Your Favorite Gameplay Speaks Volumes about You: Predicting User Behavior and Hexad Type^††thanks: The authors’ pre-print version of the paper accepted at the conference of HCI International 2023

Gamification has recently become an independent field of study in the HCI communities, and it historically has strong ties to the gaming industry. The term gamification is often defined as the use of game design elements (such as badges, points, and leaderboards) in non-game contexts (such as education, business, health, and crowdsourcing) [12, 43, 24]. According to the extant literature, gamification has the ability to change user behavior [29, 3] and can have both positive (e.g., [19, 31]) and negative effects (e.g., [19, 20, 22]) on user activities. This relevant prior knowledge motivates us to explore the possibility of using players’ game-related data to predict their most dominant user behavior in gamified (but non-game) environments.

Beyond User Behavior. Aside from user behavior, another vital topic worth discussing in any gamified environment is determining users’ gamification preferences. Here, we are curious to find out whether game-related data can also be used to predict users’ gamification preferences. To investigate users’ gamification preferences more systematically (and practically), we employ Hexad, which is a well-established empirical framework often used to identify and understand users’ gamification preferences (or Hexad type) [39, 32, 23].

Hexad Framework. According to the Hexad framework, different Hexad types include (1) Philanthropists (provoked by: purpose), (2) Socializers (provoked by: relatedness), (3) Free Spirits (provoked by: autonomy), (4) Achievers (provoked by: competence), (5) Players (provoked by: extrinsic rewards), and (6) Disruptors (provoked by: their urge to change).

In recent years, an increasing number of HCI researchers have explored many inventive methods for predicting Hexad types other than the typical surveying approach that is developed by Tondello et al. [47]. Each method has its own unique benefits and drawbacks and, when necessary, can be used in a variety of situations and contexts. For example, a research conducted by Altmeyer et al. showed that user-generated smartphone data can help predict Hexad types and potentially tailor gamification without requiring direct user interaction [1]. Nevertheless, as is evident, some people may have privacy and security concerns about sharing their smartphone data with gamification designers and developers. In another study, Kimpen et al. conducted an expert consensus study with a group of 11 experts and argued that banking data could as well be helpful in predicting Hexad types [27]. With this background in mind, our research presents a new way to predict gamers’ Hexad types based on the information about their game and gameplay preferences when such information is available.

The non-game context chosen for investigation in this article is Stack Exchange, which is a popular example of Community Question Answering Websites (CQAs). CQAs such as Stack Exchange, Quora, and Zhihu are among the most important Computer Supported Cooperative Work (CSCW) applications that have been in the interest scope of the HCI and gamification research communities for a long time [22, 20, 44]. These platforms help millions of people worldwide ask and answer questions online and serve as helpful knowledge repositories on different topics [44]. As shown in Figure 1, CQAs typically use various gamification elements such as badges, points, and leaderboards to incentivize user participation and prevent churn (i.e., user dropouts) [34].

In summary, our main contributions to the existing body of knowledge in this area are as follows:

• •

  According to our findings, there is a notable correlation between people’s preferences for specific types of gameplay and their behavioral characteristics in non-game (but gamified) contexts, as well as their preferences for specific types of gamification.

• •

  According to our results, all player archetypes can (to some extent) help predict user behavior in Stack Exchange (a non-game gamified context) and their Hexad type better than the baseline of a random labeler. However, of all the gamer archetypes studied in this paper, BrainHex is the one that has the best performance.

• •

  Our research can help eliminate (or mitigate) the need for parallel surveys of people’s preferences about games and gamification. This study shows that game-related data can be used to make reasonable inferences about people’s dominant behavioral attributes and preferences for gamification. Thus, our research findings could help game/gamification researchers and survey respondents save time, money, and energy.

• •

  Last but not least, our findings encourage HCI researchers and practitioners to think outside the box and apply their existing knowledge resources/tools (such as player archetypes) to solve broader and different types of problems (e.g., predicting Hexad user type). In the ever-expanding HCI community of today, finding creative and novel solutions to domain-related issues appears to be quite essential.

