This is a self-archived – parallel published version of this article in the 

publication archive of the University of Vaasa. It might differ from the original. 

Developing Persona Analytics Towards Persona 

Science 

Author(s): Salminen, Joni; Jung, Soon-Gyo; Jansen, Bernard 

Title: Developing Persona Analytics Towards Persona Science 

Year: 2022 

Version: Accepted manuscript 

Copyright © Authors | ACM 2022. This is the author's version of the work. It is posted 

here for your personal use. Not for redistribution. The definitive Version 

of Record was published in Proceedings of IUI '22: 27th International 

Conference on Intelligent User Interfaces, 

http://dx.doi.org/10.1145/3490099.3511144. 

Please cite the original version: 

 
Salminen, J., Jung, S-G. & Jansen, B. (2022). Developing Persona 

Analytics Towards Persona Science. Proceedings of IUI '22: 27th 

International Conference on Intelligent User Interfaces, 323-344. New 

York: Association for Computing Machinery. 

https://doi.org/10.1145/3512891 

 

 



Developing Persona Analytics Towards Persona Science 

Joni Salminen 

Qatar Computing Research Institute, Hamad Bin Khalifa University, Doha, Qatar; and University of Vaasa, Vaasa, 
Finland, joni.salminen@uwasa.fi 

Soon-gyo Jung 

Qatar Computing Research Institute, Hamad Bin Khalifa University, Doha, Qatar, sjung@hbku.edu.qa 

Bernard J. Jansen 

Qatar Computing Research Institute, Hamad Bin Khalifa University, Doha, Qatar, bjansen@hbku.edu.qa 

Much of the reported work on personas suffers from the lack of empirical evidence. To address this issue, we 

introduce Persona Analytics (PA), a system that tracks how users interact with data-driven personas. PA captures 

users’ mouse and gaze behavior to measure users’ interaction with algorithmically generated personas and use of 

system features for an interactive persona system. Measuring these activities grants an understanding of the 

behaviors of a persona user, required for quantitative measurement of persona use to obtain scientifically valid 

evidence. Conducting a study with 144 participants, we demonstrate how PA can be deployed for remote user 

studies during exceptional times when physical user studies are difficult, if not impossible. 

Keywords 

Personas, User research, Persona science, Persona analytics, Remote user studies 

1 Introduction 

The current work discusses how personas can be effectively combined with the concept of analytics, i.e., the use of 

end-user data for drawing insights into human factors [53]. We start by defining the key concepts, and we then 

explain our approach to infusing personas with analytics, which we denote as Persona Analytics (PA). We refer 

to users when we mean stakeholders that use personas for decision making (e.g., designers, software developers, 

marketers). For other terminology, we refer to end-users when we mean people on whose information personas 

are based. 

Personas, commonly applied in human-computer interaction (HCI) [16], design [8], and business domains such 

as marketing and sales [74], are fictional depictions of end-users, patients, customers, or other groups of interest 

[16]. Personas convey end-users’ needs and requirements [8], alleviate decision-makers’ self-referential bias [6], 

and enable thinking of end-users even when none are physically present [66]. Also, personas give a human face to 

analytics data [29,35], humanize segments [11], give design inspiration [57], help compare end-users [34], and 

facilitate prioritizing end-user needs [69] for system development [67]. Persona profiles show relevant 

information about end-users [58] (see Figure 1). Nielsen summarizes a good deal of personas literature on their 

creation, assessment, and use [57]. Personas are easily digestible snapshots of end-users, audiences, or customers 

for use throughout an organization. 



 

Figure 1: Example of a persona profile. Persona Analytics enables tracking users’ mouse and gaze 
interactions as the user engages with different information elements (A: Picture, name, and text 

description, B: Audience size, C: Sentiment, D: Social media quotes, E: Topics of interest, and F: Most 
viewed contents). 

Personas are increasingly being enriched with quantitative data [72], or their creation is partially or 

completely carried out by algorithmic processes, which is referred to as algorithmically generated persona 

development [50]. When quantitative data becomes part of the created persona profiles, these profiles start 

approaching other analytics systems in terms of producing end-user metrics. In fact, algorithmically generated 

personas can be seen as an alternative method of end-user understanding, as alternatives to UX analytics tools 

(e.g., Google Analytics, Adobe Analytics, HubSpot, Mixpanel, Crazy Egg, and so on), when it comes to design tasks 

that require user insights (e.g., understanding website visitors for improving usability). As such, personas 

personify the numerical data on end-user characteristics and behaviors—turning numerical reports into persona 

profiles [29]. 

This transformation from “cold numbers” into “warm people” has been denoted as a benefit of personas, in 

that personified end-user data is treated more empathetically than nameless, faceless numbers [74]. 

Algorithmically generated personas are also supported by the automation of data science pipelines—i.e., the 

process of persona creation can now be automated—as well as web technologies [52] that enable serving the 

personas to users via web browsers—i.e., via interactive persona systems [33]. As such, PA can be defined as 

follows: 

DEFINITION 1: Persona analytics refers to decision-makers (i.e., persona users) in organizations using 
personas as analytical tools to better understand their end-users or other groups of interest. 

The above definition follows the conventional understanding of algorithmically generated personas using 

quantitative data [29,30,52]. As mentioned, we define a ‘user’ as someone who uses a persona for a professional 

task, which can relate to software development, design, marketing, or any other domain where personas are 

applied. Therefore, users can be software developers, designers, marketers, or other stakeholders involved in 

user-centric decision making. Because of this connection between personas and users, there exists another aspect 

to the concept of PA. Namely, PA can be seen as a research instrument to generate lasting knowledge about 

personas and how users interact with them. Its role is to grant HCI researchers a systematic approach for 

collecting data about persona user behavior and metrics for dealing with this data. In so doing, PA paves the way 

to more effective application of persona science, defined as ‘the use of empirical scientific methods, such as 

experiments, to produce robust and generalizable information about persona creation, evaluation, use, and 

impact’. 



Indeed, in prior research [36,37], PA is defined as the systematic measurement of behaviors and interactions of 

persona users engaged with interactive persona systems. This is consistent with the second definition we put 

forth: 

DEFINITION 2: Persona analytics refers to how researchers investigate the behaviors of persona users. 

It is this second definition that motivates our current work, because examining persona users’ engagement 

with personas can generate vital insights for persona science and the design of personas and persona systems that 

better serve stakeholders’ information needs about end-users or customers. The persona research urgently needs 

a strong empirical orientation to produce knowledge that is believable and can truly push forward the boundaries 

of personas practice and theory and add up to a coherent understanding of the persona user. Advocates of the 

scientific method in persona research [4,5,11,12,23] have continuously mentioned the lack of empirical 

experiments and quantitative measurements as a bottleneck for progress in terms of theory and practice. 

Our definition of persona science implies not only collecting data and conducting research on personas but 

also making an effort to devise theories that explain the data and guide further data collection. Persona science 

deals with real user behavior and formulating theories that are relevant to the design of personas. The focus in 

these efforts lies in the study of the persona users, which we demonstrate in this work by introducing new tools 

for measuring persona user behavior. To this end, the current work concerns itself with the development of a 

novel PA system embedded within a persona system, with the purpose of more effectively 

investigating/researching the behavior of persona users. Three research questions (RQs) are posed: 

• RQ1: How to implement PA in an interactive persona system? 

• RQ2: What kind of research questions can PA address? 

• RQ3: How can PA be used for understanding persona user behavior? 

The goal of the current work is to report efforts of building analytics features into an interactive persona 

system. We demonstrate the capabilities of this PA system and discuss its value for empirical persona research. In 

practice, PA can assist in designing layouts, features, and information content in algorithmically generated 

persona profiles. To achieve these benefits, it is necessary to incorporate analytics into personas, so that the 

interaction between the users and personas (and interaction features) can be captured. Previous efforts of this 

work appeared in [36,37] – relative to these, the current work adds a full-scale case study demonstrating the 

system capabilities with a real user study (previous research only tested the system with one pilot user). The PA 

system has implications for researchers and practitioners who are increasingly adopting web-based tools for 

remote testing since social distancing hampers in-person user studies [17]. This trend is likely to continue as tools 

and practices for remote user studies evolve. 

2 Related Work 

Algorithmically generated Personas. Although quantitative personas were first created within software 

requirements engineering [7,8], the concept of personas being data-driven was introduced by McGinn and 

Kotamraju [50] and later deployed by others [40,41,51,95,97]. Though, the idea of using “data” for personas dates 

to Cooper’s [16] concept that personas should be based on real user goals instead of fiction. While data 

orientation has remained a consistent theme in the persona literature [16,17,32,9,10,50,51], three trends 

contribute to the rise of algorithmically generated personas [29,72]: (1) availability of user and customer data 

from online analytics and social media platforms; (2) democratization of data science tools and algorithms that 

enable automated persona generation; and (3) web technologies that remove the limitations of static personas via 

interactive user interfaces. These trends denote a shift from unchanging “flat file” personas into dynamic “full-

stack personas” that update automatically and are traceable to individual user-level data [29]. 

Interactive Persona Systems. From algorithmically generated personas, the next logical step of evolution is 

interactive persona systems [3,52,72], defined as interactive user interfaces (UI) that display persona profiles. This 

UI can, but not necessarily always, be accessed via web browsers [32,33,35,38]. The benefits of web technologies 



are their broad applicability and accessibility. Personas served via the web can be accessed virtually from 

anywhere using any device that supports web browsing (see Figure 2a). Supporting technologies, such as user 

account management, can be integrated with relative ease using standard libraries and best practices. 

Interactivity refers to users performing various actions on the personas, such as analyzing information on gender 

distributions, refreshing the persona quotes, filtering the quotes by sentiment and topic [80], predicting a 

persona’s interest for a given topic [2,3], and engaging in dialogue [45,47]. The interactive features are enabled by 

standard Web technologies, such as HTML, CSS, and JavaScript. 

Emerging opportunities in Literature. Following these developments in algorithmically generated personas 

and interactive persona systems that have been described as transformational [52], multiple opportunities can be 

envisioned. We highlight five such opportunities. First, (i) interaction techniques and multimedia (e.g., persona 

chat/dialogue systems [13], video, AI agents [87]…) could be incorporated into persona systems to serve various 

end-user needs [75]. Second, (ii) new features for comparing personas by design goal metrics, such as diversity 

[74] and inclusivity [21], could be added. Third, (iii) personas could be integrated into an external system to 

enable persona-based recommendations [46], content management, and customer relationship management, as 

well as facilitating online advertising [79] via application programming interfaces (APIs) [38]. Fourth, (iv) 

developers could provide explainability, transparency, and context, which are important when applying 

algorithms for persona creation [80,88], as illustrated in Figure 2b. Finally, (v) interactive systems can be used to 

drill down to the persona information and make quantitative predictions [3]. 

