Your Gameplay Says it All: Modelling Motivation in Tom Clancy's The Division - arXiv
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Your Gameplay Says it All:
Modelling Motivation in Tom Clancy’s The Division
David Melhart∗ , Ahmad Azadvar† , Alessandro Canossa† , Antonios Liapis∗ and Georgios N. Yannakakis∗
∗ Institute of Digital Games, University of Malta
{david.melhart, antonios.liapis, georgios.yannakakis}@um.edu.mt
† Massive Entertainment a Ubisoft Studio
{ahmad.azadvar, alessandro.canossa}@massive.se
Abstract—Is it possible to predict the motivation of players gameplay data—and her motivation. In particular, we assume
arXiv:1902.00040v1 [cs.LG] 31 Jan 2019
just by observing their gameplay data? Even if so, how should that solely behavioural data from a player’s gameplay would
we measure motivation in the first place? To address the above yield accurate predictors of motivation in games. To define
questions, on the one end, we collect a large dataset of gameplay
data from players of the popular game Tom Clancy’s The Division. motivation we rely theoretically on Self Determination Theory
On the other end we ask them to report their levels of competence, [13]—a well-established positive psychological framework
autonomy, relatedness and presence using the Ubisoft Perceived for motivation—and examine four core factors: competence,
Experience Questionnaire. After processing the survey responses autonomy, relatedness and presence, which is often associated
in an ordinal fashion we employ preference learning methods, with the theory in the domain of videogames [1].
based on support vector machines, to infer the mapping between
gameplay and the four motivation factors. Our key findings To study motivation quantitatively we are grounded on
suggest that gameplay features are strong predictors of player recent developments in motivation measurement tools, namely
motivation as the obtained models reach accuracies of near the Ubisoft Perceived Experience Questionnaire (UPEQ) [14],
certainty, from 93% up to 97% on unseen players. which was developed as a game-specific tool observing player
Index Terms—Self determination theory, affective computing, motivation. To infer the relationship between player motivation
digital games, player modelling, preference learning
and gameplay we collect data from more than 400 players
of Tom Clancy’s The Division (Ubisoft, 2016). We process
I. I NTRODUCTION
and aggregate this data and collect surveys on the players’
The holy grail for game design and development is to create motivation in relation to the game independently. We use
engaging experiences for the players. Players who are self- the UPEQ questionnaire to measure players’ general levels
motivated to return to the game and keep playing are critical of competence, autonomy, relatedness and presence in the
for a game’s success [1]–[3]. The central role of motivation game. Given the subjective nature of the reported notions we
for the design of games, and the experiences they elicit, adopt the second-order data processing approach [15] and
has been highlighted by a growing number of studies which we process the reported UPEQ Likert-scale values of the
adopt psychological theories of motivation within games [4]– players as ordinal data, and not as scores. We then apply
[8]. Such studies, however, follow a top-down integration of simple statistical rank-based models and preference learning
phenomenological models of motivation, which aim to identify [16] methods based on support vector machines to infer the
and explain stereotypical player behaviour. Over the last function between gameplay and reported factors of motivation.
decade, games user research and industry-based game testing Our results suggest that factors of reported motivation can be
has shifted its focus towards quantitative approaches based on predicted with high accuracy just relying on a few high-level
player analytics [9], [10] with the aim to shed more light onto gameplay features. In particular, the nonlinear machine learned
the understanding of player behaviour and experience. These models manage to predict the four motivation factors of unseen
approaches focus mainly on either clustering players based players with at least 93% accuracy; the best model reaches an
on their behavioural patterns or predicting objectively-defined accuracy of 97% for presence. The obtained results add to the
aspects of their gameplay behaviour for monetization purposes existing evidence for the benefits of ordinal data processing
(e.g. churn prediction) [3]. In that regard, the majority of on subjectively-defined notions [15] and they also validate
approaches that aim to capture aspects of player experience that motivation can be captured qualitatively with supreme
(such as engagement or motivation) based on player analytics accuracy in the examined game only based on behavioural
remain qualitative, given the complexity of measuring subjec- high-level data of playing.
tive notions of user experience in games [11]. This paper is novel in a number of ways. First, this is
Motivated by the lack of quantitative studies on the relation- the first time player motivation is modelled computationally
ship between motivation and play, in this paper we introduce a only through gameplay data in games. Second, we introduce a
data-driven player modelling approach [12] by assuming there second-order [15] methodology for treating Likert-scale scores
is an unknown underlying function between what a player which are used frequently in game testing and games user
does in the game behaviourally—as manifested through her research at large. The ordinal approach we adopt compares
the subjective scores of all players with each other and hence use Yee’s [34] player motivation typology to enhance churn
generates combinatorially very large datasets based only on prediction, while Shim et al. [35] use motivational survey data
small sets of participants. The approach is also effective in to model player enjoyment. Birk et al. [36], also incorporated
eliminating reporting biases of respondents, thereby better ap- motivational survey data along with top-down personality and
proximating the ground truth of reported motivation. Third, we player profiles to model enjoyment and effort. Although Birk
model aspects of player motivation using preference learning et al. include high-level telemetry of progression (game stage)
based solely on a small number of key gameplay features. as a control variable in their experiments, they rely mainly
Finally, for the first time we evaluate the method in Tom on survey data. In contrast to the above studies—which apply
Clancy’s The Division (Ubisoft, 2016) on over 400 players motivational survey data as top-down domain knowledge that
and the predictive capacity of the motivation models for this forms the input of enjoyment models—we instead focus on
game reach near certainty (i.e., over 93% of accuracy). predicting motivation based on game metrics alone. Similarly
to this study, Canossa et al. [5] applied regression models
II. BACKGROUND : M EASURING AND M ODELLING of telemetry to predict psychological survey data; however,
M OTIVATION that study was not aimed at player motivation modelling but
This section gives an overview of related research on player instead used the Reiss Motivational Profile [37] as a tool for
modelling and motivation studies in games (Section II-A) personality profiling based on personal motivational drives.
and then it introduces the fundamental principles of Self In this paper we focus on the structure of motivation by
Determination Theory (SDT) and the notion of presence within considering four of its core manifestations but our models do
the framework of UPEQ (Section II-B). The section ends not rely on personality profiles of the players. Instead, we
with a discussion on the strengths of preference learning for only focus on and try to infer the unknown mapping between
modelling subjectively-defined psychological constructs such gameplay features and motivation.
as motivation (Section II-C).