Personalization and customization of the gaming experience are essential to ensure that gamers have the most fun and enjoy their time. Three of the most commonly cited gamer archetypes used for this purpose in the HCI and gameplay research communities are Bartle [5], Big Five [26], and BrainHex [35]. A great deal can be gleaned about gamers from each of these archetypes.²²2In this text, the terms typology, archetype, model, and behavioral traits are sometimes used interchangeably.

$\Box$ Bartle’s Taxonomy of Player Types. According to the gamer taxonomy developed by Bartle in 1996, gamers who play in multiplayer virtual worlds can be divided into four types: achievers, socializers, explorers, and killers. As shown in Figure 2, these categories are determined by classifying gamer preferences along two dimensions, i.e., players-world spectrum (horizontal axis) and acting-interacting spectrum (vertical axis). Each gamer type is defined as follows [41]:

Achievers are defined as those who like to act in the world. They are gamers who want to feel that they have achieved something significant (often scarce) in the game; they place great value on the element of competence described in the Self-Determination Theory (SDT). According to this theory, people need a sense of competence (being skilled and effective), relatedness (being connected to others), and autonomy (being master of their behavior) in order to be satisfied enough to perform certain actions [49, 22].

Explorers, on the other hand, like to interact with the world. They are gamers who want to engage with the full scope of the system and explore its boundaries and hidden secrets; they naturally value the element of autonomy in self-determination theory the most.

Socializers enjoy interacting with other players. They are gamers who want to use the system to meet new people and make new friends; socializers, according to the self-determination theory, are interested in the element of relatedness. Finally, killers are gamers who take pleasure in doing things to other people or acting on other players. They enjoy competing with others, and the competence element of the self-determination theory appeals to them the most.

Currently, the most common empirical method for determining an individual’s Bartle type is Dr. Matthew Barr’s survey method, known as Bartle test.³³3http://matthewbarr.co.uk/bartle/

$\Box$ Big Five Personality Traits.⁴⁴4Also called Five-Factor Model of Personality In psychology, personality traits go beyond the game/gamification world [23] and are understood as patterns of thought, feeling, and behavior that are relatively persistent throughout a person’s life [17]. Developed and disseminated primarily in the 1980s and 1990s, the model of Big Five personality traits include conscientiousness, agreeableness, neuroticism, openness, and extraversion [23]. Below, we summarize the main characteristics of each personality trait:

• •

  Conscientiousness shows the proclivity for self-control, discipline, tenacity, organization, and responsibility toward others.

• •

  Agreeableness indicates the emotional feelings such as empathy and concern for the needs of others, altruism, trust, and cooperation.

• •

  Neuroticism describes a person’s proclivity to be under psychological pressure (instability) and to experience negative emotions such as depression, anxiety, and anger.

• •

  Openness⁵⁵5Also called Openness to Experience describes a person’s imaginative, creative, curious, risk-taking, and adventurous tendencies.

• •

  Extraversion (or extroversion) describes a person’s tendency to be gregarious, friendly, warm, and enthusiastic about activities that involve interaction with others, such as making new friends or participating in group projects.

TIPI (Ten Item Personality Measure) is currently the most widely used method for measuring a person’s Big Five personality traits [15].

$\Box$ BrainHex Player Typology. The BrainHex typology was developed in the early 2010s. It was inspired by several sources and references, including earlier typology approaches, discussions of play patterns, the literature on gameplay emotions, and most importantly, the relevant neurobiological studies [6, 35]. This typology divides gamers’ playing motivations into seven categories [35]: Seeker (motivated by exploration), Survivor (driven by fear), Daredevil (motivated by excitement), Mastermind (motivated by strategy), Conqueror (motivated by challenge), Socializer (motivated by social relationships), and Achiever (motivated by goal achievement). In the following, we summarize and explain each of these categories based on the explanations provided by Lennart Nacke et al. [35]:

• •

  The Seeker loves to explore the environment and have adventures. Endormorphin is released when a seeker’s brain is exposed to rich patterns of interpretable information (often sensory), activating the pleasure center in their brain.