  

(a) (b) 

Figure 2: (a) Interactive persona features, such as [A] browsing the available personas, [B] searching and 
sorting by user-defined criteria, [C] explanatory tooltips, and [D] export of usage logs. (b) Adding 

transparency to algorithmically generated personas. The first layer shows Mamdouh, a young Egyptian. 
The second information layer, accessible by clicking a chart icon, shows that Mamdouh actually comprises 

many demographic groups, of which [Egypt, 25-34 Male] is deemed the most representative by the 
algorithm 

Research Gap. While technology introduces novel opportunities for user-to-persona interaction, at the same 

time, these trends create an opportunity for better understanding of how persona users, such as designers, 

software developers, and marketers interact with personas. This better understanding of persona user behavior 

can lead to substantial advances in persona science (i.e., the academic study of personas and their usage), but it 

requires effective implementation of measurement. The lack of empirical persona user research has been noted 



by several researchers [48,72,77]. The unifying factor behind these possibilities is the need for understanding the 

persona user behavior, which requires measurement. In our solution, this measurement capability is provided by 

PA. 

3 Methodology for Persona Analytics 

3.1 Requirements Journey 

There is no standard method of building an analytics system. In our case, all the researchers had extensive 

experience of both Web analytics and personas research, which was instrumental in this process. This experience 

consisted of working with industry-leading analytics solutions, such as Google Analytics, for more than a decade 

in the case of two authors and half a decade in the case of one researcher. The persona research of the authors, 

when combined, also extends well beyond a decade and mostly consists of empirical work. Therefore, we had a 

vision of what we wanted to accomplish, what is missing from current research and practice, and what research 

questions in persona science should be addressed via empirical data. 

We started out by “drinking our own Kool-Aid,” i.e., by defining the ideal user persona for the system. This 

“persona” of a user of the PA system is a researcher that wants to conduct persona user studies in order to address 

scientifically important questions. Measuring user behaviors helps researchers tackle open research questions in 

persona science, which is the goal of this persona. To support this persona, a few requirements are posed: (a) the 

data must be accurate so that proper conclusion can be drawn from it, (b) there should be the possibility to 

include several data types to enable the comparison of different end-user inputs, and (c) the dimensionality of the 

data should not be overwhelming for the analysis task, i.e., the data needs to be exportable in a format that is 

relatively easy to analyze. 

These desiderata were considered in the design of PA by incorporating multiple data sources and by keeping 

the reporting data granularity at a user-friendly level – in other words, reports with different levels of detail and 

aggregation are provided, as explained later in this manuscript. 

Second, we brainstormed the type of questions that the PA system would need to be able to address by 

providing data for the researcher persona. The following list of scientific questions of interest (SQ) was 

collaboratively obtained among the research team members, pertaining to various persona aspects (in 

parentheses): 

• SQa: How do users interact with persona profiles? (interaction techniques) 

• SQb: What interactive features facilitate users’ discovery of personas for a given task? (interaction 

techniques) 

• SQc: What information of personas do users pay attention to? (information design) 

• SQd: What persona information influences users’ design choices and how? (information design) 

• SQe: What persona information influence users’ behaviors or attitudes about end-users? (persona 

perceptions) 

• SQf: How do users compare personas for a task? (cognitive styles, information processing) 

• SQg: How and why do users choose a persona for a given task? (cognitive styles, information processing) 

• SQh: How do users or user groups differ by their persona use? For example, are persona users with less 

experience in personas using them differently? Are there gender differences? (demographic, cultural, 

and social factors) 

Addressing these and other vital questions can provide much needed direction for persona science, addressing 

aspects of persona creation, validation, use, and value in use. Empirical, scientific inquiry is not only needed to 

produce valid knowledge for practitioners using personas, but it is also required to create robust theories on 

personas and their users. Aligned with principles of scientific inquiry, persona research can benefit from adopting 

more rigorous research designs, including hypothesis formulation based on theories in HCI, information science, 

social psychology, and other fields tangential to personas; and followed by systematic testing of those hypotheses, 



then revising the theory to adapt to persona context. Addressing these questions can help persona creators 

understand which aspects of human-to-human interactions apply to human-to-persona interactions, so design 

decisions can be made to mitigate unwanted effects (e.g., stereotyping [48] and seeing personas as irrelevant, 

abstract, or misleading [49]) as much as possible. 

3.2 Defining Metrics for Persona User Behavior 

In practice, to address these questions, the PA system needs to track various measures and use these measures to 

compute metrics. Therefore, we needed to devise PA system metrics. These metrics were defined based on their 

ability to address the types of questions posed earlier. These metrics can be divided into (a) persona-based metrics 

and (b) user-based metrics. The persona-based metrics include, for example, the following (with potential use 

mentioned after definition): 

• Time spent per persona: the duration users interacted with a given persona. Purpose: Proxy measure 

for users’ interest – it is likely that users spend more time with personas they find more interesting. 

• Number of visits per persona: the number of a given persona was visited by the users. Purpose: Proxy 

measure for users’ interest – it is likely that users visit personas they find interesting more often. 

• Persona bi- and trigrams: the number of times users visited specific two or three personas during a 

session or time period. Purpose: Bi- and trigrams can be indicative of comparative behavior, i.e., how 

users compare personas. (Technically, this is not a metric but a measure; however, we mention it here 

due to its nature of being computed.) 

In other words, these metrics communicate aggregate information about how one persona did relative to 

another – i.e., was one more popular than another, in what order where they visited, and so on. User-based 

metrics, in turn, communicate about a user or a group of users. These include: 

• Number of personas visited: the number of personas a user visited during a session. Purpose: to 

understand how thoroughly a user viewed the personas. For example, a user that only visits a small 

number of personas either quickly found what they were looking for, satisficed with the “first acceptable 

choice” [93], or was not engaged with the system and/or personas. 

• Persona coverage: the relative share of personas a user visited out of the personas available. Purpose: 

The same as previous, but as a ratio metric of the visited personas / the number of available personas. 

The higher the number of personas becomes, the more likely it is that the persona coverage per user 

decreases, as users would be unlikely to browse a very high number of personas for their professional 

tasks. 

• Average visit duration: the average time spent per persona for a given user. Purpose: can reveal if the 

user was more or less engaged relative to other users. 

• Persona rank correlation: the degree to which the order of a user visiting the personas corresponded 

with the personas’ order of presentation in the system (can be computed based on visit duration as 

well). Purpose: to test if there are order effects that affect persona use. 

Many of these metrics are inspired by similar metrics used in information theory [92], eye-tracking studies 

[25,42,71], and Web analytics [1,14,31]. While similar metrics are well established in said fields, these are metrics 

are not established in persona science and research. In fact, we are aware of no previous study that discusses 

metrics for persona user behavior – again, this hampers scientific progress in this domain. Coupled with 

participant data, the metrics can help analyze how users view different personas, if there is selection bias based 

on demographic factors of personas and their users, and so on. While these basic metrics provide a useful starting 

point for persona science, a lot more development in this domain is needed. It is also important to modify and 

adapt known metrics for the persona context, because they could be computed or interpreted differently when 

studying personas. We discuss this matter later in Section 6.8. 



3.3 Determining the Data Collection Modes 

The data collection modes were largely pre-determined by what is possible using the current Web technologies. 

Mouse-tracking is the obvious choice due to its commonality in online analytics and support provided by all Web 

browsers [39]. The advantages of mouse-tracking are three-fold: it (i) offers an unobtrusive form of tracking of 

natural user behavior, (ii) does not require calibration, and (iii) has perfect accuracy—i.e., there is virtually no 

measurement error, but the users’ movement of the mouse is perfectly traceable to specific pixels and UI 

elements. On the negative side, mouse-tracking is considered a weaker proxy for attention than gaze movement, 

i.e., eye-tracking [9,55], mainly because users might not always move their mouse when processing information 

on the screen. As processing of persona information requires eye-sight, eye-tracking is a useful data source to 

complement mouse-tracking in interactive systems [18]. The challenge of webcam-based tracking is that error 

margins can pose challenges for data quality, as there are differences in terms of hardware quality, lighting 

conditions, distance to screen and device, and a myriad of other conditions that can decrease online eye-tracking 

data quality [98]. While these issues do concern both separate hardware trackers (e.g., Tobii, MyGaze, GazePoint) 

and webcam-based eye-tracking, for the latter, the challenges in data quality are much higher because webcams 

do not provide access to infrared frequencies that the professional trackers use. 

Therefore, because mouse- and eye-tracking in a remote user study context each involve their unique 

advantages and challenges, it is appropriate to integrate both data collection modes into the PA system, which 

simultaneously completes the scope of the requirements. Figure 3 offers a conceptual overview of the PA system. 

Overall, the overall algorithmically generated persona process is as follows: end-users’ data => personas => 

persona users (e.g., marketers) => persona users’ data (collected via PA) => researchers and analysts (studying 

persona user behavior). In other words, personas serve the information needs of stakeholders, and PA serves the 

information needs of researchers interested in persona user behavior. 

 

Figure 3: Conceptual diagram of Persona Analytics. Multiple users can simultaneously interact with the 

persona system. The user’s interaction with persona profiles is captured via mouse- and eye-tracking, 
recorded in a central database (DB), and outputted via reporting interface. By analyzing the reports, 

researchers can make important discoveries for persona science. 



4 System Implementation 

4.1 Overview 

The defined questions acted as a guiding idea for requirements and implementation. In software development 

projects, requirements detail what is needed from a system [20]. Engineers or developers tend to implement 

features and functionalities according to the requirements to create the system. Two of the authors collaborated 

on creating the requirements, and one of the authors with the necessary skills implemented them. The system was 

tested internally and with a pilot user (reported in [37]), and we found it to log the data correctly. 

The implementation of PA was carried out for an interactive system called Automatic Persona Generation 

(APG) [2,3], which is a state-of-the-art system for algorithmically generated persona development. The system is 

available at https://persona.qcri.org. While we defer the reader to related work for a complete description of 

APG’s system functionalities and associated algorithms, the following subsection provides a brief explanation of 

the APG system. 