B. From Theory to Measures of Motivation
A. Player Modelling and Player Motivation The self-determination theory (SDT) is a well-established
Player modelling is one of the primary foci of AI research in positive psychology theory of the facilitation of motivation
the field of videogames [3] and generally concerned with the based on the work of Deci and Ryan [38], which has been
prediction of player behaviour or other cognitive and affective adopted to a wide variety of domains, including education
processes. As games offer a complex yet constrained environ- [39], job satisfaction [40], parenting [41], health and exercise
ment for modelling dynamic interactions, user modelling is [42], and videogames [1], [8], [43], [44]. The core theory was
also gaining more and more traction among industry practi- developed to contrast earlier frameworks of motivation as a
tioners, who use analytics-based player modelling methods to unitary concept [45], [46], by focusing on the dichotomy of
inform game design, development, and stakeholder decisions the intrinsic and extrinsic locus of causality behind motivation
[11]. Player models, among many uses, can inform and shape [47]. The latter is facilitated by external or internal rewards,
the monetization strategy of a game, they can be directly pressures, and expectations, while the former is based on
applied as drivers of personalised content generation [17], [18], the intrinsic properties of the activity itself, namely how
and they can equip agents for believable user testing [19]. well it can support the three basic psychological needs of
Among the many aims of player modelling within game competence, autonomy, and relatedness. Videogames include
development, one can distinguish the two core tasks of clus- a fair amount of pressures and rewards which can promote
tering and prediction [3]. The former is based on unsupervised extrinsic motivation [48], and yet they are generally regarded
learning methods with the aim to cluster players within groups as good facilitators of intrinsic motivation [44]. Even when
of common behavioural pattern; approaches include k-means, short-term shifts in motivation are observed during gameplay,
self-organising maps [9], matrix factorisation and archetypal games support the necessary psychological needs for the
analysis [20], [21], and sequence mining [10], [22], [23]. The facilitation of intrinsic motivation on a higher level [8]. In
latter uses supervised learning approaches to predict patterns the context of videogames, Ryan et al. [43] describe the basic
of playing such as completion time [24] and churn [25]–[27]. psychological needs underlying intrinsic motivation as:
When predicting something about the player a distinction can 1) Competence or a sense of accomplishment and a desire
be made between the prediction of the player’s behaviour (i.e., for the mastery of an action, which manifests through
what would a player do?) [28] and the prediction of aspects the proximal and distal goals of the players. This need is
of the game experience (i.e., what would a player feel?) [3], generally tied to self-efficacy and a sense of meaningful
often using behavioural data but also other modalities of player progression. It is supported through the interactions the
input such as physiological data [29], [30], independently or players have to master in order to complete the game
in a multimodal fashion [31], [32]. but not completion in itself.
The literature involving theories of motivation as applied to 2) Autonomy or a sense of control and a desire for self-
games concerns primarily theoretical models that inform the determined action, which manifests through meaningful
design of experimental protocols instead of defining target out- choices, tactics, and strategic decisions the players can
puts for predictive modelling. Indicatively, Borbora et al. [33], take. It is supported through rule systems and different
game mechanics that both structure the play experience processes—e.g. anchoring-bias [56], [57], adaptation [58],
but allow for a high degree of freedom and meaningfully habituation [59], and other recency-effects [60]—that help
different outcomes. them evaluate their own experience internally.
3) Relatedness or a sense of belonging and a desire to con- During recent years there is growing evidence supporting
nect and interact with others, which manifests through the strength of preference learning for modelling emotions
interactions with other players and believable computer and user experiences both on a conceptual [15], [55] and a
agents. It is supported by multilayer interactions, believ- technical basis [61]–[64]. Conceptually, treating subjective-
able and rich non-player characters, narrative design, and defined ground truth data as ordinal variables brings the
even interactions with other players outside the games representation of data closer to the players’ underlying true
as well. attitudes [55]. This is based on the observation that one’s
4) Presence or the feeling of a mediated experience is affective state is always relative to a certain adaptation level
a main facilitator of both competence and autonomy, [15]; i.e., emotions have a shifting baseline based on previous
and can be viewed as having physical, emotional, and experiences. Even though traditionally psychological data is
narrative components [1], [43]. Indeed, the feeling of treated as nominal categories [65], by now there is ample
presence or the pursue of immersion can be a driving evidence suggesting higher validity and reliability when labels
force behind the motivation of gameplay [8], [49], [50]. of emotion or experience are treated as ordinal [61], [66], [67].
Based on the strong relationship between STD and pres- Indicatively, Martinez et al. [61] compared classification and
ence, both the Player Experience of Need Satisfaction PL across a number of modelling tasks found that ranking
Questionnaire [43] and UPEQ [14] use it to measure a yields more robust models than classification. A number of
level of involvement with the game which can facilitate other studies compared the processing of affective annotations
other positive psychological needs. as both ratings (e.g., Likert items) and rankings and found
It is important to note that the above factors are not that 1) first-order data processing (i.e. ranks) yields higher
contributing equally to the formulation of intrinsic motivation; reliability and inter-rater agreement and 2) second-order pro-
while competence or relatedness are regarded as the core cessing of the absolute rating values was also beneficial
catalysts, autonomy generally plays a supporting role in the fa- with regards to both reliability and validity [15], [62], [63].
cilitation of motivation. Nevertheless, in absence of autonomy, Recently Camilleri et al. [32] applied PL for obtaining general
motivation can only be considered introjected or compulsive models of players across games and Melhart et al. [64] showed
[51]. Within games the main drive of intrinsic motivation is that ranking Support Vector Machines (SVMs) are more robust
generally competence because of how the activity is structured, than support vector classifiers in predicting arousal across
while relatedness contributes to enhancing the experience [1]. dissimilar affective corpora.