• •

  The Survivor takes pleasure when exposed to a high level of suspense induced by fear or anticipation of frightening situations, at least in the context of fictional activities (e.g., horror movies and video games). Among survivors, the arousal neurotransmitter epinephrine amplifies the impact of dopamine triggered by the receipt of rewards. However, it still needs to be determined whether the pleasure of fear should be evaluated by the intensity of the fright itself or by the relief felt afterward.

• •

  The Daredevil revels in the thrill and peril of playing on the edge and taking risks (e.g., playing at extreme heights or participating in a car race at extreme speeds). Similar to Survivors, epinephrine is a reward enhancer in daredevils.

• •

  The Mastermind takes pleasure in solving complex puzzles, developing clever strategies, and making the most efficient decisions. The pleasure and decision centers of a mastermind’s brain are closely connected, and good decisions are inherently rewarded. The mastermind archetype is very similar to Bartle’s explorer type.

• •

  The Conqueror describes a challenge-seeking gamer who is not content with easy victories. Difficult situations stimulate the release of epinephrine (adrenalin), associated with arousal and excitement, and norepinephrine, associated with anger. Regardless of gender, testosterone is also thought to play a role in this archetype. According to the literature, the conqueror archetype is weakly related to Bartle’s killer type.

• •

  The Socializer enjoys interacting with others, hanging out with trusted people, talking with them, and assisting them. This archetype is linked to the social center of the brain, which is the primary neural source of oxytocin, a neurotransmitter associated with trust. The name of this archetype alludes to Bartle’s socializer archetype.

• •

  The Achiever describes a goal-oriented gamer motivated by long-term achievements. In this archetype, the release of dopamine and stimulation of the brain’s pleasure center contribute to a sense of personal fulfillment. The name of this archetype alludes to Bartle’s achiever archetype.

Currently, most HCI and gameplay researchers and practitioners use the official International Hobo Questionnaire [33] to determine the BrainHex types of their research participants.

In recent years, gamification has emerged as an essential research topic in the fields of HCI, CSCW, and Learning at Scale. Simply put, gamification is the application of game design elements such as badges, points, and leaderboards in non-game contexts [12]. However, the mere addition of badges, points, and leaderboards to a system/service does not guarantee user engagement or lead users toward a desired behavior [20, 22, 19]. According to Nicholson, a good gamification design should support its target users in making meaningful connections between their personal goals and the use of gamification [36]. This is important because it can help improve users’ behavioral outcomes and experiences. The literature suggests that a tailored gamification design can help promote the formation of such perceptual connections [36, 23, 22].

$\Box$ Hexad [47] is a framework widely used by the gamification research community to adapt gamification in various systems and applications. It is based on the self-determination theory (SDT) [49]. Hexad framework helps gamification researchers and designers understand users’ motivations for interacting with different types of gamification elements in different environments. Below, we review the main characteristics of each Hexad type (according to [47]):

• •

  Philanthropists seek purpose and meaning. They are altruistic and willing to help or give without expecting anything in return.

• •

  Socializers are motivated by relatedness (as described in SDT). They want to interact with others and form social bonds.

• •

  Free Spirits are motivated by autonomy (as described in SDT). They enjoy creation and exploration without any controls or restrictions.

• •

  Achievers seek competence (as described in SDT). They want to prove their worth by taking on challenging tasks and moving up in a system.

• •

  Players are after the extrinsic rewards. They are willing to engage in any activity (regardless of the type) in order to receive a reward within a system.

• •

  Disruptors are driven by their will to initiate a change (either positive or negative) in the system. They enjoy challenging the system.

The survey approach developed by Tondello et al. [47] is the most common method for determining a user’s Hexad type. However, as the Hexad framework becomes more popular, new approaches to predicting the Hexad types are being developed that can be useful in a variety of situations based on the availability of different types of data [2, 1, 27]. Our research here contributes to this field of study by introducing a novel method for predicting Hexad types based on people’s gamer archetypes.