4.2 Algorithmic Approach for Persona Generation 

APG generates personas from online analytics data—e.g., from YouTube audience statistics or Google Analytics log 

data on end-users. APG infers demographically and behaviorally distinct patterns from user datasets [27,28]. APG 

has been previously applied to datasets on social media users [89], ad target groups [84], online news audiences 

[2,3], and video game players [90]. The APG persona creation relies on three main steps [30]: (a) identify unique 

user behavioral patterns using non-negative matrix factorization (NMF) [43], (b) associate these behavioral 

patterns with representative demographics (age, gender, country) to form “skeletal personas”, and (c) enrich the 

skeletal personas with personified information that matches the demographics (name, picture, job, education 

level, relationship status, topics of interest). Figure 4 summarizes the algorithmic persona generation process. 

 

Figure 4: (A) applying the NMF algorithm [43] to the user dataset V that consists of demographic groups 
(g) and content (c). This matrix is decomposed to W and H, both involving the hyperparameter p that 
indicates the number of personas. Epsilon describes the error term. Through enrichment process (B), 

explained in the body text, APG produces a set of p personas (C) that have personified information, 
conceptually known as “personification of big data” [96]. 

4.3 Measurement Paradigm 

The key insight going from APG (interactive persona system) to PA (persona analytics) is that, when personas are 

provided through a web browser, PA takes place via mouse- (and eye-)tracking that records the persona users’ 

mouse (or gaze) movements and clicks (eye fixations) on the persona profiles and their information elements. 

https://persona.qcri.org/


This enables empirical persona research, such as building click paths, persona visit sequences, dwell time 

analyses, and so on. In PA, we track both the information usage behavior within the persona profiles and the 

transitions between the personas. One can think of personas as “pages” in the conventional Web analytics terms, 

and then information reports are the pages’ content. To answer research questions that advance persona science, 

we need both levels of tracking. 

To this end, the PA system records the user’s mouse and gaze movements simultaneously during the session. 

The data is stored in a backend database. To support analysis, the screen coordinates are automatically converted 

to the corresponding information elements in the logs using JavaScript. In other words, the PA system records 

that a given mouse hovering or fixation was targeting, e.g., “Persona picture”. The duration of hover or fixation is 

calculated based on the “in” and “out” timestamps. In total, the PA system has 120 predefined HTML elements 

describing all information in the persona profile page (see Figure 5). The main elements include Headline (name, 

gender, age, country), About (picture, text description, job, education level, relationship status), Sentiment, Topics 

of Interest, Viewed Conversations (Quotes), Viewed Contents, and Audience Size, corresponding to typical 

information in persona templates [58]. 

 

Figure 5: Because interactive persona systems serve the personas via a web browser, web technologies, 
such as HTML and JavaScript, can be used for tracking how users interact with the personas. The frames in 
the figure illustrate how information elements in the persona profile are tagged for user tracking. In total, 

PA tracks 120 elements in the persona profile. 

4.4 Online Eye-Tracking 

The eye-tracking is implemented using WebGazer.js1 [62–64], a webcam-based gaze tracker developed at Brown 

University. WebGazer is available in JavaScript as an open-source library2. Multiple alternative frameworks were 

compared, but we chose WebGazer based on four reasons (a) it provides a relatively good accuracy based on our 

pilot testing, (b) is actively developed based on the update frequencies in the GitHub repository, (c) the source 

                                                                            
1 https://webgazer.cs.brown.edu/  
2 https://github.com/brownhci/WebGazer  

https://webgazer.cs.brown.edu/
https://github.com/brownhci/WebGazer


code is publicly available and can be integrated into systems such as APG, and (d) the software is provided free of 

charge. These properties make WebGazer a feasible online eye-tracker for research-based systems such as APG. 

4.5 Administrative Features 

In the APG’s UI, system administrators can enable either mouse tracking, eye tracking, or both for all users or for a 

subset of users. From a user’s point of view, the only difference is that, when eye-tracking is enabled, every 

session starts with calibration (see Figure 6a and b). The PA system processes mouse- and eye-tracking data 

identically, which means that both types of interaction are recorded in the database and can be exported in a 

single file in order to improve usability for the researcher using the PA system. When the coordinates of hovering 

or gazing correspond to a predefined element, this even, along with timestamp (in/out) and meta-data (User and 

Session ID), is sent by the client browser via Ajax (Asynchronous JavaScript and XML) to the backend database. 

The PA system maps the coordinates to the persona information element the user is interacting with. 

Administrators can download the log files for data analysis (see Figure 6d). They can also create new user studies 

from the persona system’s backend. 



  

(a) (b) 

 

 
(c) (d) 

Figure 6: (a) Eye-tracking calibration dialogue and (c) how it shows to users; (c) example of a user’s eye-
tracking pattern before the data is converted to each specific information element in the persona profile 

(denser color indicates more gaze fixations in a given area), and (d) data export dialogue shown for 
researchers to export user logs. The logs are provided in CSV files which can be downloaded via the 

Download button. 

To prepare the data exports, the system uses Pandas (i.e., Python Data Analysis Library) after retrieving the 

logs from the backend database. It computes the duration of each interaction based on “in” and “out” timestamps 

and generates a comprehensive data report, in which each mouse and eye fixation event, its timestamp, its target 

information element in the persona profile, and meta-information (Session ID and User ID), are saved into a file. 

The logs can be downloaded for further analysis. Information about the variables logged by the PA is included in 

Supplementary Material3. These variables were determined based on the metrics and questions detailed in the 

previous sections. As a result, the data recorded by the PA system enables the calculation of various metrics of 

persona user behavior. We now illustrate, though an example user study, how PA can serve persona researchers. 

                                                                            
3 https://www.dropbox.com/s/yeu1jrohs6lbpg8/central%20variables.docx?dl=0  

https://www.dropbox.com/s/yeu1jrohs6lbpg8/central%20variables.docx?dl=0


5 Validation Study 

A remote user study was conducted in which 114 participants used an interactive persona system with PA 

enabled, to browse a set of 10 personas, created from a dataset of a tourism-promoting organization with 1.8M 

(1,795,115) user likes over 5,312 Instagram posts. The participants’ task was to choose a persona to target for 

tourism marketing (i.e., promoting a specific destination, in this case, a country). Basic demographics of 

participants are provided in Table 1. 

Table 1: Participant demographics. 

Age  Male Female Non-binary  

M SD 62 55 1 N = 118 

(81.9%*) 

35.35 9.08 52.5% 46.6% 0.01%  

*For 18.1% of the participants, we did not have demographic information. 

The participants were recruited using an online data collection service called Prolific [61]; the same service has 

been used in various other persona user studies [82,83,88]. We used the platform’s industry categories as a 

sampling criterion, including “Art/Design”, “Graphic Design”, and “Market Research”, in order to reach people that 

work in industries were personas are relevant. Students were excluded. All participants were provided with a 

definition of personas, and a task description prior to their use of the system. The study flow is illustrated in 

Figure 7. The study dealt with testing the effect of simple and complex explanations on user behavior and 

perceptions, served via a product walk-through (i.e., a process that sequentially shows different parts of the 

system to a user that is logging in for the first time). The participants were directed from the online service to the 

system, which randomly allocated each participant into one of the three experimental conditions. The participants 

were randomly assigned by the APG system to one of the three experimental conditions (simple explanations 

about the system, complex explanations, and no explanations at all). 

 

Figure 7: An example of how the APG and PA systems can be synchronized to conduct a full remote user 
study. In this case, we used an online platform to recruit participants to the study [A]. After using the 

system to browse the personas [B] (no time limit was imposed), the participants completed a survey [C] 
and were redirected back to the data collection platform [D] that logged successful study completions. PA 

recorded both the survey and behavioral data [E]. 

After using the persona system, which consistent viewing as many personas as they wanted out of the ones 

created by the algorithm (see Table 2) for as long as they wanted, the users could click on a banner to indicate 

they are ready to complete their task, after which the participant is transferred to a survey platform that collects 



data about their task completion, perceptions, and demographic variables. Upon completing the survey, the 

participant is automatically redirected back to the online data collection platform, where their participation is 

marked complete. Both APG and the survey platform record the UTM parameters4 that identify (in an anonymous 

way) the participant so that participants that pass the data validation stage (i.e., researchers validating that their 

responses were genuine) can be easily compensated, and their system usage data is linked with their survey 

responses for further analysis. During system usage, PA was enabled, and data was collected on the users’ 

interactions with the persona system. 

Table 2: Personas the algorithm created for the user study. 

Persona Age Gender Country 

Mamdouh 28 Male Egypt 

Rahul 34 Male India 

Ashley 25 Female United States 

Muhammad 34 Male Pakistan 

Alaa 18 Female Egypt 

John 26 Male United States 

Abdalaziz 23 Male Egypt 

Rizky 18 Male Indonesia 

Putri 20 Female Indonesia 

Chris 40 Male United States 

6 Explorations of Persona User Behavior 

6.1 Overview 

In this section, we analyze the collected data to demonstrate how PA can serve empirical persona user research. 

We do not explicitly test any hypotheses, although the data obtained from PA could be used for that, but for the 

sake of demonstration, we inductively analyze the data and provide exploratory findings about persona user 

behavior. We then synthesize these findings in the form of propositions that future research could test, with 

complementary theorization, as hypotheses. In other words, we illustrate how PA can be of service towards 

persona science. 

The results of the effect of the three conditions will be reported in a future publication; here, we focus on 

demonstrating how PA can be used to analyze the data obtained from user experiments. For parsimony, the 

following analyses focus on the mouse-tracking data. Because the eye-tracking data is logged in the precisely same 

data structure as the mouse-tracking data, the exact same analyses and metrics can be obtained from eye-

tracking. Based on our piloting of the eye-tracking module, the accuracy strongly varies (from ~16% to ~80% in 

our testing) by the user, condition, and equipment. This is also why it is more reliable to carry out this 

demonstration with the mouse-tracking data. 

6.2 Descriptive Statistics 

Descriptive statistics about participants’ engagement with the system (see Table 3) indicate that, on average, 

participants spent around 8.5 minutes browsing the personas for their task and visited persona profiles on 

average 14 times. What is striking is the high dispersion among the participants – the standard deviations are 

high for both the dwell time (SD = 8.3 minutes) and visit counts (SD = 11). The participant with the shortest dwell 

time only used the system for 20 seconds, while the participant with the longest dwell time used the system for 

                                                                            
4 Urchin Tracking Module parameters are a standard technique for tracking source and meta-data of Web traffic. 



more than an hour (62.6 minutes). The shortest visit path only included visiting one persona, whereas the longest 

path consisted of visiting the persona profiles 60 times, which equals 6 visits per profile on average. These results 

indicate a major dispersion in engagement, with some participants being “persona power users” while others lack 

significant engagement with the system. Future analyses could investigate how these two extreme user types 

differ (e.g., demographic or industry variables that might explain the differences) and why (e.g., low task 

motivation, not perceiving personas as relevant or useful). 