In this paper we rely on SDT and recent advances on Given the above theoretical framework on the ordinal nature
measurement tools to quantify the four above-mentioned as- of experience and the large body of recent empirical evidence
pects of motivation. For that purpose we use UPEQ, a game- on the benefits of the ordinal modelling approach [15], in this
tailored questionnaire designed to measure the factors of SDT paper we view player motivation as an emotional construct
as affected by the gameplay experience. UPEQ was developed [1] with ordinal properties. As a result we compare player
by researchers at Massive Entertainment [14] specifically to feedback on relative grounds and use PL to model the ranking
predict gameplay outcomes relevant for industry designers and between the levels of reported motivation in players as mea-
stakeholders. Earlier work [14] has demonstrated that UPEQ sured by the factors of UPEQ. In that regard, we do not focus
is able to predict playtime, money spent on the game, and on the internal validity of the survey data—as this has been the
group playtime based on measured factors of SDT. Beyond its focus or earlier work [14]—and we instead consider the UPEQ
utility, UPEQ also addresses the limitations of prior domain- scores as the underlying ground truth we need to approximate.
specific SDT questionnaires, such as the Game Engagement After acquiring a general score for all the measured factors for
Questionnaire [52], BrainHex [53], and the Player Experience each participant, we return to analysing and modelling the data
of Need Satisfaction [43], while focusing on the adaptation of as ordinal values, thereby following a second-order modelling
the Basic Need Satisfaction Scale(s) [54] into a survey specific approach [15].
to videogame play. The result is a reliable and consistent
III. P REFERENCE L EARNING FOR M ODELLING
assessment tool with a strong theoretical foundation in SDT.
M OTIVATION
C. The Ordinal Nature of Motivation Preference Learning (PL) is a supervised machine learning
In the proposed methodology for this study we opt for pref- technique [16], in which an algorithm learns to infer the pref-
erence learning (PL) methods because of the strong connection erence relation between two variables. PL is a robust method,
between this ordinal machine learning paradigm and how which relies on relative associations instead of absolute values
player experience operates in games. In essence, PL models or class boundaries and is instead based on the pairwise
certain psychological processes by focusing on the differences transformation of the original dataset into a representation of
between occurrences instead of their absolute values [15], the differences between feature vectors in the query [68]. This
[55]. This approach falls much closer to the players’ cognitive transformation of the dataset reformulates the original problemA. Tom Clancy’s The Division
The data we analyse in this study is in-game behavioural
data (player metrics) and survey questionnaire responses from
players of Tom Clancy’s The Division (Ubisoft, 2016), here-
after The Division, collected between 2016 and 2018. The
Division is an online multiplayer action-role playing game
(Fig. 1), that combines a character progression system with
third-person, cover-based, tactical shooting combat mechanics.
The game is set in a post-apocalyptic New York which is hit
by a smallpox epidemic. Players, as government agents, have
to work together (and against each other) to scavenge and
Fig. 1. An example of the gameplay of Tom Clancy’s The Division (Ubisoft, investigate the city, which fell into chaos in the aftermath of
2016). Image taken from: store.steampowered.com/app/365590. No copyright the pandemic and the rise of organised crime activity.
infringement intended.
The core of the game is a progression system, in which
players gain new levels by participating in different in-game
activities including story-focused and optional missions to un-
in a way that a binary classifier can solve it. In our new dataset,
lock new abilities, and gain new equipment including weapons
we associate the direction of the preference relation with one
and armour. The strength of a player can be measured by their
of two classes. As an example, we observe the preference
level (up to 30) and the quality of their equipment is expressed
relation: xi xj ∈ X (xi is preferred over xj ) based
in Gear Score points. In the player versus environment (PvE)
on their associated output: yi > yj . Through the pairwise
sections of the game, players can group up and complete
transformation two new features are created: x01 = (xi − xj ),
missions together. The game also features a competitive player
associated with y10 = 1 and x02 = (xj − xi ), associated with
versus player (PvP) area—called the Dark Zone—which has
y20 = −1. This comparison between each pair of feature
its own progression system. In this special area players can still
vectors provides X 0 ⊆ X×(X−1) 2 new datapoints. X 0 is a
group up to complete missions for better equipment; however,
subset of all possible unique combinations because a clear
they can also turn on each other and become Rogue by
preference relation is not always inferable.
killing other players and taking their rewards for themselves.
This study uses ranking Support Vector Machines (SVM) After reaching the maximum level, players can participate in
[69] as they are implemented in the Preference Learning Incursions, which are particularly difficult missions for groups.
Toolbox1 [70], which is based on the LIBSVM library [71]. Ubisoft also released a number of expansions for the game in
We choose SVMs in this initial study as they have proven the form of downloadable content (DLC), which added new
their ability to yield robust models even with a limited amount areas, equipment, and both PvE and PvP content to the game.
of data and input features. SVMs were originally employed The game was not only well received (80/100 Metacritic
to solve classification tasks by maximizing the margins of a score on consoles2 ) but was also the best selling game of
hyperplane separating the datapoints projected into a higher Ubisoft at the time of its release3 . As the game integrates
dimensional feature space [72] but were later adopted to solve different systems from massively multiplayer online role play-
PL tasks as well [69]. We use both linear and non-linar SVMs ing games and multiplayer shooters and supports different
with radial basis function (RBF) kernels. Unlike linear SVMs, play styles and interaction modes (i.e. player-environment and
which aim for a linear separation between datapoints, RBF player-player), it provides a rich and complex game test-
SVMs emphasize the local proximity of datapoints, fitting the bed for research on motivation. As the popularity of online
maximum-margin hyperplane in a transformed feature space multiplayer games is on the rise and the second instalment
[72]. For tuning our algorithms, we rely on the C regulariza- of The Division set to release on March 15, 20194 , a study
tion term which controls the trade-off between maximizing the on how the gameplay of The Division shapes the motivational
margin and minimizing the classification error of the training factors of its players is both timely and relevant.
set, and—in case of RBF kernels—the γ hyperparameter,
which controls how each comparison between datapoints is B. Dataset and Preprocessing
weighted in the non-linear topology by limiting the variance The collected data consists of aggregated information on
of the similarity measure between points. the in-game activity of players over a long period of time and
their corresponding UPEQ survey scores. These two types of
IV. T HE G AME AND THE DATA data were collected independently, with the gameplay features
recorded between 2016 and 2018 and the survey data collected
In this section we first present the examined game in detail
2 https://www.metacritic.com/game/xbox-one/tom-clancys-the-division
and we then discuss the nature of the collected data and the
3 https://news.ubisoft.com/en-us/article/313497/
ways we pre-processed it for modelling motivation.
the-division-sets-new-sales-records/
4 https://news.ubisoft.com/en-us/article/342463/
1 http://plt.institutedigitalgames.com/ the-division-2-pc-features-specs-detailed/through a web interface separately in 2018. As such, the survey
data measures a general disposition of the players.
The dataset consists of one datapoint per player and, in
total, 443 players participated in the above-mentioned data
collection process. Approximately 51% of the subjects were
young adults, between 18 and 34 years old, while 9% were
underage (15-17), 34% between 35 and 54, and 6% above 55.