To increase the diversity of our research participants, we used a variety of sources to recruit them, including Stack Exchange chat rooms and social media platforms such as Facebook and Twitter.⁶⁶6Two moderators from Stack Exchange assisted us throughout the recruitment process. Inclusion criteria were 1) age over 18 years; 2) more than six months of active participation in at least one Stack Exchange forum;⁷⁷7This measure allows us to avoid studying the behavior of churned users [20]. and 3) having at least some occasional engagement with games and different gamification schemes (self-reported via seven-point Likert scales).⁸⁸8i.e., 3/7 on the Likert scale, where 7/7 points out the maximum engagement.

In this way, we recruited 67 Stack Exchange users (23 female) with different demographics and varying levels of exposure to games and gamification. The age range of the participants is between 18 and 45 years (mean = 28.21, SD = 9.36). The ethnic distribution of our research participants is as follows: Asian (n = 27), White (n = 14), Hispanic and Latino/a (n = 10), Black or African American (n = 7), Native American (n = 2), mixed race (n = 2), and other races (n = 5). According to our initial screenings, on the seven-point Likert scale, the average user engagements of our participants with games and various gamification schemes (prior to our research) are 4.82 (SD = 1.12) and 4.30 (SD = 0.93), respectively.⁹⁹9Again, 7/7 shows the maximum engagement.

All of our research procedures have been approved by the IRB at the Hong Kong University of Science and Technology. After obtaining informed consent from all research participants, we asked them (via email) to complete four online surveys (three for gamer archetypes and one for Hexad) within one to two days of receiving the email. The HKUST Qualtrics website was used to collect all responses, allowing participants to take breaks between (and during) surveys and to measure the time it took them to complete each survey. Upon successfully receiving the answers, each participant was given HK$ 40 as a token of appreciation. All survey responses for each participant were stored under their unique Stack Exchange IDs for further analysis. The average completion times (in minutes) for the Bartle, Big Five, and BrainHex are 22.83 (SD = 7.38), 15.25 (SD = 6.94), and 29.50 (SD = 5.66). The average time to complete the Hexad survey is 20.25 minutes (SD = 4.18).

Stack Exchange Data Explorer¹⁰¹⁰10https://data.stackexchange.com/ was used to collect each participant’s CQA data using their Stack Exchange IDs. This data includes information about each participant’s posting behavior, i.e., their posts’ exact timestamps and peer-evaluated posting quality scores.¹¹¹¹11The difference between the number of up-votes and down-votes Different types of posts in the CQA of Stack Exchange include questions, answers, and comments. In this context, user posts closed and removed by community moderators are regarded as spam. Closed posts with negative scores are also considered spam.

In addition to users’ posting data, we also collected their metadata on other types of activities, i.e., their voting behavior and edit services.

The oldest post in our dataset was made on July 10th, 2015, and the most recent post was made on September 18th, 2022. The average reputation of research participants in this work is 9,656.04 (SD = 10,067.18). And the number of users in the categories of novice users, low reputation users, established users, and trusted users are n = 8, n = 17, n = 23, and n = 19 respectively. The average number of questions, answers, comments, and spam posts per individual is 43.76 (SD = 26.12), 52.10 (SD = 29.18), 68.11 (SD = 35.92), and 4.08 (SD = 2.23), respectively. Furthermore, the average number of votes and edits per person is 496.92 (SD = 311.55) and 30.74 (SD = 26.62), respectively. The most dominant behavior (action) of each participant in our study is determined by counting the frequency with which user behaviors are repeated.

In keeping with the exploratory spirit of our study, we first conduct and present the results from a thorough Pearson correlation analysis on data from collected game and gamification surveys and data from the participants’ CQA activities.