Table 3: Dwell time (in seconds) and number of visits to the persona profiles by the participants. 

 Mean SD Min Max Median 

dwell time (i.e., system usage 

time) 

514.3 496.8 19.5 3756.8 410.8 

visits (i.e., number of times 

loading a persona profile) 

14.0 11.0 1 60 12 

As a whole, exactly half (50.0%) of the users viewed at least 9 of the personas (see Figure 8a), i.e., achieving a 

persona coverage of 90%, while less than a third (31.9%) viewed three or fewer personas. The fact that close to 

half (47.2%) viewed all 10 personas implies that users have a need for viewing a variety of personas—10 was the 

highest number in this study, but it seems likely that given the choice, users would have viewed more than 10 

personas. Concerning information viewing patterns (Figure 8b), the persona profiles contain 8 parent information 

elements (About, Audience Size, Headline, Sentiment, Timeline, Topics of Interest, Viewed Contents, and Viewed 

Conversations). Only a minority of persona visits contained viewing all 8 information elements (0.5%). While 

some visits included only viewing one information element (10.5%) – perhaps an indication of rapid verification 

of a recalled detail – more than half of the persona visits (52.3%) contained the viewing of at least four parent 

information elements (i.e., at least half of the main information in persona profiles). Section 6.5 investigates 

further what information was most viewed. 

  
(a) (b) 

Figure 8: (a) Persona coverage (i.e., the number of personas a participant viewed during their whole 
sessions), (b) Information coverage (i.e., how many parent information elements were viewed by the 

users during each viewing of a persona). 

6.3 Correlations and Gender Effects 

There was no notable correlation between participant age and their system usage time (r = 0.08) or age and 

number of visits (r = -0,16). In terms of dwell times, results from a t-test indicate that females were using the 

system longer (M = 634.8 seconds) than males (M = 493.4 seconds), t(114) = -1.55, p = 0.06. In terms of visit 

count, there was no significance difference, with females (M = 14.8) and males (M = 16.3) visiting a roughly even 

number of personas, t(114) = 0.79, p = 0.21. (Both these tests were based on the 118 participants for which we 



had gender information; there was one participant who indicated non-binary gender and who was therefore 

excluded from the analysis.) For males, with 95% confidence, the population mean for persona visit duration is 

between 33.4 and 59.4 seconds, based on 61 samples. For females, with 95% confidence, the population mean for 

persona visit duration is between 46.5 and 80.3, based on 55 samples. Finally, there is also no significant effect 

based on persona gender, with both male (M=37.8 seconds) and female (M=35.3 seconds) personas being 

frequented roughly an even amount of time, t(1996) = 0.50, p = 0.31. 

6.4 Persona Viewing Patterns 

One of the researchers conducted an exploratory data analysis (EDA) on 21 participants’ patterns of viewing the 

personas – by pattern, we mean how long a participant viewed a persona in their sequence of browsing the 

personas. This EDA revealed several different patterns of viewing the personas (see Figure 9), including (a) shark 

fin, (b) u-shape, (c) stabilizing, (d) sporadic, (e) linear declining, and (f) triangle shapes. 

   

(a) Shark fin (b) U-shape (c) Stabilizing 

   
(d) Sporadic (increasing) (e) Linear declining (f) Triangle 

Figure 9: Different persona viewing duration patterns based on an exploratory analysis. 

The variety of patterns indicates that it might be difficult to find general “laws” that would govern how 

individual users explore a set of personas. However, among the manually reviewed samples, we observed that the 

dwell times tend to decrease over the number of visits – 16 out of 21 had such a trend (76.2%) (e.g., a and c in 

Figure 9), while only three participants (14.3%) had an increasing dwell time trend (e.g., d). The two remaining 

had trends that could not be categorized as either decreasing or increasing (b and f). Thus, it appears that the time 

spent reviewing persona profiles decreases over the number of visits. In cases where the dwell time appears to 

“resurge” (e.g., b), the participant may be returning to a persona they found interesting earlier in order to verify, 

learn more, or compare information. Some patterns remain highly sporadic till the end of the session (e.g., d), 

while others seem to stabilize early (e.g., c). 

6.5 Persona Information Viewing Behavior 

We investigated where users focused their attention as a proxy for attention and interest. Results in Figure 10a 

show that the users were most interested in social media quotes in the persona profile (68.4% of the total dwell 

time), followed by the personas’ basic information (“About”, 14.1%) and audience size (8.1%) that indicates how 

many people there are on Facebook and Twitter similar to the persona. Plotting the data shows that users’ 

attention is unevenly distributed, with the quotes garnering over five times more dwell time than the second most 

popular information, i.e., the persona’s basic information. Unlike for persona visits, dwell time and visit count is 

strongly correlated (r = 0.63) for persona information. It is known from previous persona studies that quotes are 



very impactful for users’ perceptions of personas [86], but it is interesting that the mouse-tracking shows the 

quotes overshadow other information this strongly. 

 
 

(a) (b) 

Figure 10: Information viewing behavior of the participants. (a) Dwell time (bars) and visit counts (line). 
(b) Most common transitions between the parent information elements across all personas (Mamdouh is 

used for illustration). Users start browsing the persona’s basic information, including picture, text 
description and sociographics (State 0). They then move to comments (S1), most viewed content (S2), 
which is viewed repeatedly (S3), before moving to topics of interest (S4), audience size (S5), persona’s 

name and demographics (S6), and back to audience size (S7). 

On the other hand, if we instead of dwell time focus on the number of visits, personas’ basic information 

(“About,” as indicated by the orange line in Figure 10a) becomes the most important information element. This 

element contains the text description and picture of the persona that, again, previous research has been found 

influential for persona profiles [58]. The information viewing sequence (see Figure 10b) seems to move 

diagonally from top left to bottom right, then up, then bottom left, and back up and finally down (↘↑↙↑↓). This 

sequence was obtained by calculating the most common parent elements in states S0...S7 across all participant-

persona pairs. 

Two takeaways can be elicited from these findings: (a) that personas’ quotes, text description, and picture are 

among the most impactful information based on users’ mouse engagement, and (b) measuring dwell time and 

visit counts can give different results, which is why measuring these two separately makes sense – an information 

element with a high visit count but low dwell time is frequented in short bursts, whereas an information element 

with a high dwell time is focused on for a longer time; it is logical that quotes interest people because they are 

seen to reflect the persona’s attitudes and are information-rich for various user tasks. 

6.6 Effect of Order of Personas 

Order of presentation has been shown to affect how the information is accessed, used, and recalled [19]. Among 

the notable effects in this line of work are, e.g., the primacy effect that implies first seen information is the most 

impactful [68] and serial effect, implying that first and last items in a list are given special attention [54]. In 

persona system context, these effects can matter, because the personas are shown in a list, and such effects can 

therefore cause that, e.g., the needs of the first and last personas in the list would be considered more strongly 

than those of other personas. 

When plotting the data to investigate, three observations can be made from Figure 11: (1) there is a first 

persona effect, i.e., the first shown persona (Mamdouh) gets substantially more attention than the others, (2) 

there is no strong pattern of primacy effect in terms of declining dwell time based on persona’s order of being 

displayed in the system. However, (3) the fact that the last persona (Chris) is the second most viewed implies 



serial effect in which the first and last items of a list garner the most attention. Spearman rank correlation 

between persona order and dwell time is negligible (r = 0.176). Correlation between system order and number of 

visits is moderate (r = 0.576). Correlation between dwell time and visits is also moderate (r = 0.455). However, 

when we compute the most common personas visited (i.e., S1 = the first persona the user visits, S2 = the second 

persona they visit, ... S10 = the tenth persona they visit), we find that the TOP-10 path is precisely identical to the 

order of presenting the personas in the system. A further check reveals that 31 users (21.5%) follow this sequence 

when using the system. That is, about one fifth of the users browse all the personas in the order in which they 

were presented. 

 
 

(a) (b) 

Figure 11: (a) Dwell time distribution among personas. (b) First persona effect. Mamdouh (the first 
persona shown in the system) received almost one third of all dwell time from the users. If dwell time was 
distributed evenly, he should only receive 10%, which means an excess of 229% from this equal baseline, 

some of which likely stems him being the default persona in the system. 

Overall, these results indicate that (a) the system’s default persona garners the most attention, and (b) a 

sizeable portion of the users visit the personas in the exact order that the system shows them. Perhaps the system 

should evenly fluctuate the default position among the personas to mitigate for “discrimination” arising from this 

effect. The current logic is that the system loads the persona by default that has the highest audience 

representation, i.e., the most engagements in the baseline user data the personas are created from. Testing 

various rationale for the default persona and the effect of rotation are excellent ideas for future work. We also 

observed a potential cultural effect, which is a possible manifestation of the users better identifying with the 

personas from their own cultural sphere (see Table 4). Cultural aspects in personas remain an important area of 

future work, such as the special role of default and exit personas, i.e., those with whom users finish their browsing 

session (see Figure 12). 

Table 4: Ethnic bias? The only three Western personas rank the highest in terms of average view time 
(apart from Mamdouh that is the default persona). 89.8% of the participants were from Western 

countries (Europe and United States). Users may feel more comfortable identifying with personas from 
their own culture and ethnicity. This also implies persona studies should employ culturally and ethnically 

diverse samples to obtain internationally valid results. 

Persona Avg time per visit  

Mamdouh (rank = 1) 69.10 

Chris (rank = 10) 37.78 

John (rank = 6) 33.78 

Ashley (rank = 3) 33.49 

Alaa (rank = 5) 32.98 



Persona Avg time per visit  

Rizky (rank = 8) 28.92 

 

Putri (rank = 9) 28.13 

Abdalaziz (rank = 7) 26.53 

Muhammad (rank = 4) 24.85 

Rahul (rank = 2) 20.79 

 

Figure 12: Exit personas. The number of times a user stopped their session after viewing a given persona. 
Because the task dealt with considering a specific persona, it is possible that the exit indicates a higher 

likelihood of the persona being chosen for the task. 

6.7 Modeling Persona User Behavior 

Because the PA system records the user’s transition from one persona information element to another (based on 

mouse and gaze movements), as well as capturing transitions from one persona to another, there are important 

opportunities for modeling user behavior, some of which we illustrate here. Figure 13 describes these 

opportunities through the concept of persona-gram, which refers to a string of letters depicting a user’s path of 

visiting either the personas or the information elements within a persona profile. This information can be stored 

as a state transition matrix (see bottom of Figure 13) which can be further used for computing the probability of a 

user transition from one state to another. 