Country-wise, 23% of the respondents were from the United
Kingdom, 15% from Australia, 14% from Sweden, 9% from
Denmark, 5% from Finland, 5% from Norway, 1% from New
Zealand, with the remaining 28% not providing an answer.
The dataset is cleaned of datapoints with missing values,
corrupted entries, and outliers to prevent skewing any statisti-
cal analysis process. An extensive pruning was necessary due
to outliers distorting the distribution of general game metrics
(see Section IV-C) and due to noise generated by the data
logging service which inflated playtime. After the cleaning
process the dataset contains 298 players.
C. Extracted Features
To represent computationally the player’s behaviour within
the game, we ad-hoc design 30 high-level gameplay features.
While most of these are simple aggregated game metrics Fig. 2. Rank correlation matrix of the game metrics, play style classes
describing the time allocation and progression of the player, (motivation model input), and motivation factors (motivation model output).
4 of them are exclusive categories of distinct play styles
based on sequence-based profiling of the player’s in-game
and treat these scores as ordinal data through pairwise
activities [10]. Additionally, the dataset contains 4 Likert
comparisons across all players.
scores that represent the four motivation factors of each player
as measured by the UPEQ survey. The three types of data D. Ordinal Dataset Processing
considered in this study as detailed as follows: As the original dataset contains one datapoint per player,
1) Game Metrics: These features can be categorised as re- individual feature vectors used for the preference learning task
lating to general playtime (Days Played, Days in Groups, are independent.This means that during the preparation of the
Days in the Dark Zone, Sessions, Playtime, Group PL experiments, each datapoint is compared to every other
Playtime, Dark Zone Playtime, Playtime as Rogue); point during the pairwise transformation of the dataset. This
completion (Non-Daily Missions, Daily Missions, Side transformation applies a preference threshold (Pt ) parameter,
Missions, Days with Incursions, Incursions); progres- which controls the margin of significance under which two
sion (Gear-Score, Dark Zone Rank, Level, Early Level datapoints are considered equal. The purpose of a threshold Pt
30, Reached Level 30); early gameplay (Level < 30, is to counter the noise in the ground truth data which can skew
Early Playtime, Early Group Playtime, Early Dark Zone modelling results. Additionally to translating the relationship
Playtime, Early Playtime as Rogue); and DLC game- of datapoints into preference relations, this step also creates
play (Underground Playtime, Survival Playtime, Season- new datapoints for the machine learning task. The size of
Pass). the dataset is nearing a quadratic proportion to the original
2) Player Types: The 4 different player types are named dataset, with 64, 705 training and 775 testing points on average
Adventurer, Elite, PvE All-Rounder, and Social Dark depending on the ground truth and the optimal Pt parameter
Zone Player. These types have been derived through a (see Section VI-A). Furthermore, because each pairwise com-
traditional k-means clustering of highly aggregated data parison creates two new datapoints—describing the preference
and have been discontinued. relation in both directions—the transformation balances the
3) Motivation Factors: UPEQ scores the four factors of baseline of the classification task to 50% accuracy.
motivation in the form of averaged Likert-scale values.
While computing the mean of ordinal data can be prob- V. D ESCRIPTIVE S TATISTICAL A NALYSIS
lematic conceptually [15], [63], average survey scores In this section we present a descriptive statistical analysis
are a wide-spread method of using Likert-like data as which offers a first overview of the properties of the dataset.
they can still show certain tendencies within the scores In particular, we use Kendall’s τ to explore connections
(e.g., a higher score is assumed to correspond to an between individual input features (game metrics and play
overall more positive response). As mentioned in Section styles) and the four corresponding target outputs (motivation
II-C, we adopt a second-order modelling approach [15] factors measured by UPEQ). We choose τ over Spearman’sρ because of its lower gross error-sensitivity and asymptotic former describes a form absorption by the game world, the
variance [73]. In the analysis of this Section we use two latter often requires communication outside of the game.
statistical significance thresholds (α) of 0.005 and 0.001. While the above results showcase the descriptive capacity
As evidenced by Fig. 2, the data yields significant cor- of UPEQ to a good degree, the limited number of significant
relations between and within game metrics and play styles, effects obtained for the four reported factors of motivation
however, there is little to no linear relationship between highlight that the mapping between gameplay and reported
these features and motivation factors. The analysis finds that motivation is complex and far from linear. As more sophis-
between 88.30% (α = 0.001) and 91.07% (α = 0.005) ticated methods are needed to capture that mapping, in the
of the observed correlations are significant within the game remainder of this paper we present experiments using machine
metrics, which is not surprising given that most of the data is learning for deriving both linear and non-linear relationships
focused on the time players spend on different activities. As between gameplay features and motivation factors.
such, many of the more general features (e.g., Days Played)
should correlate with the frequency of specific activities (e.g., VI. M ODELS OF P LAYER M OTIVATION
Completed Daily Missions) and vice versa. Looking at the This section presents the results of the machine learned
correlations between game metrics and play styles, we observe models of player motivation based on different feature sets.
fewer significant connections, between 46.15% (α = 0.001) As the extracted play styles are more complex descriptors of
and 52.88% (α = 0.005). The Adventurer and the PvE All- the player behaviour than the aggregated game metrics, we
Rounder are the play style categories which have the most examine their capacity of predicting motivation independently
significant correlations with the game metrics. This result (see Section VI-C and VI-B) but also in fusion (Section
makes intuitive sense as the Adventurer class denotes novice VI-D). Before delving into the details of the obtained results
players with generally lower amounts of playtime and PvE All- we outline the validation and parameter tuning process we
rounder represents players with high completion rates, group followed for our PL models.
playtime, and participation in endgame content.
