Screen Reading Regions in Social Media Comments: An Eye-Tracking Analysis of Visual Attention on Smartphones

Our study examines how formatted text influences reading behavior concerning the smartphone screen regions users utilize on social media comments. Aiming to research visual attention in scrollable content, we investigate the differences in the distribution of visual attention in screen regions between social media comment sections and the users’ preferred reading regions across all cases examined. An experiment was conducted with participants (n = 47) engaging in reading activities on the comment section of four social media -Twitter, YouTube (in two versions), Facebook, and Instagram- chosen due to their popularity and the users’ familiarity with them. Results showed that users’ visual attention is distributed differently between social media. Different reading styles were observed, with users not utilizing the entirety of the screen but instead focusing on specific screen regions that varied between users.


INTRODUCTION AND BACKGROUND
The shift from reading print media to consuming digital content has introduced changes in the reading behavior of people engaging with reading material.Contrary to the more in-depth and concentrated reading of traditional print, the screen-based reading behavior consists of "more time spent on browsing and scanning, keyword spotting, one-time reading, non-linear reading, and reading more selectively" [12].
Eye tracking technology captures and analyses eye movement to determine the user's visual attention [6].It has been extensively utilized in the field of Human-Computer Interaction (HCI) for usability [5] and user experience research [1], as well as an input method [14,15], among others.Additionally, the robustness of eye tracking technology has been demonstrated by its extensive use in other areas beyond HCI, such as economics [11], marketing [21], education [20], and Psychology [17].
Whereas there is a rich history in eye tracking techniques [4], there is ongoing research that aims at improving accuracy with [18] comparing algorithms for identifying fixations and saccades and work on adaptive algorithms [16] for better fixation, saccade, and glissade detection in eye tracking data.The eye tracking capacity to discern and record eye gaze patterns with high accuracy is ideal for our work which aims to identify how users interact with digital reading content.
One characteristic that distinguishes digital content from paper media is that the reading interaction forgoes turning pages for a more natural -in a mobile phone's context-vertical scrolling for presenting additional content onto the screen.The introduction of scrolling instituted a significant transformation in how users engage with and read content as it enabled them to reposition content -while maintaining the original formatting-within the physical constraints of the screen.In contrast, reading print media requires traversing the entirety of the page first before introducing additional content by turning to the next page.Scrolling allows the use of specific screen regions for reading as it enables the positioning of the content, and several studies are focusing on novel scrolling techniques that could influence the preferred screen reading regions.
Ishak and Feiner introduced content-aware scrolling (CAS) that considers the content's position and other document characteristics to create a path through the text that the user scrolls through that follows the natural sequence of the written text even in a two-column format [7].Furthermore, to allow the identification of points of high interest in a web page in long documents on smartphones, Kim et al. developed content-aware kinetic scrolling that delivers dynamically pseudo-haptic feedback in the form of friction around such content [8].Although the previous research focuses on enhancing the scrolling experience by introducing content awareness, another approach is utilizing the user's gaze as an input method.Kumar et al. work is based on the notion that the gaze incorporates additional information regarding the user's intention to scroll and explores the user's gaze as a primary and augmented input for improving scrolling [9].
Notwithstanding the ongoing research for enhancing and developing new scrolling techniques, there is some evidence that reading from web pages utilizing a scrolling format can adversely affect readers' understanding of complex topics, especially those with low working memory capacity [19].The finding mentioned above indicates that more research is needed to understand all facets of scrolling, and in our case, we focus on how scrolling content can influence the users' preferred screen reading regions.
Eye-tracking studies can identify users' visual attention and examine their reading characteristics concerning scrolling content, with several works focusing on desktop and mobile phones.Regarding research that focuses on preferred screen regions, [3] showed that the users' reading behavior of documents and web pages on desktop computers is to focus their visual attention in a preferred region and scroll content into this region for them to read.Furthermore, they found a correlation between the distance of the user's scrolling in the vertical axis and the height of the preferred reading area users utilized.Lorigo et al. indicated an "F" shape pattern that corresponds to users' visual attention on the screen during users' evaluation of results in search engines [13].
Shifting to mobile phones, Lagun et al. showed that "user attention is focused on the center and top half of the screen" in web results and concluded that "user attention behavior on mobile phones is very different from that on desktops" [10].Finally, Biedert et al. examined the visual attention of users with formatted text on mobile phones and, upon visual inspection of eye-tracking recordings, identified three types of readers with different reading characteristics in terms of preferred screen regions and stated that "not the full height of the screen was generally being used" [2].Most users that participated in the previous study exhibited a "blockwise" reading behavior where they utilized certain parts of the screen for reading (blocks) and scrolled new content inside; the rest of the users were equally split between reading the entire page before scrolling and reading line by line.
Additionally, there is some evidence that the reading task might influence reading behavior, since Buscher et al. examined two distinct reading tasks (shopping and reading) and found two types of reading behaviors; concentrated reading and scanning [3].
Our study examines how formatted text influences reading behavior concerning the screen regions users utilize on social media comments.The purpose of choosing such a task is that it is typical and familiar to most users and that the formatting employed has several text segments (comments around two lines long) that should enable users to operate at a certain part of the screen.More extensive content (more lines per comment that exceeds the vertical space of the screen) could force them to deviate from their reading style, as Biedert et al. observed that users tend to change their reading style when a paragraph does not fit on the screen [2].Four social media were chosen for this study due to their popularity and the users' familiarity with them: Twitter, YouTube, Facebook, and Instagram.YouTube was examined in two versions: one with the comment section in full screen mode and one with the comments as a user interface (UI) element below the video.
Our aim is twofold: to examine if there are differences in the distribution of visual attention across regions between the chosen social media and to corroborate previous findings with quantitative measurements regarding preferred screen regions in a different context (social media comments) that constitute the following research questions: RQ 1 : Are there any differences in users' distribution of visual attention in the screen regions examined between the chosen social media?RQ 2 : Are there any differences between users' preferred reading regions across all chosen social media regarding the screen regions examined?