In Figure 13, the names on the left illustrate a user’s path of visiting the personas. On the right-hand side (User 

1 = U1), the same path is transformed into a string. The string format enables the comparison of different users 

using Levenshtein’s edit distance (ED) [73]. For example, User 2 (U2) differs from User 1 (U1) in only two string 

states (bolded in Figure 13), yielding ED1,2 = 1. Users that have a low edit distance are similar to each other in 

terms of their persona use behavior, whereas users with a high edit distance are behaviorally more different. 

When computing the distances of all users, it becomes possible to identify “average” behaviors and distinct outlier 

behaviors. (Moreover, it is possible to consider the duration of each visit to get a higher dimensional 

representation of the user’s viewing behavior.) 



 

Figure 13: Illustration of persona-grams using imaginary data. The series on the left describes a user’s 
transition from one persona to another. User 1 on the right is the same sequence transformed into a 
string. User 2 has the same sequence except for two differences (A-N, bolded), which means the edit 

distance is 2, i.e., one needs to make two edits to make the strings identical. The fewer changes one needs 
to make, the more similar the sequence of viewing a person is between two users. 

We computed the average edit distance across the dataset obtained and found the number highly dispersed. In 

other words, two users would rarely view the personas (or the information elements within the personas) in the 

same or similar order. This finding is interesting in itself – it implies users’ processing of persona information is 

more idiosyncratic than anticipated. To give a simple example, Table 5 shows three randomly chosen participants 

that each have a path length of 10, i.e., they visited persona profiles 10 times during their session. 

Table 5: Persona-grams of three randomly chosen users who each visited persona profiles 10 times. The 
color codes indicate the same persona being visited in the same sequence: yellow is shared by all three 

users, green is shared by Users 1 and 3, and turquoise by User 2 and 3. User 3 has a more similar browsing 
behavior with User 1 than with User 2, and User 1 and User 2 share the least similarity. 

User 1 User 2 User 3 

Mamdouh 

Alaa 

Putri 

Alaa 

Rizky 

Abdalaziz 

Alaa 

Putri 

Rizky 

Alaa 

Mamdouh 

Rizky 

Chris 

Rahul 

Alaa 

Muhammad 

Abdalaziz 

Ashley 

John 

Putri 

Mamdouh 

Alaa 

Putri 

Alaa 

Ashley 

Chris 

Abdalaziz 

John 

Mamdouh 

Ashley 

As can be seen from Table 5, the behaviors are almost completely unique—for example, the only shared visit 

among the three is to the first persona, which is the default shown by the system. Hence, edit distance is a 

troublesome metric, because in this case, we would need 9 edits to make User 1 and User 2 — ratio-wise, this is 

9/10, so 90% change rate (i.e., 9 out of 10 paths are different). Due to high uniqueness demonstrated by this 

example (which is also accentuated by the fact that the strings are of different lengths across the dataset!), the 

similarity of behaviors could perhaps be measured using other options. For example, User 2 and User 3 viewed 7 

same personas (Mamdouh, Chris, Alaa, Abdalaziz, Ashley, John, and Putri) and 3 different personas (Rizky, Rahul, 

Muhammad). So, even though their exact viewing sequences are very different, the users actually view more the 

same than the different personas, i.e., there is likeness in their browsing behavior. To quantify this likeness, we 

can apply set theory to form an intersection (i.e., an overlap of paths). The intuition is that if two users visited 



more same personas than two other users, their persona browsing behavior was more similar. To can quantify 

this by calculating the Jaccard coefficient (J), which simply indicates the overlap between two sets. This metric is 

commonly used in information theory to compare sets [44]. 

Applying J to our examples from Table 5, we can observe that User 2 and User 3 are more similar to each other 

(J=0.7) than User 1 and User 2 (J=0.5) or User 1 and User 3 (J=0.5). (For replication, the sets are: User 1 – M, A, P, 

A, R, B, A, P, R, A; User 2 – M, R, C, H, A, U, B, S, J, P; and User 3 – M, A, P, A, S, C, B, J, M, S.) Unlike ED, which is only 

applicable to pairwise comparison, sets can be expanded from pairwise comparisons to multiple sets (see Figure 

14). 

 

Figure 14: Among the 10 available personas, 4 (40%) were viewed by all three users. Users 2 and 3 
viewed 3 personas that User 1 did not view, whereas Users 1 and 2 viewed 1 persona that User 3 did not 
view. User 2 viewed 2 personas that neither of the other users viewed, which indicates this user had the 

most diverse viewing pattern. 

Obtaining the number of intersecting elements (i.e., shared personas that any number of participants viewed) 

is trivial and easy using basic functions in scripting languages like R and Python, which increases the practical 

appeal of modeling persona behavior using sets. One can also use sets to compare the behaviors of different 

groups. We illustrate some of these cases in Figure 15. For example, (a) union can be used for identifying all 

personas that a group of users viewed, which can be beneficial when the number of personas exceeds a handful, as 

would be the case for large and heterogeneous online audiences. (b) Intersection shows common elements of 

two or more users or groups. Intersection can reveal common personas of interest, i.e., that most participants 

engaged with. (c) Difference can show personas that one group viewed exclusively, e.g., those that were unique to 

more experienced persona users. Finally, (d) subset and superset can help make comparisons on the variety of 

personas visited. For example, in our previous case, User 2 is a superset of User 1 (i.e., User 2 visited all the 

personas that User 1 did and more). 



    
(a) Union (b) Intersection (c) Difference (d) Sub- and supersets 

Figure 15: Examples of basic set operations and how they can be used for investigating persona user 
behavior. 

While approaching the analysis of persona user behavior using set theory seems fruitful, we can include even 

more information in the comparison. Namely, the set representation ignores that the visits are typically unevenly 

distributed among the personas, both by count and duration. Some personas are viewed more than once; some 

are viewed considerably longer than others. A set would not consider this information at all. 

For a representation that considers such information, we can turn to empirical distributions or probability 

distributions. These indicate how the time or number of visits is allocated between different personas during a 

user session. For example, if a user visits Alaa 5 times, Putri 5 times, and John 2 times, the empirical distribution is 

5 / 12; 5 / 12; 2 / 12; or [0.42, 0.42, 0.16]. For a finite set of personas and users, we can compute a complete 

probability distribution for each user, where complete implies that each persona-user pair will have a value. Then, 

from information theory, we can use several metrics to compare the obtained probability distributions. These are 

known as statistical distance metrics (e.g., Kullback–Leibler (KL) divergence or Jensen-Shannon distance). The 

smaller the distance between two users, the closer their behavior is in terms of how they divide their time among 

the available personas. Applying this logic to our dataset, we obtain the probability distributions indicated in 

Table 6. 

Table 6: Distribution of example users’ visits among the shown personas. Non-zero values are highlighted. 

 Mamdo

uh 

Rah

ul 

Ashl

ey 

Muham

mad 

Ala

a 

Joh

n 

Abdala

ziz 

Riz

ky 

Put

ri 

Chr

is 

Us

er 1 

10% 0% 0% 0% 40

% 

0

% 

10% 20

% 

20

% 

0% 

Us

er 2 

10% 10

% 

10% 10% 10

% 

10

% 

10% 10

% 

10

% 

10

% 

Us

er 3 

20% 0% 20% 0% 20

% 

10

% 

10% 0% 10

% 

10

% 

The distance D between two distributions can be computed using the following equation, 

,

1

1
  |  ( ) ( ) |

N

p q i i

i

D p x q x
N =

= − ,  

where p and q are the distributions to compare (e.g., User 1 and User 2). The formula calculates the absolute 

difference for both users for each persona X, and then takes the average as the distance number. Unlike KL 

divergence, which is non-symmetrical (i.e., the distance between User 1 and User 2 might not be the same), this 

formulate gives symmetrical results (i.e., Dpq = Dq,p). (As a sidenote, symmetry is, of course, desirable for our 

purpose, because there is no reason to assume that the results should differ when comparing User 1 to User 2 or 

vice versa; in both cases, the sequences are the same.) When inputting the fractions from Table 6 to this formula, 

we can see the results aligning with the J comparison, so that Users 2 and 3 are the most similar (D = 0.06), 

whereas User 1 is equally distant to User 2 (D = 0.10) and User 3 (D = 0.10). This example illustrates how concepts 

from information science can be leveraged for persona science, namely, by understanding persona viewing 



behaviors as probability distributions and then computing distance. Essentially, the smaller these distance values 

are, the closer two persona viewing behaviors would be (as the behaviors are represented as probability 

distributions). 

6.8 Interpreting the Metrics 

It is well known that general rules about whether a metric value is “good” or “bad” are difficult to draw – for 

example, for an entertainment website or social media service, it is desirable that users spend a lot of time on the 

site, because time is positively correlated with revenue models [100]. However, the opposite applies for 

government information websites or search engines – the user is expected to find the information as quickly as 

possible and then leave the site. So, in some cases, small engagement time is optimal; in other cases, it is not 

optimal. For personas, the same applies, with even more nuance – in many professional tasks, users are time-

pressed; they have deadlines, they want the information immediately, and so on. For these scenarios, low 

engagement time with a persona system would be considered a good sign (given that the user was able to 

complete their task successfully or that the personas helped the user). However, in scenarios where the user is 

conducting end-user research (e.g., market research), they might be interested in dwelling deep into the persona 

details: a low engagement time would therefore not indicate that the personas were useful for the user. 

We provide further examples, as the matter of defining and computing various metrics is not trivial. 

Example 1: Consider that a user visits 5 personas during the session and does overall 20 visits. Now, if we 

apply the average, we get 20 / 5 = 4. however, the comparative aspect is if user visiting these 5 personas during 

their first 5 visits, versus visiting the 5 personas during their first 10 visits, the behavioral pattern is different. In 

the former case, the user is first visiting many personas, and then spending the rest of the time comparing them. 

In the latter case, the user is engaging in comparative behavior already during visiting the first personas. These 

two strategies would be different, but a simple average would miss the nuance. To quantify such patterns, we can 

compute x = how many personas the user visited and y = how many visits it took the user to visit each of the x 

personas at least once. This metric can be called “persona scanning tendency” or PST. Using the above example, 

the scores would be different: 

PST5 = 5 / 5 = 1 versus PST10 = 5 / 10 = 0.5. 