High correlations between game metrics and motivation A. Validation and Parameter Tuning
factors are not as frequent: only between 10.57% (α = 0.001) All PL models are validated with 10-fold cross validation.
and 17.3% (α = 0.005) of the observed correlations are sig- To prevent data leakage, the training and test folds are sep-
nificant. The motivation factor correlated with the most game arated before the normalisation and pairwise transformation
metrics is relatedness. As evidenced by Fig. 2, relatedness— of the data. We apply z-normalisation to both the training
a motivation factor related to connection and interaction— and the test set before the transformation. To preserve the
reveals higher correlations with features denoting both group independence of the test set, we assume that it is drawn from
play (e.g., Group Playtime) and PvP actions (e.g., Playtime the same distribution as the training set and apply the same
as Rogue). Autonomy is also correlated with a number of transformation to the corresponding test set as well.
gameplay features, but mostly with completion (e.g., Daily The optimal parameters of the RankSVMs are found
Missions) and early play (e.g., Playtime as Rogue) aspects of through exhaustive search within value bounds. In particular
gameplay. we search exhaustively the triplets of C, gamma and Pt values
Given that play styles are conceived as high level distinct that yield the highest 10-fold cross-validation accuracies. The
clusters of behavioural patterns, it is not surprising that the C regularisation term is searched within C ∈ {1, 2, 3, 4, 5},
number of correlations within the group are low (33% (α = the gamma RBF parameter in γ ∈ {0.1, 0.5, 0.75, 1, 2}, and
0.001) and 50% (α = 0.005)) and only separate Adventurer— the optimal preference threshold in Pt ∈ {0, 0.5, 1}. While
the novice—from the other classes. Correlations between play the best gamma parameter was found to be 0.5 over all
styles and motivation factors are between (6.25% (α = 0.001) experiments, C and Pt were more sensitive to the topology of
and 18.75% (α = 0.005). The Adventurer and the PvE All- the data; see Fig. 3. The C and Pt parameters selected were
Rounder are the classes that correlate with competence and C = 2, Pt = 1 for competence; C = 1, Pt = 0 for autonomy;
relatedness, suggesting that these factors are easier to predict C = 4, Pt = 1 for relatedness; C = 3, Pt = 0.5 for presence
linearly via play styles. for linear SVMs and C = 3, Pt = 0.5 for competence; C = 4,
Finally, there is a high number of correlations between Pt = 0 for autonomy; C = 4, Pt = 0 for relatedness; C = 2,
the motivation factors (66% for both α values). The only Pt = 0.5 for presence for RBF SVMs.
correlations which are not significant are between relatedness
and autonomy and between relatedness and presence. This B. Models Based on Play Styles
result supports quantitatively the theoretical assumption— In the first round of experiments we only use the four play
see Section II-B—that the catalyst of intrinsic motivation in styles as input of the SVM model to test their predictive
games is the competence of a player; both player autonomy capacity of motivation; Figure 4a shows the final results
and relatedness appear to have supporting roles as factors obtained. Despite the low dimensionality of the input feature
of motivation. As the UPEQ questionnaire measures mainly set, both linear and non-linear models are able to surpass the
spatial presence [14], it makes intuitive sense that there are is 50% baseline respectively with 3.7% and 3.57% accuracy on
no effect between presence and relatedness because while the average across all models. Although the best models surpass(a) Play styles
Fig. 3. Test accuracy sensitivity of models based on all available features to
two core RankSVM parameters: C and Pt . When RBF SVMs are used, γ is
0.5.
the 60% threshold, the obtained performance on average is not
impressive. We can safely conclude that the four play styles on
their own are baseline-level predictors of player motivation and
they do not suffice for building accurate models of motivation.
C. Models Based on Game Metrics (b) Game metrics
In the second round of PL experiments we train the models
only considering the 26 game metrics as model inputs—as
discussed in Section IV-C. As evidenced by the bar plots of
Figure 4b, game metrics are fairly successful in predicting the
reported motivation factors. In particular, linear SVM models
are successful with an average accuracy of 65.89% across
all models while the best models for individual factors are
performing at almost 80% accuracy on certain folds: 79.66%,
75.62%, 71.69%, and 79.68%, respectively, for competence,
autonomy, relatedness, and presence. Relatedness appears to
be the easiest factor to predict for the linear models, which is
not surprising given that relatedness correlated with the most
individual game metrics during the statistical analysis. On
the other hand, linear SVMs struggle with autonomy, which (c) All features
can be explained by the low amount of correlations between
Fig. 4. Average accuracy of the best linear and RBF SVMs. The dotted line
autonomy and the other features found during the descriptive shows the baseline. The error bars represent 95% confidence intervals.
statistical analysis of the data.
Non-linear kernels further improve the model’s performance
to a 75.62% accuracy on average across all models. The D. Models Based on All Features
best individual models vastly outperform the corresponding
linear models reaching almost 90% accuracy (competence: We assume that high level play style profiles can enhance
86.73%; autonomy: 89.31%; relatedness: 89.95%; and pres- the predictive capacity of game metrics by adding domain-
ence: 87.60%). Compared to linear models, RBF SVMs appear specific information. As it is evidenced by Figure 4c, including
to be more robust across any motivation factor as they manage all available features for the PL task, improves the accuracy of
to improve greatly even the poor-performing linear models the non-linear models beyond the capabilities of models based
(i.e., autonomy. Unlike the poor performances obtained with on game metrics alone. On the one hand, the linear models
the models based solely on play styles, models based on game are only reaching 65.92% on average across all tests (79.66%,
metrics are very accurate and robust across all four factors. 70.94%, 71.79%, and 76.52%, respectively, for competence,autonomy, relatedness and presence) which is comparable to competence—there are observable patterns within the distribu-
the performance obtained by the models based on game met- tion of playtime between high and low tier players. While low
rics. On the other hand, models using a non-linear RBF kernel tier players are spending their time in groups and focusing on
reach acurracies of 82.36% on average and achieve accuracy PvE content (i.e., spending little time in the Dark Zone; not
values above 93% in their highest performing folds: 93.01%, playing as a rogue), the high tier players are focusing more on
94.35%, 95.02% and 96.83%, respectively, for competence, PvP action even during the early game. These players are also
autonomy, relatedness and presence. spending their time more diversely on average, experiencing
Interestingly, despite relatedness being one of the easier fac- more options that the game has to offer, and thus having
tors to predict for linear SVMs models, the RBF kernel SVMs better completion rates. Daily Missions, which are recurring
yield relatively lower accuracies on average. Nevertheless, but optional activities for extra in-game equipment, also appear
similarly to non-linear models based on game metrics, RBF to be much more prominent in the top autonomy players.