METHODOLOGY
In this study, we followed a within-subject design, and each participant was exposed to 5 experimental conditions.The participants were instructed to read through comments on various social media platforms using a mobile phone, which was handed to them while their gaze was being recorded.The social media platforms selected for this experiment, including YouTube, Facebook, Twitter, and Instagram, were their June 2022 versions.The experiment was conducted at the Human-Computer Interaction Laboratory of our Department.A participation form -with a QR code-was handed out to the University to attract volunteers and inform them about the nature of the experiment and information for scheduling a session.A total of 53 participants participated in the experiment, 19 females and 34 males, aged between 18 to 57 (Mean = 24.6 and SD = 8.45).

Tasks
For each platform studied, we prepared a sample post and accompanying comments (Figure 1) in the participants' native language (Greek) that remained the same throughout the experiment; no comments were added or deleted, and their sequence remained unchanged.The comments' lengths were comparable on each platform (around two lines long), with only Facebook having a few lengthier comments (as is accustomed to Facebook's comments).
The experiment consisted of 5 tasks (Table 1), one for each platform (Facebook, Twitter, and Instagram) and two for YouTube, covering two cases: the comment section filling the entire screen (hereafter mentioned as YouTube full screen) or having the comments underneath the video area (hereafter mentioned as YouTube).For each task, the participants were instructed to read the post or watch the video first and then read through the comment section, trying to identify specific comments that contained a piece of information that was given prior to the task by the researcher.
The task for each condition was distinct and required careful reading of the comments to identify the correct one.In the screenshot of Facebook shown in Figure 1, the third comment on the comments page is the correct one, as the second sentence mentions that not a long while back, this person bought the same fridge.The eye tracking analysis has shown that all the participants that were included in the study have read all the comments as instructed, indicating that the task and the comments were complex enough to compel the participants to read all the comments carefully in order to identify the correct one.
A pilot study was conducted (n = 4) to enable the evaluation of the experiment procedures, such as completion time, number of comments in each task, and the complexity of instructions given in

Social media
Task description Facebook Read all the comments carefully and "like" the comment where the user mentions that they own the same fridge.

Instagram
Read all the comments carefully and "like" the comment where the user says they disagree with the race's outcome.Twitter Read all the comments carefully and "like" the comments (there are more than one) where the users suggest an ingredient change to the plate of the photo.YouTube (full screen) Read all the comments carefully and "like" the comment where the user mentions that they prefer another competing company.

YouTube
Read all the comments carefully and "like" the comment where the user mentions that they did a similar build for other furniture besides the shelf.
each task.The data captured from the participants during piloting was excluded from the results.