A lower PST score would indicate a smaller tendency to scan all or many personas first and then dwell into 

their information, and vice versa. A high PST score could also be associated with a linear user behavior, in which 

the user sequentially visits the available personas. In contrast, a low PST score could be associated with non-

linear user behavior. For example, if it took the user ten visits to see all the five personas, they were likely doing 

comparisons along the way. Therefore, this PST metric can reveal insights about different users’ tendency to 

compare personas against one another and proceed in an organized manner when browsing the available 

personas. 

Example 2: How to measure how diversely a user browses the personas? Highly diverse behavior would be 

one that looks at many personas. For example, if there were a total of 10 personas available, 5 out of 10 personas 

visited is less diverse than 10 out of 10 personas visited. That is, persona coverage would seem like a good metric 

(being 5 / 10 = 0.5 and 10 / 10 = 1 for the above cases). But, if the user only made 5 visits and visited 5 personas 

(out of the 10), then that is more diverse than making 12 visits and visiting 6 personas. In other words, when 

assessing the diversity of visit behavior, understood in this manner, the denominator should be the number of 

visits the particular user made, not the number of available personas. So, even if persona coverage would be 

higher for the user visiting 6 personas (6 / 10 = 0.6), their visit diversity would still be lower than that of the user 

visiting one persona less (i.e., 6 / 12 = 0.5 vs. 5 / 5 = 1). 

There are also cases where different granularity would be needed, for example, considering all information 

elements (of which there are 120) versus only parent information elements (of which there are 8), or considering 

forward-only movement versus both forward and backward movements (i.e., across information elements and/or 

personas). Depending on these choices, the dimensionality of the analyzed dataset can greatly increase or 

decrease. Overall, these examples highlight the non-trivial nature of measuring and understanding persona user 



behavior. Intepreting PA metrics follows the general patterns in UX research: a high value can indicate either a 

desirable or undesirable effect depending on the task type and user goals: for example, longer dwelling time may 

be a sign of more interest, but it could just as well be a sign of confusion and disorientation. Aligning the metrics 

with expressed user perceptions (e.g., “I was confused when using the personas”) can be useful in this regard. 

Overall, case-dependent interpretation is required. 

7 Discussion 

7.1 Highlights and Novelty 

Even though there is some empirical evidence showing the effectiveness of personas for specific tasks [78], as a 

whole, the body of knowledge remains underdeveloped, even after 20 years of personas being part of HCI 

research and practice. The purpose of PA is to help generate new knowledge on persona user behavior towards 

the advancement of persona science. Examples include behaviors related to order of presenting the personas, 

revisit frequency, users’ styles of browsing and comparing personas, and persona choice – i.e., how and why 

people choose a specific persona for their decision-making task. Analyzing users’ typical viewing patterns and 

dwell time per persona information can inform persona information design (i.e., what information users most 

interact with), and help deduce understanding of persona usage based on a real system and real or realistic use 

cases and scenarios. Combined with algorithmically generated persona systems, this opens the possibility of 

updating persona profiles in real-time based on expressed user needs. 

It is well known that the body of knowledge on personas relies on a high number of case studies as opposed to 

repeated experiments with independent samples (i.e., different organizations, locations, users, etc.). Even when a 

case study focuses on a specific context, there is a need for multiple case studies to establish something more 

generalizable than the one case can do [99]. Without having more rigor, persona research cannot proceed much, 

and it risks going in circles instead of establishing evidence-based empirical phenomena. Examples of this “going 

on circles” include conflicting findings about personas being applicable and not being applicable (e.g., [70] vs. 

[60]) – to date, nobody has explained when personas are applicable and when not. The only way to achieve 

plausible explanations is to carry out repeated testing and measure the results empirically, for which PA provides 

a way. 

Users’ interaction with personas can be measured in many ways. Here, we focused on two commonly used 

technologies: mouse tracking and eye tracking, as these two technologies have unique strengths and weaknesses, 

and both can be implemented in a Web-based interactive persona system. As acknowledged in many HCI studies, 

eye- and mouse-tracking are helpful techniques for studying user engagement with interactive systems [15,91], 

and personas in particular [26,76,78,85,86]. When integrated directly into an interactive persona system, these 

data collection methods can deliver rich datasets describing persona use and modeling for complex user 

behaviors [18,22]. 

Moreover, integrating these techniques also enable user studies during exceptionally difficult times (e.g., 

during a global pandemic) when it is either not possible or very difficult to conduct in-person user studies. 

Highlights of this work include the following: 

• Providing conceptual underpinnings of persona analytics and persona science, two emerging and 

promising concepts for HCI and user segmentation. 

• Developing a novel persona analytics system embedded within a persona system, by integrating mouse- 

and eye-tracking functionalities. Measuring users’ mouse and gaze activities grants an understanding of the 

behaviors of a persona user. Studying the behavior of the persona user is essential to generate empirical 

knowledge and theories of personas and human-persona interaction. 

• Demonstrating an end-to-end experimental loop for empirical persona studies, using an interactive 

persona system and data collection platforms with a user study of 144 participants. Remote user studies such as 

this help scale persona user studies from the conventional 30-or-so to hundreds of participants. 



• Providing exploratory findings of persona user behavior, based on the use of PA, its metrics, and various 

statistical techniques. 

The trends contributing to interactive persona systems are likely to continue, including the evolution of (1) 

digital user data from online analytics platforms through APIs, (2) data science algorithms and libraries that 

integrate into quantitative persona creation process, and (3) web medium that surpasses the limitations of paper 

for persona delivery and user engagement. Interactive persona systems commonly rely on web standards [33] 

that enable users to access personas from any device with an internet connection and make it possible to record 

persona users’ behavior and analyze with using PA, a customized solution for tracking users’ interaction with 

persona profiles. This system, therefore, has value and potential for advancing persona science and the 

development of interactive persona systems for years to come. 

The current work puts forward a new artifact for studying the behavior of persona users. While previous 

persona user studies [73,76,78] have analyzed dwell time and users’ information viewing sequence, these metrics 

have, as far as we know, not been captured directly from an interactive persona system in any previous research. 

Thus, the presented solution has novelty in its field. As far as we know, the study presented the first “online 

laboratory” solution for interactive persona systems during a period in history where there is a need for remote 

user study solutions. 

7.2 Novel Opportunities to Push Forward Persona Science 

Persona science needs progress on all fronts, eyeing on long-term theory formulation but also investing in short-

term returns through the use of empirical methods. Persona science can contribute to a much-needed transition 

beyond the general claims that “personas work” or “personas do not work”, or the repetition of their “benefits” 

and “problems,” into systematically examining the conditions where the effects emerge. 

Many human factors have not been investigated empirically. Based on case studies, variables of special interest 

include at least (a) Experience [81], (b) Task type [4], (c) Job role [59], and (d) Culture [56]. Investigation of 

combination of these human factors would be based on specific research objectives. 

First, the effect of users’ experience with personas on behaviors; this tends to be reported in persona studies 

but not included as a variable. How does novice persona users’ use of personas differ from more experienced 

users? Can the behaviors of more experienced users be used for guiding the novice users to learn to use personas 

more efficiently? The task and task type, more specifically. This is reported but rarely controlled – most typically, 

only one task type is deployed and in only one empirical setting, without repetition to achieve robustness. As a 

result, the current body of literature has, for example, no comparison between different task types – e.g., design, 

content creation, ad targeting, etc. Personas can be deployed for a range of professional tasks, but no study 

compares what kind of personas are ideal for the various task types and if users approach the personas differently 

based on the task type. 

The users’ job roles – again, reported, but comparison among different roles and organizational units are 

rarely conducted, even though it is common sense that a person’s job position would greatly affect how they use 

personas to support their work. Studies tend to mention “designers” but looking deeper into these users’ job 

positions, it is revealed that they work in multiple departments, have multiple different perspectives to the end-

user, and require much different information for their decision making. Overall, systematic analysis of these 

variables in experimental studies can produce long-lasting, consistent, and robust knowledge on personas and 

their users, extending the boundaries of persona research. Finally, culture. Previous research has hinted at 

cultural effects related to personas use [83], but there is no adequate understanding of how the cultural match 

between the shown personas and the users mediate the interaction and whether personas themselves can help 

bridge cultural gaps for design. 
Table 7 proposes a preliminary research roadmap. Naturally, due to the enormous scope of potential research 

topics, this proposal is not a complete one. However, it maps some relevant questions that can be addressed via 

PA. 



Table 7: Research roadmap for empirical persona research. This roadmap includes research questions 
that PA can help address. Through solving multiple research questions, researchers can start formulating 
a unified theory of personas that would explain, with robust empirical when and when not personas work, 

what factors govern their successful use, and how can persona creators and champions increase the 
likelihood persona projects. 

 Open Research Question (ORQ) and Sub-

Questions (SQ) 

Useful for…* Variation 

by... 

ORQ1 What types of personas are most/least viewed? 

• SQ1a: Are there differences based on 

personas’ gender, age, nationality, 

ethnic background? 

• SQ1b: Are some types of personas 

systematically disadvantaged in terms 

of how frequently and how long users 

interact with them? If so, can system 

features, e.g., rearranging order of 

showing, mitigate such disadvantages? 

Persona Creation and Use ⇒ How do users 

interact with persona profiles? 

experience 

task type 

job role 

culture 

ORQ2 What persona information was most/least viewed? 

• SQ2a: How is users’ consumption of 

persona information affected by 

position and screen size of the 

information? 

• SQ2b: Is some information considered 

redundant regardless of its position 

and screen size? 

Persona Information Design ⇒ What 

information do persona users pay attention 

to? 

ORQ3 How does the user transition between (a) the 

personas and (b) the information elements in the 

persona profiles? 

• SQ3a: What is the degree of linearity / 

predictability? 

• SQ3b: Are the dwell times 

consistently increasing / decreasing? 

Persona Information Design ⇒ How can we 

model users’ cognitive styles of using 

personas? 

ORQ4 How does increasing/decreasing the number of 

personas affect user behavior? 

• SQ4a: What interaction techniques can 

help users cope with more than a 

handful of personas? 

• SQ4b: What is the extra cognitive cost 

of adding a persona? 

Persona User Behavior ⇒ What is the optimal 

number of personas? 

ORQ5 How do the viewed (a) personas and (b) persona 

information influence users’ design choices? 

• SQ5a: What is the effect of persona 

use of desirable outcomes such as 

increased usability, satisfaction, 

profitability? 

• SQ25b: How can these outcomes be 

measured? 

Persona Impact, Value of Personas ⇒ What is 

the real value of personas? 