SVMs trained on all available features show a comparable The top and bottom ranked players based on relatedness
predictive capacity across the four different motivation factors. is shown on Figure 5c. The order based on relatedness is
Among the four, autonomy seems to be substantially more mainly affected by player-player interaction, which manifests
difficult to predict when using linear models. as either group play (e.g. Group Playtime, Days in Groups,
Incursion Days, etc.) or PvP action (Rogue Playtime), while
E. Visualising the Motivation Models other features remain comparable between the top and bottom
This section demonstrates the applicability of the models for tiers. Interestingly, highly ranked players are not possessing
game design, by re-examining the features describing players, a season pass for DLCs (save for one), while almost all the
ordered by the predictions of each model. To create a rank bottom 10 players do. Because these players have no logged
order between players, the models are refit to the whole dataset playtime with the DLC content, however, the effect can be
with the best parameters. We then create the global order attributed to a latent variable and not the DLC content.
across players’ motivation by comparing the output of the Finally, the order of players based on predicted presence is
best obtained SVM models through a round robin tournament depicted in Figure 5d. The order of the feature vectors appear
among all players in the dataset. We run the tournament for to be uniformly distributed when ranked by presence as there
each of the four motivation factors separately. Finally we order are not many observable features that differ between the top
the players based on the outcome of these tournaments and— and the bottom players. Of note is that higher tier players
for the sake of readability—we visualise the feature sets for show more involvement in the social activities of the game,
the top 10 and the bottom 10 players in the order for each of mostly in terms of time spent in groups. Moreover, lower
the four motivation factors (Fig. 5). tier players are more focused on progression and expansion
As evidenced by Figure 5a, the global order of predicted content. Since the UPEQ questionnaire measures physical
competence suggests that most of the top players ordered presence [14], it is not surprising that activities which require
by this factor are playing more in general, team up early extra-game coordination such as interaction with other players
(and generally spend more time in groups), and have higher can decrease this factor.
completion and progression rates compared to the players that Across all motivation factors, the feature varying the most
rank at lowest competence. This is particularly evident in PvP, between the top and the bottom ranked players is the com-
end-game PvE, and daily missions. It is not surprising that pletion of Daily Missions, and PvP activity (either spending
the players topping the competence rank are mostly associated time in the Dark Zone or going Rogue). While early gameplay
with the PvE All-Rounder play style, whereas the players at alone is rarely a strong predictor, game metrics in general seem
the bottom tier have been classified mostly as novices (i.e., to be appropriate predictors and useful for both analytics and
Adventurer). An interesting observation is that players rushing player modelling efforts. It is also largely evidenced that even
through the content (Early Level 30 coupled with lower general in cases of no obvious linear relationship between individual
playtime) report lower competence than players building their features and motivation factors, non-linear PL techniques can
characters more slowly. The connection between features and provide efficient methods for predicting motivation and offer
predicted ranking, however, is not always observable in a linear an insightful qualitative tool for game design.
fashion. As players rate their experience without an external
frame of reference, higher values in certain features do not VII. D ISCUSSION
necessarily mean higher ranking as well. As an example, the
model selects to place three Adventurer players at the top of In this study we applied preference learning via SVMs to
the competence rank. While these users have no end-game model psychological constructs that make up self determi-
experience (no completed Incursions), they spent more time nation theory and demonstrated the supreme efficiency and
in groups and playing in the Dark Zone (i.e., the game’s PvP robustness of our models to predict motivation factors of play-
environment), than novice players in the bottom rank. ers based solely on their in-game data. We showed that these
The players ranked by predicted autonomy are illustrated models can be used directly in a games user research context
in Figure 5b. While playtime in general does not have such to predict factors of players’ motivation based on aggregated
an effect on the order of autonomy—as it has on the order of data, collected through multiple play sessions, which can beThe Division and defined motivation quantitatively using the
UPEQ questionnaire. We converted the Likert scores to ordinal
values using a second-order data processing approach [15]
that yielded data corpora of over 60, 000 samples, representing
the differences between players’ UPEQ Likert scores in pairs.
We predicted each player’s levels of competence, autonomy,
relatedness and presence with more than 80% accuracy on
average, whereas the best models achieve accuracies of at least
93%.
(a) Competence The presented research offers insights for game industry
professionals and stakeholders, who aim to leverage and en-
hance the positive psychological effects of gameplay but also
for researchers in affective computing and game user research.
The paper offers a novel approach that utilises a machine (pref-
erence) learning approach to model psychological constructs
such as motivation by treating the subjectively defined notions
as an ordinal phenomenon. To the best of our knowledge,
this is the first study, which attempted to quantitatively model
and predict constructs of STD based on player behaviour in a
commercial-standard game and with that level of success.
(b) Autonomy A core limitation of this work is posed by the processing of
the motivation factors (i.e., the ground truth). In particular, we
opted for a simple experimental design and took the aggregated
UPEQ scores at face value; aggregating Likert scale scores in
this way, however, poses a number of methodological issues.
The most important of these limitations is the processing of
ordinal values as interval data [63]. As the distance between
ordinal points is arbitrary, their mean value is not necessarily
meaningful and introduces subjective reporting bias to the
measurements [55]. To address this issue, future research
should focus on new ways of transforming Likert-like data
(c) Relatedness to preference relations. One possible approach would be to
observe how individual player’s baseline shifts through the
survey and scoring the 4 observed factors depending on
significant changes in the responses.
Another limitation is posed by the nature of the dataset. As
the dataset contains aggregated gameplay data, we cannot ob-
serve clearly how each player’s motivation changes over time
and thus the relationship between the player’s “lifetime” and
a one-time post-experience survey is rather broad. Although
this type of data is easier to handle in a qualitative analysis,
machine learning models are able to handle larger datasets
with a higher dimensional feature space. Future research could
focus on collecting a new dataset, recording multiple sessions
and questionnaires per participant, and creating models which
can predict temporal changes in the player’s motivation.
It is important to note that the ordinal data processing
method we propose in this paper yields large datasets which
are generated through the pairwise comparison of all players.
(d) Presence
As such, the method is feasible in current game development
Fig. 5. Feature sets of the top and bottom 10 players ordered based on
predicted motivation factors. Each row represents a player and each column
settings since it only requires a small sample of player expe-
is a different input feature. The colour shows the normalised distribution. rience annotations—such as those usually available in qual-
ity assurance departments of game studios. Although SVMs
already showcased high levels of efficiency in predicting
the subject of further qualitative analysis. In particular, we col- motivation on sets of several thousand datapoints, it can be
lected and processed the data of 298 players of Tom Clancy’s argued that they might not be as robust when faced withmuch larger datasets of that type. In such instances alternative [14] A. Azadvar and A. Canossa, “Upeq: ubisoft perceived experience
methods derived from deep preference learning [74], [75] are questionnaire: a self-determination evaluation tool for video games,”
in Proceedings of the International Conference on the Foundations of
directly applicable to the modelling task. Digital Games (FDG). ACM, 2018.
