical analysesignore, the peril in informative computer interfaces is not that they may feed us bad informationagainst our will, it is that the form in which they present information to use will subtly change ourexpectations of what we want to look for. The classic ideal observer in web browsing - Pirolli &
Journal, Vol. 1, No. 1, Article . Publication date: December 2018. ew tab page recommendations cause a strong suppression of exploratory web browsingbehaviors :23
Card’s information forager - models humans as observers optimizing information acquisition underinformation-theoretic constraints [25–27]. Our demonstration of large priming effects caused byas innocuous a source as the new tab page interface could help further nuance such models byincorporating the effect of psychological biases in the same way the influential cognitive biasesliterature has influenced the design of microeconomic models [35].Our description of the delicate interaction between mind and machine in this most humdrum oftasks - web browsing - should also stimulate further exploration of the negative externalities ofpersonalization, recommender systems and other methods of preference influence common in HCIapplications. At present, such influence is sometimes justified by discriminating between malevolentand benevolent deception in UI design, with the basic difference being that a benevolent deceptionis meant to benefit the user [1]. Our study presents a clear example of the inadequacy of suchethical definitions - in our case, the developer, in principle, gains nothing by presenting frequentlyviewed web pages as clickable icons; the UI is designed the way it is to minimize the user’s needto type. Nonetheless, this benevolent deception, conflating accessibility with preference via thesubconscious impact of presentation on memory, ends up affecting user behavior substantially, asour results show.
A natural limitation of this work is the size and representativeness of the population samplewe used to establish the basic fact that presentation of commonly used websites’ page icons onthe new tab display interferes with the retrieval of infrequently viewed websites. Nonetheless,the empirical results we describe in this paper are statistically significant by all conventionalstandards of measurements, and our sample is quite diverse with respect to gender, education, andoccupational status. While one can never have too many participants in a study making claimsabout internet browsing behavior, the large effect sizes we find, consistently seen both acrossand within participants, suggest that our study was adequately powered to discern the specifichypothesis it investigated.It is also possible to question how well conclusions from this study, with a mean sample age of28, may generalize to the experience of older adults. At the same time, it must be remembered thatwhile our sample is not representative of the general population, it is certainly more representativeof heavy internet users [3, 33], who are the population most likely to be affected by the UI designdecisions critiqued in this paper. We also note that the average age of our population is very closeto the median age of the country this study was conducted in, suggesting that there is no significantage-bias in our recruitment.Our results show that changes in the new tab page interface affect the diversity of users’ browsingexperience, and that this change is brought about specifically by interacting with the new tabpage. But they are unable to shed light on the psychological phenomenon underlying this effect.Whereas we initiated this experiment inspired by the analogy between web browsing and memoryretrieval, successful retrieval from memory requires both motivation and memory performance.Our experiment design does not differentiate these two variables. Thus, we cannot claim rigorouslythat it is memory impairment caused by competitive interference from displayed web page iconsthat is driving the difference in website browsing patterns across our manipulations. We can merelysuggest that this is likely to be true, to the extent that motivation to access different websitesdoes not change as a function of the UI display modes and offer as supporting evidence the factthat browsing behavior changed in the same direction and by about the same amount whetherinfrequent websites or no websites were shown on the new tab page. Identifying the mechanismby which the UI changes are transforming into behavioral changes is essential to identify usefulsolutions.
Journal, Vol. 1, No. 1, Article . Publication date: December 2018.
24 Homanga Bharadhwaj and Nisheeth Srivastava
In this paper, we contribute two basic results. One, we show empirically that subtle UI manipulationsof the new tab page in common browsers can create large relative changes in users’ propensity toexplore sites outside the repertoire of sites they habitually visit. Two, we show using a simulationstudy informed by empirically measured parameters that large relative changes in explorationpropensity manifest linearly in large absolute changes in the visibility of diverse informationsources in a user’s browsing experience. Together, these results present striking evidence that thecurrent design practice of showing frequently visited webpages on new tab pages is suppressingthe expression of web surfers’ exploratory tendencies on the web.Considering how ubiquitous computers and the internet have become, this effect would spaneven mobile devices. In fact, since small screen sizes in hand held mobile phones or tablets arelikely to bias users towards clicking the new tab page icons over typing in the address bar, theexploration suppression is likely to be exacerbated, as an admittedly small post hoc cohort analysisof our data shows. Further research is needed to quantify the extent of this difference.Finally, we note that all common browsers offer users the ability to use a blank page as theirnew tab page, making our conclusions immediately actionable. Based on our results, people shoulddecide for themselves whether the convenience of clicking through to commonly visited web pagesis worth the potential curtailment of their curiosity.
REFERENCES [1] Adar, E., Tan, D. S., and Teevan, J. Benevolent deception in human computer interaction. In
Proceedings of theSIGCHI conference on human factors in computing systems (2013), ACM, pp. 1863–1872.[2] Aldous, D. An introduction to covering problems for random walks on graphs.
Journal of Theoretical Probability 2 , 1(1989), 87–89.[3] Assael, H. A demographic and psychographic profile of heavy internet users and users by type of internet usage.
Journal of advertising research 45 , 1 (2005), 93–123.[4] Bakshy, E., Messing, S., and Adamic, L. A. Exposure to ideologically diverse news and opinion on facebook.
Science348 , 6239 (2015), 1130–1132.[5] Carr, N.
The shallows: How the internet is changing the way we think, read and remember . Atlantic Books Ltd, 2010.[6] Charness, G., Gneezy, U., and Kuhn, M. A. Experimental methods: Between-subject and within-subject design.
Journal of Economic Behavior & Organization 81 , 1 (2012), 1–8.[7] De Choudhury, M., Gamon, M., Counts, S., and Horvitz, E. Predicting depression via social media.