 Open Research Question (ORQ) and Sub-

Questions (SQ) 

Useful for…* Variation 

by... 

*NOTES: The categories of Persona Creation, Validation, Use, and Impact originate from [77]. ⇒ indicates a 

higher abstraction of the ORQ in question. 

7.3 Practical Implications 

Benefits for Researchers. Persona science is the application of scientific methods to persona research, with the 

goal of producing empirically valid knowledge about persona creation, validation, use, and impact for design 

outcomes, individual decision making, and organizations. PA supports persona science by aiding in the scientific 

process, such as generation and testing of hypotheses: data exploration → inductive analysis → propositions → 

hypotheses → independent data collection → hypothesis testing → theory generation of personas and their users. PA 

can be deployed for both inductive and deductive research. Deductive research typically formulates hypotheses 

and relies on experiments to test those hypotheses. The hypotheses can be inspired by works in HCI, social 

psychology, economics, etc. Manipulating variables in the interactive persona system, such studies can be 

conducted in-person or remotely (see Figure 7 illustrating an end-to-end loop for remote participation). Inductive 

studies rely on freely exploring user behaviors to formulate propositions that can later be tested in controlled 

settings. Here, we conduct a series of exploratory analyses in order to demonstrate the type of information, 

metrics, and analyses that the PA can afford to researchers. 

Benefits Relative to Pre-Existing Solutions. From a practical point of view, the reader might pose the 

question, “Why not just use Google Analytics for this purpose? Why develop a new system?”. There are various 

reasons for that. Compared to pre-existing industry solutions, such as Google Analytics (GA), which is the 

dominant Web analytics service [65], PA has four major benefits, summarized in Table 8 and explained thereafter: 

Table 8: PA vs. GA – a practical comparison. 

 Persona Analytics Google Analytics 

Data ownership x - 

Tracking 

customization 

x Partial 

Clickstream logs x Requires Premium version 

Persona metric 

exports 

x - 

• Data ownership: the data is recorded to our serves, not to third party servers such as Google’s. 

• Customized event and information tagging: As we know the exact functionality of the APG system, we can 

label the information elements and events with suitable names from the onset. 

• Clickstream data: Unlike in the standard GA installation, we are able to obtain raw log data of all actions 

taken by a user in the APG system. The standard GA installation only provides aggregate data export. 

• Tailored reports for persona user analysis: GA reports are designed for websites, not for custom-built systems 

like interactive persona systems. Therefore, the available metrics and the reports are not suitable for analytical 

questions related to personas. The development of our reports was inspired by analytical tasks for which the 

data would be used. Thus, the reports serve persona research better than reports in GA. 

With these benefits, the PA system is more equipped for tracking persona user behavior than GA, and therefore 

has the potential to serve researchers more adequately. 

Benefits for end-users. Personas are an important tool for professionals in various domains, which is why 

understanding how people use them is instrumental for the creation of better end-user insights tools and 

therefore advocate more customer- and user-centric thinking in organizations. The practical implementation of 

PA also requires consideration for the users’ privacy, to let them know their usage of the system is being tracked. 



Therefore, we notify users of this tracking in APG’s terms of service, which is similar to any other website tracking 

user behavior. Additional ethical considerations include acquiring consent from users when conducting user 

studies to track their system usage behavior, as well as acceptance from an institutional review board (IRB) when 

there is a reason to suspect that the research deals with topics that warrant ethical scrutiny (in persona user 

studies, harmful scenarios tend to be rare but nonetheless could exist for some study topics). 

7.4 Specific Application Areas for Future Research and Development 

There are several areas to further deepen the level of analysis. The data produced by the PA system can be 

examined using many computationally advanced approaches, for example, by creating a persona state-transition 

matrix and applying Markov Chain modeling or neural networks to the constructed matrix to model historical 

dependencies (similar to [94]). Similarly, future work could look into predicting user outputs (e.g., task success) 

using behavioral features. There are several architectures to deal with sequential data, including recurrent neural 

networks (RNNs) [10], that could be used for modeling purposes that can result in persona recommenders or user 

behavior classification. While we leave these additional considerations for future work, we do want to point out 

two prominent opportunities: 

• Prediction. By varying the personas, their information, or interactive features, the effect of such 

variables on outcome variables such as persona choice, task completion success, quality, time, perceived 

usefulness, or design task impact (e.g., usability improvement) can be measured. For example, when 

there are many personas available, users often choose a specific persona for their task. This choice 

matters because selecting one persona over another can influence how different user groups’ needs are 

considered (or not considered). 

• Persona recommenders. For large and heterogeneous populations, the appropriate number of 

personas can be in the hundreds or more to correctly represent the diversity of the user population [11]. 

Therefore, there is a need, on the one hand, to show more personas to users, and on the other hand, to 

build tools and methods for users or systems to narrow down the number of candidate personas for a 

given task, without taking away users’ choice of reaching beyond this candidate pool to browse 

marginalized or fringe user groups [21]. Approaches for achieving this may include recommending users 

a specific set of personas based on task criteria or user modeling. Simple examples that have been 

already implemented in interactive persona systems include sorting the personas based on their 

segment size (i.e., how many people they represent in the baseline data), either in decreasing order 

(when wanting to see the most representative personas) or in increasing order (when wanting to see 

outlier personas). However, more elaborate persona recommenders are missing to date. The 

recommenders’ value increases as the number of personas increases because large persona sets with 

hundreds or even thousands of personas [27] pose a manageability problem for a human efficiently to 

deal with. 

Finally, the PA approach could be applied to other forms of profile systems, including social media profiles, 

gaming avatar profiles, and so on. Although the system was designed with personas in mind, if the use case of 

wanting to learn how people interact with profiles is adequately similar, then PA-type of analytics can be 

deployed. 

8 Conclusion 

In this work, we demonstrated a fully functional persona analytics system embedded within a fully functional 

interactive persona system. We demonstrated ways in which persona analytics can reveal structural patterns in 

user interaction with personas. These patterns can lead to advancement of persona science and theoretical 

propositions for persona-user interaction. Persona research lacks empirical studies, so there is plenty of room for 

contributions by keen researchers. To this end, we hope our research encourages others to pursue empirical 

research questions in the persona domain. The techniques we demonstrated have special value during 



exceptional times when physical user studies are hindered by social distancing. Finally, we proposed seven 

metrics that can be computed from the persona analytics data. In addition to these metrics, devising new 

quantitative metrics for persona studies is a valuable research direction. 

References 
<bib id="bib1"><number>[1]</number> Mohamed Aboelmaged and Samar Mouakket. 2020. Influencing models and determinants in big data 
analytics research: A bibliometric analysis. Information Processing & Management 57, 4 (July 2020), 102234. 
DOI:https://doi.org/10.1016/j.ipm.2020.102234</bib> 
<bib id="bib2"><number>[2]</number> Jisun An, Haewoon Kwak, Soon-gyo Jung, Joni Salminen, and Bernard J. Jansen. 2018. Customer 
segmentation using online platforms: isolating behavioral and demographic segments for persona creation via aggregated user data. Social 
Network Analysis and Mining 8, 1 (2018), 54. DOI:https://doi.org/10.1007/s13278-018-0531-0</bib> 
<bib id="bib3"><number>[3]</number> Jisun An, Haewoon Kwak, Joni Salminen, Soon-gyo Jung, and Bernard J. Jansen. 2018. Imaginary People 
Representing Real Numbers: Generating Personas from Online Social Media Data. ACM Transactions on the Web (TWEB) 12, 4 (2018), 27. 
DOI:https://doi.org/10.1145/3265986</bib> 
<bib id="bib4"><number>[4]</number> F. Anvari, D. Richards, M. Hitchens, and M. A. Babar. 2015. Effectiveness of Persona with Personality 
Traits on Conceptual Design. In 2015 IEEE/ACM 37th IEEE International Conference on Software Engineering, 263–272. 
DOI:https://doi.org/10.1109/ICSE.2015.155</bib> 
<bib id="bib5"><number>[5]</number> Farshid Anvari, Deborah Richards, Michael Hitchens, Muhammad Ali Babar, Hien Minh Thi Tran, and 
Peter Busch. 2017. An empirical investigation of the influence of persona with personality traits on conceptual design. Journal of Systems and 
Software 134, (December 2017), 324–339. DOI:https://doi.org/10.1016/j.jss.2017.09.020</bib> 
<bib id="bib6"><number>[6]</number> Farshid Anvari, Deborah Richards, Michael Hitchens, and Hien Minh Thi Tran. 2019. Teaching User 
Centered Conceptual Design Using Cross-Cultural Personas and Peer Reviews for a Large Cohort of Students. In 2019 IEEE/ACM 41st International 
Conference on Software Engineering: Software Engineering Education and Training (ICSE-SEET), 62–73. DOI:https://doi.org/10.1109/ICSE-
SEET.2019.00015</bib> 
<bib id="bib7"><number>[7]</number> M. Aoyama. 2005. Persona-and-scenario based requirements engineering for software embedded in 
digital consumer products. In Proceedings of the 13th IEEE International Conference on Requirements Engineering (RE’05) , Washington, DC, USA, 
85–94. DOI:https://doi.org/10.1109/RE.2005.50</bib> 
<bib id="bib8"><number>[8]</number> M. Aoyama. 2007. Persona-Scenario-Goal Methodology for User-Centered Requirements Engineering. In 
Proceedings of the 15th IEEE International Requirements Engineering Conference (RE 2007) , Delhi, India, 185–194. 
DOI:https://doi.org/10.1109/RE.2007.50</bib> 
<bib id="bib9"><number>[9]</number> Ernesto Arroyo, Ted Selker, and Willy Wei. 2006. Usability tool for analysis of web designs using mouse 
tracks. In CHI’06 extended abstracts on Human factors in computing systems, 484–489.</bib> 
<bib id="bib10"><number>[10]</number> Homanga Bharadhwaj. 2019. Explainable recommender system that maximizes exploration. In 
Proceedings of the 24th International Conference on Intelligent User Interfaces: Companion , 1–2.</bib> 
<bib id="bib11"><number>[11]</number> Chris Chapman, Edwin Love, Russell P. Milham, Paul ElRif, and James L. Alford. 2008. Quantitative 
Evaluation of Personas as Information. In Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 1107–1111. 
DOI:https://doi.org/10.1177/154193120805201602</bib> 
<bib id="bib12"><number>[12]</number> Chris Chapman and Russell P. Milham. 2006. The Personas’ New Clothes: Methodological and Practical 
Arguments against a Popular Method. In Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 634–636. 
DOI:https://doi.org/10.1177/154193120605000503</bib> 
<bib id="bib13"><number>[13]</number> Eric Chu, Prashanth Vijayaraghavan, and Deb Roy. 2018. Learning Personas from Dialogue with 
Attentive Memory Networks. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Association for 
Computational Linguistics, Brussels, Belgium, 2638–2646. Retrieved June 20, 2019 from https://www.aclweb.org/anthology/D18-1284</bib> 
<bib id="bib14"><number>[14]</number> Theresa Bilitski Clarke and Bernard J. Jansen. 2017. Conversion potential: a metric for evaluating 
search engine advertising performance. Journal of Research in Interactive Marketing 11, 2 (2017), 142–159.</bib> 
<bib id="bib15"><number>[15]</number> Argyris Constantinides, Marios Belk, Christos Fidas, and Andreas Pitsillides. 2020. An eye gaze-driven 
metric for estimating the strength of graphical passwords based on image hotspots. In Proceedings of the 25th International Conference on 
Intelligent User Interfaces, 33–37.</bib> 
<bib id="bib16"><number>[16]</number> Alan Cooper. 1999. The Inmates Are Running the Asylum: Why High Tech Products Drive Us Crazy and 
How to Restore the Sanity (1 edition ed.). Sams - Pearson Education, Indianapolis, IN.</bib> 
<bib id="bib17"><number>[17]</number> Giuseppe Desolda, Rosa Lanzilotti, Danilo Caivano, and Maria F. Costabile. 2021. An experience on 
remote testing exploiting new web technology. In INTERACT’21 Workshop on Remote user Testing – Experiences and Trends, Bari, Italy.</bib> 
<bib id="bib18"><number>[18]</number> A. T. Duchowski. 2009. Eye Tracking Methodology: Theory and Practice. Springer, London.</bib> 
<bib id="bib19"><number>[19]</number> Hermann Ebbinghaus. 2013. Memory: A contribution to experimental psychology. Annals of 
neurosciences 20, 4 (2013), 155.</bib> 
<bib id="bib20"><number>[20]</number> Achim Ebert, Shah Rukh Humayoun, Norbert Seyff, Anna Perini, and Simone D.J. Barbosa (Eds.). 2016. 
Usability- and Accessibility-Focused Requirements Engineering. Springer International Publishing, Cham. DOI:https://doi.org/10.1007/978-3-319-
45916-5</bib> 
<bib id="bib21"><number>[21]</number> Joy Goodman-Deane, Sam Waller, Dana Demin, Arantxa González-de-Heredia, Mike Bradley, and John 
P. Clarkson. 2018. Evaluating Inclusivity using Quantitative Personas. In In the Proceedings of Design Research Society Conference 2018, Limerick, 
Ireland. DOI:https://doi.org/10.21606/drs.2018.400</bib> 
<bib id="bib22"><number>[22]</number> Thomas Grindinger, Andrew T. Duchowski, and Michael Sawyer. 2010. Group-wise Similarity and 
Classification of Aggregate Scanpaths. In Proceedings of the 2010 Symposium on Eye-Tracking Research & Applications (ETRA ’10), ACM, New York, 
NY, USA, 101–104. DOI:https://doi.org/10.1145/1743666.1743691</bib> 
<bib id="bib23"><number>[23]</number> Jonathan Grudin. 2006. Why Personas Work: The Psychological Evidence. In The Persona Lifecycle, 
John Pruitt and Tamara Adlin (eds.). Elsevier, 642–663. DOI:https://doi.org/10.1016/B978-012566251-2/50013-7</bib> 