[15] G. N. Yannakakis, R. Cowie, and C. Busso, “The ordinal nature of
VIII. C ONCLUSION emotions: An emerging approach,” IEEE Transactions on Affective
Computing, 2018.
[16] J. Fürnkranz and E. Hüllermeier, “Preference learning,” in Encyclopedia
In this paper we attempted to infer the computational of Machine Learning. Springer, 2011, pp. 789–795.
mapping between gameplay data and aspects of player moti- [17] G. N. Yannakakis and J. Togelius, “Experience-driven procedural content
vation. In particular we used the Ubisoft Perceived Experience generation,” IEEE Transactions on Affective Computing, vol. 2, no. 3,
pp. 147–161, 2011.
Questionnaire to survey about the competence, autonomy, re- [18] N. Shaker, G. N. Yannakakis, and J. Togelius, “Towards automatic
latedness and presence of almost 300 players of Tom Clancy’s personalized content generation for platform games.” in Proceedings
The Division. We also logged these players’ in-game data of the Conference on Artificial Intelligence and Interactive Digital
Entertainment (AIIDE), 2010.
and we examined the degree to which such data can be
[19] C. Holmgård, A. Liapis, J. Togelius, and G. N. Yannakakis, “Evolving
a powerful predictor of players’ survey responses. For that personas for player decision modeling,” in Proceedings of the Confer-
purpose, we converted survey scores to ordinal values and ap- ence on Computational Intelligence and Games (CIG). IEEE, 2014,
plied preference learning to create efficient and robust support pp. 1–8.
[20] C. Bauckhage, A. Drachen, and R. Sifa, “Clustering game behavior
vector machine models of the four factors of motivation. The data,” IEEE Transactions on Computational Intelligence and AI in
ordinal machine learning approach we took proved extremely Games, vol. 7, no. 3, pp. 266–278, 2015.
successful in predicting such complex psychological constructs [21] A. Drachen, R. Sifa, C. Bauckhage, and C. Thurau, “Guns, swords and
data: Clustering of player behavior in computer games in the wild,”
by eliminating the reporting bias of the questionnaires. Our in Proceedings of the Conference on Computational Intelligence and
core findings suggest that the mapping between high-level Games (CIG). IEEE, 2012, pp. 163–170.
gameplay metrics and survey-based annotations of complex [22] H. P. Martı́nez and G. N. Yannakakis, “Mining multimodal sequential
patterns: a case study on affect detection,” in Proceedings of the
emotional and cognitive states is not only possible to infer International Conference on Multimodal Interfaces. ACM, 2011, pp.
but the inferred models have predictive capacities that reach 3–10.
certainty levels. [23] G. Wallner, “Sequential analysis of player behavior,” in Proceedings of
the Annual Symposium on Computer-Human Interaction in Play. ACM,
2015, pp. 349–358.
R EFERENCES [24] T. Mahlmann, A. Drachen, J. Togelius, A. Canossa, and G. N. Yan-
nakakis, “Predicting player behavior in tomb raider: Underworld,”
[1] S. Rigby and R. M. Ryan, Glued to games: How video games draw us in Proceedings of the Symposium on Computational Intelligence and
in and hold us spellbound. Praeger, 2011. Games (CIG). IEEE, 2010, pp. 178–185.
[2] G. N. Yannakakis and A. Paiva, “Emotion in games,” Handbook on [25] J. Runge, P. Gao, F. Garcin, and B. Faltings, “Churn prediction for high-
affective computing, pp. 459–471, 2014. value players in casual social games,” in Proceedings of the Conference
[3] G. N. Yannakakis and J. Togelius, Artificial Intelligence and Games. on Computational Intelligence and Games (CIG). IEEE, 2014, pp. 1–8.
Springer, 2018. [26] Á. Periáñez, A. Saas, A. Guitart, and C. Magne, “Churn prediction
[4] J. M. Weinhardt and J. B. Vancouver, “Computational models and in mobile social games: towards a complete assessment using survival
organizational psychology: Opportunities abound,” Organizational Psy- ensembles,” in Proceedings of the International Conference on Data
chology Review, vol. 2, no. 4, pp. 267–292, 2012. Science and Advanced Analytics (DSAA). IEEE, 2016, pp. 564–573.
[5] A. Canossa, J. B. Martinez, and J. Togelius, “Give me a reason to dig [27] M. Viljanen, A. Airola, J. Heikkonen, and T. Pahikkala, “Playtime
minecraft and psychology of motivation.” Proceedings of the Conference measurement with survival analysis,” IEEE Transactions on Games,
on Computational Intelligence and Games (CIG), vol. 2013, pp. 1–8, vol. 10, no. 2, pp. 128–138, 2018.
2013. [28] S. C. Bakkes, P. H. Spronck, and G. van Lankveld, “Player behavioural
[6] J. Hamari and J. Tuunanen, “Player types: A meta-synthesis,” Transac- modelling for video games,” Entertainment Computing, vol. 3, no. 3,
tions of the Digital Games Research Association, vol. 1, no. 2, 2014. pp. 71–79, 2012.
[7] A. Canossa, J. B. Badler, S. Tignor, and R. C. Colvin, “In your face(t) [29] H. P. Martı́nez, M. Garbarino, and G. N. Yannakakis, “Generic phys-
impact of personality and context on gameplay behavior.” in Proceedings iological features as predictors of player experience,” in Proceedings
of the International Conference on the Foundations of Digital Games of the International Conference on Affective Computing and Intelligent
(FDG), 2015. Interaction (ACII). Springer, 2011, pp. 267–276.
[8] D. Melhart, “Towards a comprehensive model of mediating frustration [30] V. Georges, F. Courtemanche, M. Fredette, P.-M. Lger, and S. Sncal,
in videogames,” Game Studies, vol. 18, no. 1, 2018. “Developing personas based on physiological measures,” in Proceedings
[9] A. Drachen, A. Canossa, and G. N. Yannakakis, “Player modeling using of the International Conference on Physiological Computing Sysytems,
self-organization in tomb raider: Underworld,” in Proceedings of the 2018, p. 131136.