Apparatus
A Samsung Galaxy S20 FE with a 6.5-inch display and a resolution of 2400x1080 pixels was used as the mobile device.The phone was attached to a stand, as shown in (1) of Figure 2.For the eye-tracking data collection, we used Tobii Pro Glasses 3, a wearable eye-tracking device with a Full HD resolution scene camera with a 106°field of view and two eye-tracking cameras per eye.The experiment was recorded with Tobii Pro Lab that accompanies the eye-tracker.

Procedure
In the beginning, each participant was informed about the experiment and asked to sign a consent form.Afterward, participants had to put on the eye-tracker and find an appropriate position across the stand where the mobile phone was attached.Then, for each participant, the phone's height and angle were individually modulated, and they were encouraged to find a natural position to hold and read from the phone, followed by the eye-tracking calibration process.
Participants had to go through all 5 tasks in random order, with each task having a specific description that remained the same for all participants.After completing each task, participants had to repeat the calibration process to continue with the next task, ensuring accurate data by having separate recordings.
A task was considered complete when the desired comment was found or when there were no more comments to read.At the end of the experiment, participants were asked to complete a questionnaire about the procedure, with the total duration of the experiment at 15-20 minutes for each participant.

Data Cleaning and Statistical Analysis
The eye-tracker records a video file and the raw eye movement data points during each recording according to its sampling rate.Each video was examined with the overlaying raw data for each recording using Tobii Pro Lab, and invalid data were identified and removed.This process resulted in the removal of 6 participants, while data for some tasks were also removed from the remaining 47 participants.As a result, the following are the total numbers of valid data for each social media: 44 for Twitter, 41 for Instagram, 39 for Facebook, 40 for YouTube (full screen comments), and 42 for YouTube.
Figure 2: A user reading comments while wearing eye-tracking glasses (1), the specified AOI when the screen is divided into two regions (2), and the AOI when the screen is divided into three regions (3).

Areas of Interest and Metrics.
A quantitative analysis of the users' visual attention was conducted by assigning Areas of Interest (AOI) and dividing the screen into separate regions.Two separate cases were examined, with the screen divided into two parts (top and bottom half, (2) of Figure 2) and divided into three (top, middle, and bottom thirds, (3) of Figure 2).Only the comment section was included in the AOI on each social platform examined while also excluding static parts of the UI, such as the top and bottom navigation bars.Afterward, the Tobii software calculated the desired metrics within the AOI that were defined in the previous steps, and the measurements were exported in an Excel report.The results of the following section are derived from the analysis of the total visit duration1 of each AOI, which is the total time that the user visited each with their gaze, and was conducted with IBM SPSS Statistics v28.0.

RESULTS
This section is comprised of two parts; the first presents the results of the comparisons between all social media and the respective screen regions regarding the percentage of visit duration (RQ 1 ).It comprises ten groups of social media comparisons (all unique combinations of the 5 cases tested) and the comparisons between the corresponding five screen regions examined.The second section presents the results regarding differences between users' preferred screen reading regions (RQ 2 ).

Results Regarding RQ 1
Of 47 participants who conducted the experiment, 29 of them had valid data in all five social media examined, and the Shapiro-Wilk tests showed that the normality assumption was violated in all visit duration percentages of all regions examined for each social media.Nonparametric tests were performed with the Friedman test showing that there are differences in the visit duration percentages between social media in the first region when the screen is split in three, x 2 (4) = 10.608,p = 0.031, and no significant differences were found for other screen regions; top half (x 2 (4) = 6.785, p = 0.148), bottom half (x 2 (4) = 6.785, p = 0.148), middle third (x 2 (4) = 9.159, p = 0.057), and bottom third (x 2 (4) = 7.969, p = 0.093).
The next sections provide the results of the post-hoc pairwise comparisons of each pair of social media for each region examined, as depicted in Figure 3. Since the comparisons were conducted with only two social media, the number of participants with valid data in each pair is higher, which delivers more accurate results.

Comparing Twitter with
Instagram.Shapiro-Wilk tests indicated that the normality assumption was violated for all regions and all social media, with the exception of Instagram's middle third screen region (W(39) = 0.958, p = 0.155) that followed a normal distribution; thus, nonparametric tests were performed.A Wilcoxon Signed-Rank test showed significant differences in the percentage of visit duration between Twitter and Instagram in the middle third screen region, n = 39, T = 249, z = -1.968,p = 0.049, r = 0.32 (blue bar, Figure 3).Wilcoxon Signed-Rank tests were performed to compare Twitter and Instagram for the rest of the regions but showed no significant differences.