<bib id="bib24"><number>[24]</number> Jonathan Grudin and John Pruitt. 2002. Personas, Participatory Design and Product Development: An 
Infrastructure for Engagement. In Proceedings of Participation and Design Conference (PDC2002), Sweden, 8.</bib> 
<bib id="bib25"><number>[25]</number> Jacek Gwizdka, Rahilsadat Hosseini, Michael Cole, and Shouyi Wang. 2017. Temporal dynamics of eye-
tracking and EEG during reading and relevance decisions. Journal of the Association for Information Science and Technology 68, 10 (October 2017), 
2299–2312. DOI:https://doi.org/10.1002/asi.23904</bib> 
<bib id="bib26"><number>[26]</number> Charles G. Hill, Maren Haag, Alannah Oleson, Chris Mendez, Nicola Marsden, Anita Sarma, and 
Margaret Burnett. 2017. Gender-Inclusiveness Personas vs. Stereotyping: Can We Have it Both Ways? In Proceedings of the 2017 CHI Conference, 
ACM Press, Denver, Colorado, USA, 6658–6671. DOI:https://doi.org/10.1145/3025453.3025609</bib> 
<bib id="bib27"><number>[27]</number> Bernard J. Jansen, Soon-gyo Jung, and Joni Salminen. 2019. Creating Manageable Persona Sets from 
Large User Populations. In Extended Abstracts of the 2019 CHI Conference on Human Factors in Computing Systems, ACM, Glasgow, United Kingdom, 
1–6. DOI:https://doi.org/10.1145/3290607.3313006</bib> 
<bib id="bib28"><number>[28]</number> Bernard J. Jansen, Soon-gyo Jung, and Joni Salminen. 2020. From flat file to interface: Synthesis of 
personas and analytics for enhanced user understanding. Proceedings of the Association for Information Science and Technology 57, 1 (October 
2020). DOI:https://doi.org/10.1002/pra2.215</bib> 
<bib id="bib29"><number>[29]</number> Bernard J. Jansen, Joni Salminen, and Soon-gyo Jung. 2020. Data-Driven Personas for Enhanced User 
Understanding: Combining Empathy with Rationality for Better Insights to Analytics. Data and Information Management 4, 1 (2020), 1–17. 
DOI:https://doi.org/10.2478/dim-2020-0005</bib> 
<bib id="bib30"><number>[30]</number> Bernard Jansen, Joni Salminen, Soon-gyo Jung, and Kathleen Guan. 2021. Data-Driven Personas (1st 
ed.). Morgan & Claypool Publishers. Retrieved February 10, 2021 from 
https://www.morganclaypool.com/doi/abs/10.2200/S01072ED1V01Y202101HCI048</bib> 
<bib id="bib31"><number>[31]</number> Joel Järvinen and Heikki Karjaluoto. 2015. The use of Web analytics for digital marketing performance 
measurement. Industrial Marketing Management 50, Supplement C (October 2015), 117–127. 
DOI:https://doi.org/10.1016/j.indmarman.2015.04.009</bib> 
<bib id="bib32"><number>[32]</number> Soon-gyo Jung, Jisun An, Haewoon Kwak, Moeed Ahmad, Lene Nielsen, and Bernard J. Jansen. 2017. 
Persona Generation from Aggregated Social Media Data. In Proceedings of the 2017 CHI Conference Extended Abstracts on Human Factors in 
Computing Systems (CHI EA ’17), ACM, Denver, Colorado, USA, 1748–1755.</bib> 
<bib id="bib33"><number>[33]</number> Soon-Gyo Jung, Joni Salminen, Jisun An, Haewoon Kwak, and Bernard J Jansen. 2018. Automatically 
Conceptualizing Social Media Analytics Data via Personas. In Proceedings of the International AAAI Conference on Web and Social Media (ICWSM 
2018), San Francisco, California, USA, 2.</bib> 
<bib id="bib34"><number>[34]</number> Soon-gyo Jung, Joni Salminen, and Bernard J. Jansen. 2019. Personas Changing Over Time: Analyzing 
Variations of Data-Driven Personas During a Two-Year Period. In Extended Abstracts of the 2019 CHI Conference on Human Factors in Computing 
Systems - CHI EA ’19, ACM Press, Glasgow, Scotland Uk, 1–6. DOI:https://doi.org/10.1145/3290607.3312955</bib> 
<bib id="bib35"><number>[35]</number> Soon-Gyo Jung, Joni Salminen, and Bernard J. Jansen. 2020. Giving Faces to Data: Creating Data-Driven 
Personas from Personified Big Data. In Proceedings of the 25th International Conference on Intelligent User Interfaces Companion  (IUI ’20), 
Association for Computing Machinery, Cagliari, Italy, 132–133. DOI:https://doi.org/10.1145/3379336.3381465</bib> 
<bib id="bib36"><number>[36]</number> Soon-gyo Jung, Joni Salminen, and Bernard J Jansen. 2021. Implementing Eye-Tracking for Persona 
Analytics. In ETRA ’21 Adjunct: ACM Symposium on Eye Tracking Research and Applications, ACM, Virtual conference, 1–4. 
DOI:https://doi.org/10.1145/3450341.3458765</bib> 
<bib id="bib37"><number>[37]</number> Soon-gyo Jung, Joni Salminen, and Bernard J. Jansen. 2021. Persona Analytics: Implementing Mouse-
tracking for an Interactive Persona System. In Extended Abstracts of ACM Human Factors in Computing Systems - CHI EA ’21, ACM, Virtual 
conference. DOI:https://doi.org/10.1145/3411763.3451773</bib> 
<bib id="bib38"><number>[38]</number> Soon-gyo Jung, Joni Salminen, Haewoon Kwak, Jisun An, and Bernard J. Jansen. 2018. Automatic 
Persona Generation (APG): A Rationale and Demonstration. In CHIIR ’18: Proceedings of the 2018 Conference on Human Information Interaction & 
Retrieval, ACM, New Jersey, USA, 321–324. DOI:https://doi.org/10.1145/3176349.3176893</bib> 
<bib id="bib39"><number>[39]</number> Pascal J. Kieslich, Felix Henninger, Dirk U. Wulff, Jonas MB Haslbeck, and Michael Schulte-
Mecklenbeck. 2019. Mouse-Tracking: A Practical Guide to Implementation and Analysis 1. In A handbook of process tracing methods. Routledge, 
111–130.</bib> 
<bib id="bib40"><number>[40]</number> Ari Kolbeinsson, Erik Brolin, and Jessica Lindblom. 2021. Data-Driven Personas: Expanding DHM for a 
Holistic Approach. In International Conference on Applied Human Factors and Ergonomics, Springer, 296–303.</bib> 
<bib id="bib41"><number>[41]</number> Dannie Korsgaard, Thomas Bjørner, Pernille Krog Sørensen, and Paolo Burelli. 2020. Creating user 
stereotypes for persona development from qualitative data through semi-automatic subspace clustering. User Model User-Adap Inter 30, 1 (March 
2020), 81–125. DOI:https://doi.org/10.1007/s11257-019-09252-5</bib> 
<bib i