ICWSM 13 (2013), 1–10.[8] Dezsö, Z., Almaas, E., Lukács, A., Rácz, B., Szakadát, I., and Barabási, A.-L. Dynamics of information access onthe web.
Physical Review E 73 , 6 (2006), 066132.[9] Epstein, R., and Robertson, R. E. The search engine manipulation effect (seme) and its possible impact on theoutcomes of elections.
Proceedings of the National Academy of Sciences 112 , 33 (2015), E4512–E4521.[10] Flaxman, S., Goel, S., and Rao, J. M. Filter bubbles, echo chambers, and online news consumption.
Public opinionquarterly 80 , S1 (2016), 298–320.[11] Foundation, M. Frecency algorithm. Blog, Aug 2014. Retrieved September 19, 2017 from https://developer.mozilla.org/en-US/docs/Mozilla/Tech/Places/Frecency_algorithm.[12] Grinberg, N., Dow, P. A., Adamic, L. A., and Naaman, M. Changes in engagement before and after posting tofacebook. In
Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems (2016), ACM, pp. 564–574.[13] Haim, M., Graefe, A., and Brosius, H.-B. Burst of the filter bubble? effects of personalization on the diversity ofgoogle news.
Digital Journalism 6 , 3 (2018), 330–343.[14] Huberman, B. A., Pirolli, P. L., Pitkow, J. E., and Lukose, R. M. Strong regularities in world wide web surfing.
Science 280 , 5360 (1998), 95–97.[15] Ikeda, S., Kubo, I., and Yamashita, M. The hitting and cover times of random walks on finite graphs using localdegree information.
Theoretical Computer Science 410 , 1 (2009), 94–100.[16] Leonard, T. C. Richard h. thaler, cass r. sunstein, nudge: Improving decisions about health, wealth, and happiness,2008.[17] Leonard, T. C. Richard h. thaler, cass r. sunstein, nudge: Improving decisions about health, wealth, and happiness,2008.Journal, Vol. 1, No. 1, Article . Publication date: December 2018. ew tab page recommendations cause a strong suppression of exploratory web browsingbehaviors :25 [18] Liao, Q. V., and Fu, W.-T. Beyond the filter bubble: interactive effects of perceived threat and topic involvement onselective exposure to information. In
Proceedings of the SIGCHI conference on human factors in computing systems (2013), ACM, pp. 2359–2368.[19] McLuhan, M., MCLUHAN, M. A., and Lapham, L. H.
Understanding media: The extensions of man . MIT press, 1994.[20] Myrberg, C., and Wiberg, N. Screen vs. paper: what is the difference for reading and learning?
Insights 28 , 2 (2015).[21] Newell, B. R., and Shanks, D. R. Unconscious influences on decision making: A critical review.
Behavioral and BrainSciences 37 , 1 (2014), 1–19.[22] Nguyen, T. T., Hui, P.-M., Harper, F. M., Terveen, L., and Konstan, J. A. Exploring the filter bubble: the effect ofusing recommender systems on content diversity. In
Proceedings of the 23rd international conference on World wide web (2014), ACM, pp. 677–686.[23] Niforatos, E., Cinel, C., Mack, C. C., Langheinrich, M., and Ward, G. Can less be more?: Contrasting limited,unlimited, and automatic picture capture for augmenting memory recall.
Proceedings of the ACM on Interactive, Mobile,Wearable and Ubiquitous Technologies 1 , 2 (2017), 21.[24] Pariser, E.
The filter bubble: What the Internet is hiding from you . Penguin UK, 2011.[25] Pirolli, P., and Card, S. Information foraging in information access environments. In
Proceedings of the SIGCHIconference on Human factors in computing systems (1995), ACM Press/Addison-Wesley Publishing Co., pp. 51–58.[26] Pirolli, P., and Card, S. Information foraging.
Psychological review 106 , 4 (1999), 643.[27] Pirolli, P., and Fu, W.-T. Snif-act: A model of information foraging on the world wide web.
User modeling 2003 (2003),146–146.[28] Resnick, P., Garrett, R. K., Kriplean, T., Munson, S. A., and Stroud, N. J. Bursting your (filter) bubble: strategies forpromoting diverse exposure. In
Proceedings of the 2013 conference on Computer supported cooperative work companion (2013), ACM, pp. 95–100.[29] Schmidhuber, J. Formal theory of creativity, fun, and intrinsic motivation (1990–2010).
IEEE Transactions onAutonomous Mental Development 2 , 3 (2010), 230–247.[30] Sinha, R., and Swearingen, K. The role of transparency in recommender systems. In
CHI’02 extended abstracts onHuman factors in computing systems (2002), ACM, pp. 830–831.[31] Slamecka, N. J. An examination of trace storage in free recall.
Journal of experimental psychology 76 , 4p1 (1968), 504.[32] Sood, V., Redner, S., and Ben-Avraham, D. First-passage properties of the erdős–renyi random graph.
Journal ofPhysics A: Mathematical and General 38 , 1 (2004), 109.[33] Soule, L. C., Shell, L. W., and Kleen, B. A. Exploring internet addiction: Demographic characteristics and stereotypesof heavy internet users.
Journal of Computer Information Systems 44 , 1 (2003), 64–73.[34] Srivastava, N., Kapoor, K., and Schrater, P. R. A cognitive basis for theories of intrinsic motivation. In
Developmentand Learning (ICDL), 2011 IEEE International Conference on (2011), vol. 2, IEEE, pp. 1–6.[35] Tversky, A., and Kahneman, D. Judgment under uncertainty: Heuristics and biases. science 185 , 4157 (1974),1124–1131.[36] Uchiyama, K., et al. The first hitting time of a single point for random walks.