Following that, we train the computational models for our prediction tasks (i.e., predicting the dominant user behavior and Hexad type). To this end, we employ a high-performance machine learning model known as Extreme Gradient Boosting (which is also called XGBoost). The reason for using XGBoost is its simplicity and strength in reducing the training bias [21]. We build and compare our models based on various input features derived from our participants’ previous survey responses (i.e., Bartle, Big Five, and BrainHex characters). We compare our models (also) against the baseline of a random labeler. The reason for using a random labeler here is because, to the best of our knowledge, there are currently no other (suitable/better) baselines that can be used for comparison with our work. Due to the limited amount of our sample data (for training purposes), we use the metric of five-fold cross-validation error to evaluate and report the performance of all models in this paper. After dividing our data into different folds (for the purpose of cross-validation), we use an up-sampling technique (based on Bootstrapping) to address the issue of imbalanced data labels. All models in our study are created and tested on a local PC equipped with an Intel i7-5820k CPU (6 cores @3.3GHz) and 32GB RAM using the popular Python library scikit-learn.

We conducted n = 192 Pearson correlation tests to uncover the interesting relationships between participants’ game-based characters and behavioral profiles in the CQA and between participants’ game-based characters and gamification-based characters in general (i.e., Hexad type). For the purpose of illustration, Figure 3 depicts the correlation plots (tested with a statistical confidence of 99.95) between seeker participants’ BrainHex measures (intensities) and the frequencies of seeker users’ posting behaviors, i.e., questions (Pearson’s r value = + 0.24), answers (Pearson’s r value = + 0.61), comments (Pearson’s r value = - 0.08), and spam posts (Pearson’s r value = - 0.02). We observe a statistically significant positive correlation between the seekers’ BrainHex measures and their tendencies to post (more) questions and answers inside the CQA community. However, the statistical significance of the negative correlation between seekers’ BrainHex measures and their frequencies of commenting and spamming still needs to be proved.

Likewise, Figure 4 shows the correlation plots (tested with a statistical confidence of 99.95) between seeker participants’ BrainHex and Hexad type measures (intensities). The Pearson r values resolved for the Hexad type correlations are as follows: Philanthropist (r = + 0.01), Socializer (r = - 0.03), Free Spirit (r = + 0.25), Achiever (r = + 0.19), Player (r = + 0.31), and Disruptor (r = + 0.18). In the case of the seeker archetype (from BrainHex), only the positive correlations reported for free spirit and player Hexad measures are statistically significant.

It is worth noting here that different surveys often yield different scales for indicating user characteristics and that there is often a considerable discrepancy between the frequency with which users engage in various activities. With this point in mind, we have therefore normalized all numbers in all figures (and analyses) to lie within the range [0,1] for greater clarity and better comparability.

The following are the most important statistically significant findings and conclusions from our correlation analysis:

Bartle. The frequency of questions and answers in the CQA correlates positively with Bartle’s achiever and Socializer measures. We found that the amount of comments and spam posted in the CQA is positively correlated with Bartle’s explorer and killer measures. Bartle’s achiever measure correlates positively with Hexad’s free spirit, achiever, and player intensities. There is also a positive correlation between Bartle’s killer measure and Hexad’s achiever, player, and disruptor metrics (indicators).

Big Five. The number of questions asked, answers given, comments written, and edits made in the CQA correlate positively with the Big Five measures of openness and extraversion. One of the Big Five dimensions, conscientiousness, correlates positively with the frequencies of voting and editing, whereas neuroticism correlates negatively with both. Furthermore, the Hexad measures socializer and philanthropist correlate positively with the Big Five levels of extraversion and agreeableness. Also, there appears to be a positive relationship between openness and Hexad’s free spirit measure.