Comparing
YouTube with Facebook.Shapiro-Wilk tests indicated that the normality assumption was violated for all regions and all social media; thus, nonparametric tests were conducted.A Wilcoxon Signed-Rank test showed significant differences in the percentage of visit duration between YouTube and Facebook (n = 37) in the top half screen region T = 202, z = -2.058,p = 0.04, r = 0.34 (yellow bar, Figure 4), bottom half screen region, T = 464, z = 2.058, p = 0.04, r = 0.34 (red bar, Figure 4), and the top third screen region, T = 116, z = -2.587,p = 0.01, r = 0.43 (purple bar, Figure 3).Wilcoxon Signed-Rank tests were performed to compare YouTube and Facebook for the rest of the regions but showed no significant differences.

Comparing YouTube (full screen comments) with Facebook.
Shapiro-Wilk tests indicated that the normality assumption was violated for all regions and all social media; thus, nonparametric tests were performed.A Wilcoxon Signed-Rank test showed significant differences in the percentage of visit duration between YouTube (full screen comments) and Facebook (n = 37) in the bottom third screen region, T = 152, z = -2.094,p = 0.036, r = 0.34 (yellow bar, Figure 3).Wilcoxon Signed-Rank tests were performed to compare  YouTube (full screen comments) and Facebook for the rest of the regions but showed no significant differences.

Comparing YouTube (full screen comments) with Instagram.
Shapiro-Wilk tests indicated that the normality assumption was violated for all regions and all social media with the exception of Instagram's middle third (W(35) = 0.952, p = 0.132) screen region that followed a normal distribution; thus, nonparametric tests were conducted.A Wilcoxon Signed-Rank test showed significant differences in the percentage of visit duration between YouTube (full screen comments) and Instagram (n = 35) in the middle third screen region, T = 172, z = -2.342,p = 0.019, r = 0.4 (green bar, Figure 3) and the bottom third screen region, T = 320, z = 2.664, p = 0.008, r = 0.45 (red bar, Figure 3).However, regarding the percentage of visit duration, the rest of the regions showed no significant differences when Wilcoxon Signed-Rank tests were performed to compare YouTube (full screen comments) and Instagram.
3.1.5The Rest of the Social Media Comparisons.For the comparisons of Twitter with YouTube, Twitter with YouTube full screen comments, Twitter with Facebook, YouTube with YouTube full screen comments, YouTube with Instagram, and Facebook with Instagram, nonparametric tests were conducted as the Shapiro-Wilk Tests showed that our data deviated significantly from normality and results showed that there were no significant differences when comparing the same regions for each of the social media groups.

Results regarding RQ 2
If we examine the visiting time percentages of individual screen regions, the data show that some participants have not visited certain areas with their gaze.Calculating the mean visiting time percentages of each user in all social media for each region revealed that several participants are near 0% and a few close to 100%.Table 2 presents the number of participants with a mean visit duration percentage (across all social media) of less than 5% and more than 95%.This distinction can reveal if some users are using only specific parts of the screen while reading scrollable content in the social media examined.A notable mention regarding the table is that the top and bottom thirds have 15 and 16 cases, respectively, revealing that a relatively high number of users, from 47 in total, refrain from using these regions.