BrainHex. The frequency of CQA posts (of any type) correlates positively with the BrainHex mastermind and conqueror measures. The frequency of questions and answers is proportional to the BainHex seeker measure. The BrainHex daredevil metric is positively correlated with the frequency of votes and spam. The socializer measure, on the other hand, is positively correlated with the frequency of comments and answers. The number of questions a participant asks correlates positively with their BrainHex survival measure, while the number of answers correlates negatively with it. Nonetheless, the BrainHex achiever measurement is positively correlated with the number of answers provided by the participants identified with that character. In addition, the BrainHex achiever metric is positively correlated with Hexad’s philanthropist, achiever, and player intensities. The seeker archetype in BrainHex, as noted before, is positively correlated to the measurements of free spirit and player archetypes in Hexad. In BrainHex, the daredevil metric is positively correlated with the player and disruptor measurements in Hexad. The socializer measure in the BrainHex is positively related to the philanthropist and socializer measures in Hexad.

Table 1 summarizes the performances of the different prediction models based on their average cross-validation error. Overall, the combination of XGBoost and the BrainHex user type measures outperforms all other competing models in predicting the most dominant user behavior. This combination outperforms the Big Five and Bartle models by 26.13% and 7.38%, respectively. In addition, BrainHex consistently outperforms all other models at the categorical level, with the exception of the editing and voting tasks (where the BrainHex model falls slightly behind the Big Five model). Furthermore, we can see that all gamer archetypes can significantly outperform the prediction of a random labeler. Hence, even in the absence of BrainHex data, the Big Five and then the Bartle model could be used as helpful alternatives. However, the accuracy issue should be considered for more sensitive tasks and decisions.

Similarly, Table 2 summarizes the performances of the various prediction models when predicting the (most dominant) Hexad type. Again, the combination of XGBoost and the BrainHex user type measures outperforms all other competing models in accurately predicting the Hexad types. According to the total performance results, this model (XGBoost + BrainHex) reduces error by 15.60% compared to the Bartle archetype and 3.77% compared to the Big Five model. Furthermore, with the exception of the philanthropist user type (where BrainHex falls 1.85% short of the Big Five model), the BrainHex model (also) outperforms all other competing models at all categorical levels. We should mention here that all gamer archetypes (i.e., Bartle, Big Five, and BrainHex) can significantly outperform the prediction of a random labeler. Hence, in the absence of BrainHex data, the Big Five and then the Bartle model could still be used as viable (helpful) substitutes.

There are several critical limitations to our research in this paper that should be mentioned. The first and most obvious limitation is the small number of participants in the study. Unfortunately, it is becoming increasingly difficult to recruit participants for research studies such as ours that require lengthy and time-consuming surveys to be completed. Also, due to the limited number of data samples, we decided not to use Deep Learning or other advanced neural network models for our prediction tasks. Future work could focus on collecting more data (on a larger scale) and improving the machine-learning component of the models. In addition to these caveats, the context of our research study itself is a major limitation. Obviously, Stack Exchange is not the best representative for all kinds of online gamified platforms on the Internet. Hence, it is necessary that future studies cross-check the findings of our work also on other (and more diverse) platforms.

In the end, we present four interesting research directions for future work. 1) Future player surveys and gamification archetypes (or data collection methods) will need to be less intrusive and collect as little personal information as possible. In addition, the advent of platforms such as the metaverse has necessitated the development of newer and more secure techniques for collecting and disseminating human data. As an idea for future research, we would like to see if it is possible to use people’s metaverse-based peripheral data to infer aspects of their personality in order to provide them with better and more tailored gaming/gamification services. 2) Understanding how users and gamification experts perceive and reason about the relationships between different gamer archetypes, CQA behaviors, and Hexad is another interesting research direction for future work that merits being thoroughly explored. Comprehensive qualitative research methods, such as inductive and deductive content analysis, are likely to yield better results in this type of investigation. 3) People’s interest in the playful aspects of a gamified platform may be a good indicator of their willingness to stay on that site for a more extended period of time. Future research could examine and explore this idea further to determine if information about users’ gamer and gamification archetypes could help gamified platforms and services detect when a user will churn their websites/services (or become less active/loyal), perhaps in a more timely and accurate manner. 4) Although we found some interesting correlations between gamer archetypes, user behaviors, and Hexad types, our study does not demonstrate or assert any causal relationships. Future research could use controlled experimental designs to determine the presence (or absence) of any causal relationships.