DISCUSSION
The statistical analysis of the visit duration data revealed significant differences between some social media while investigating the participants' distribution of visual attention in the screen regions examined.Specifically, when examining the top and bottom half of the screen, only one comparison (YouTube vs Facebook) revealed significant differences in the distribution of visual attention, with most of the social media -except Facebook-having virtually evenly split percentages of visit duration between regions.Facebook is an outlier, with 40.7% of the mean visit duration percentage being in the top half and 59.3% in the bottom half, suggesting that the participants focused primarily on the bottom part of the screen.One possible explanation may lie to the fact that Facebook had a number of comments that were larger compared to the other social media, some of them having multiple lines per post, and this factor might have influenced the results, suggesting that users would not scroll until they had read the entire post while also indicating a shift of the reading area bellow the usual region that was utilized in the other social media.Having only one comparison revealing differences might indicate that if there are differences in the reading behavior, splitting the screen into two reading zones (top and bottom) is not precise enough to reveal all the potential differences.
When examining the results of the participants' distribution of visual attention with the screen divided into three parts, the comparisons reveal 5 cases with significant differences between the social media examined.The differences found are related to all screen regions (top, middle, and bottom thirds), with 1 case for the top third and 2 cases for each of the other thirds.In addition, all social media had at least one screen region that had significant differences with another, with Facebook having 2, while YouTube (full screen comments) and Instagram had 3.
These results indicate that there is a notable change in the distribution of visual attention in screen regions of the users when interacting with different social media, and it could be related to multiple reasons that need to be further examined.For example, each social media's user interface and specifically UI elements other than plain text, such as the button design and the padding between elements, might have influenced the results since all social media examined have some UI design differences.Additionally, these findings might indicate that there are inherited differences in how each social media is utilized by the users that are not related to UI.A notable finding that to some extent substantiates the previous points is that YouTube with comments below the video compared to YouTube with comments in a full screen mode did not reveal differences in any screen region, indicating that a lower vertical dimension (a reduction of around 25%) did not influence the reading behavior when other UI elements were identical.
Regarding the results that indicate that a relatively high number of users were not utilizing some screen regions for reading, we can infer that these users scrolling and reading behavior differed from the rest of the participants.Specifically, 16 participants did not use the top third region of the screen (had less than 5% mean visiting duration percentage), and 15 did not use the bottom third region of the screen, suggesting that users exhibiting this behavior were evenly split.While this might seem like a small percentage of users, combining them reveals that 31 of the 47 participants were not using a third of the screen for reading.Not using a part of the screen suggests that the users have a preferred reading screen region and that when they have finished reading the region's content, they prefer to scroll additional content into this region instead of reading above or past it.These findings support the observations made by Biedert et al. that most users exhibited blockwise reading behavior [2].

CONLUSSION AND FUTURE WORK
The paper examines the differences in the distribution of visual attention in screen regions between 4 social media (Twitter, YouTube, Facebook, and Instagram) while also investigating the differences between users' preferred screen reading regions.An experiment was conducted by dividing the screen into two (top and bottom) and three regions (top, middle, and bottom), respectively, and capturing the visit duration for each user.Regarding the comparisons of the social media examined, the analysis of the results revealed differences between the platforms examined in the preferred reading regions, with the division into three regions revealing a higher number of cases that differentiate from another.Concerning the evaluation of users' preferred screen reading regions, results showed that most users do not utilize the whole screen for reading but opt to use the top and middle or the middle and bottom part of the screen.
Our work provides a quantitative analysis of the users' visual attention, including more participants, compared with previous works that focused on qualitative analysis and a smaller sample of users.Furthermore, the results of our study substantiate previous findings that observed diverse reading styles and users not utilizing the entirety of the screen, which was also validated in the context examined (social media).Specifically, our results showed that users tend to focus on certain regions of the screen -that can differ between the platforms examined-which could impact how content is presented and delivered on social media.For example, if users tend to focus more on the top and middle regions of the screen, social media platforms could consider highlighting or prioritizing content in those areas.Finally, since variations regarding the preferred screen regions are observed between users, social media platforms could provide a way for the user to adapt certain UI elements to their preferences (e.g., a user who prefers reading in the top region could choose that the comments should be in this region and other "secondary" UI elements could be in other regions).
A limitation of this study is that while we had a substantial number of participants that shed light on user reading behavior, it might not be sufficient to make broad conclusions.Additionally, our findings only apply to the social media platforms and tasks that were examined and might not be applicable to other platforms.Future works include examining whether these findings are true in other tasks containing scrollable content and investigating the cause of differences in seemingly similar UIs (the social media examined).The results of our work can incentivize future research in these areas and have the potential to inform future designs regarding the comment sections.

Figure 1 :
Figure 1: Snapshots of the social media comments section that were used in the study (in Greek).

Figure 3 :
Figure 3: The figure combines descriptive statistics (mean visit duration of the top, middle, and bottom third regions) with the results of the hypothesis testing (the bar-connected regions indicating significant statistical differences) of all social media comparisons.

Figure 4 :
Figure 4: The figure combines descriptive statistics (mean visit duration percentage of the top and bottom half regions) with the results of the hypothesis testing (the bar-connected regions indicating significant statistical differences) of all social media comparisons.

Table 1 :
Task description for each of the examined social media

Table 2 :
The number of participants (n = 47) that have less than 5% or more than 95% mean visit duration percentage across all social media