Symposium on Computational Intelligence and Games (CIG). IEEE, [31] N. Shaker, S. Asteriadis, G. N. Yannakakis, and K. Karpouzis, “Fusing
2009, pp. 1–8. visual and behavioral cues for modeling user experience in games,” IEEE
[10] S. Makarovych, A. Canossa, J. Togelius, and A. Drachen, “Like a dna Transactions on Cybernetics, vol. 43, no. 6, pp. 1519–1531, 2013.
string: Sequence-based player profiling in tom clancys the division,” [32] E. Camilleri, G. N. Yannakakis, and A. Liapis, “Towards general models
in Proceedings of the Artificial Intelligence and Interactive Digital of player affect,” in Proceedings of the International Conference on
Entertainment Conference. York, 2018. Affective Computing and Intelligent Interaction (ACII). IEEE, 2017,
[11] M. S. El-Nasr, A. Drachen, and A. Canossa, Game analytics. Springer, pp. 333–339.
2016. [33] Z. Borbora, J. Srivastava, K.-W. Hsu, and D. Williams, “Churn prediction
[12] G. N. Yannakakis, P. Spronck, D. Loiacono, and E. André, “Player in mmorpgs using player motivation theories and an ensemble approach,”
modeling,” in Dagstuhl Follow-Ups, vol. 6. Schloss Dagstuhl-Leibniz- in Proceedings of the International Conference on Privacy, Security, Risk
Zentrum fuer Informatik, 2013. and Trust (PASSAT) & Proceedings of the International Conference on
[13] R. M. Ryan and E. L. Deci, “Self-determination theory and the fa- Social Computing (SocialCom). IEEE, 2011, pp. 157–164.
cilitation of intrinsic motivation, social development, and well-being.” [34] N. Yee, “Motivations for play in online games,” CyberPsychology &
American Psychologist, vol. 55, no. 1, p. 68, 2000. Behavior, vol. 9, no. 6, pp. 772–775, 2006.[35] K. J. Shim, K.-W. Hsu, and J. Srivastava, “An exploratory study of player [58] H. Helson, Adaptation-level theory: an experimental and systematic
performance, motivation, and enjoyment in massively multiplayer online approach to behavior. Harper and Row: New York, 1964.
role-playing games,” in Proceedings of the International Conference [59] R. L. Solomon and J. D. Corbit, “An opponent-process theory of
on Privacy, Security, Risk and Trust (PASSAT) & Proceedings of the motivation: I. temporal dynamics of affect.” Psychological Review,
International Conference on Social Computing (SocialCom). IEEE, vol. 81, no. 2, p. 119, 1974.
2011, pp. 135–140. [60] S. Erk, M. Kiefer, J. Grothe, A. P. Wunderlich, M. Spitzer, and H. Walter,
[36] M. V. Birk, D. Toker, R. L. Mandryk, and C. Conati, “Modeling “Emotional context modulates subsequent memory effect,” Neuroimage,
motivation in a social network game using player-centric traits and vol. 18, no. 2, pp. 439–447, 2003.
personality traits,” in Proceedings of the International Conference on [61] H. Martinez, G. Yannakakis, and J. Hallam, “Dont classify ratings of
User Modeling, Adaptation, and Personalization. Springer, 2015, pp. affect; rank them!” IEEE transactions on Affective Computing, no. 1,
18–30. pp. 1–1, 2014.
[37] S. Reiss and S. M. Havercamp, “Toward a comprehensive assessment [62] G. N. Yannakakis and H. P. Martinez, “Grounding truth via ordinal
of fundamental motivation: Factor structure of the reiss profiles.” Psy- annotation,” in Proceedings of the International Conference on Affective
chological Assessment, vol. 10, no. 2, p. 97, 1998. Computing and Intelligent Interaction (ACII). IEEE, 2015, pp. 574–
[38] E. Deci and R. M. Ryan, Intrinsic motivation and self-determination in 580.
human behavior. Springer Science & Business Media, 1985. [63] G. N. Yannakakis and H. P. Martı́nez, “Ratings are overrated!” Frontiers
[39] E. L. Deci, R. J. Vallerand, L. G. Pelletier, and R. M. Ryan, “Moti- in ICT, vol. 2, p. 13, 2015.
vation and education: The self-determination perspective,” Educational [64] D. Melhart, K. Sfikas, G. Giannakakis, G. N. Yannakakis, and A. Liapis,
Psychologist, vol. 26, no. 3-4, pp. 325–346, 1991. “A motivational model of video game engagement.” Proceedings of
[40] M. Gagné and E. L. Deci, “Self-determination theory and work motiva- Machine Learning Research, 2018 IJCAI workshop on AI and Affective
tion,” Journal of Organizational Behavior, vol. 26, no. 4, pp. 331–362, Computing, vol. 86, pp. 26–33, in print.
2005. [65] A. Kapoor, “Machine learning for affective computing: Challenges and
[41] M. Joussemet, R. Landry, and R. Koestner, “A self-determination theory opportunities,” The Oxford Handbook of Affective Computing, p. 406,
perspective on parenting.” Canadian Psychology/Psychologie Canadi- 2015.
enne, vol. 49, no. 3, p. 194, 2008. [66] M. Junge and R. Reisenzein, “Indirect scaling methods for testing
[42] M. S. Hagger and N. L. Chatzisarantis, Intrinsic motivation and self- quantitative emotion theories,” Cognition & Emotion, vol. 27, no. 7,
determination in exercise and sport. Human Kinetics, 2007. pp. 1247–1275, 2013.
[67] ——, “Metric scales for emotion measurement,” Psychological Test and
[43] R. M. Ryan, C. S. Rigby, and A. Przybylski, “The motivational pull
Assessment Modeling, vol. 58, no. 3, p. 497, 2016.
of video games: A self-determination theory approach,” Motivation and
[68] J. Fürnkranz and E. Hüllermeier, “Pairwise preference learning and
Emotion, vol. 30, no. 4, pp. 344–360, 2006.
ranking,” in Proceedings of the European Conference on Machine
[44] A. K. Przybylski, C. S. Rigby, and R. M. Ryan, “A motivational model
Learning. Springer, 2003, pp. 145–156.
of video game engagement.” Review of General Psychology, vol. 14,
[69] T. Joachims, “Optimizing search engines using clickthrough data,” in
no. 2, p. 154, 2010.
Proceedings of the SIGKDD International Conference on Knowledge
[45] C. L. Hull, “Principles of behavior: An introduction to behavior theory.” Discovery and Data Mining. ACM, 2002, pp. 133–142.
Journal of Philosophy, vol. 40, pp. 558–559, 1943. [70] V. E. Farrugia, H. P. Martı́nez, and G. N. Yannakakis, “The preference
[46] B. F. Skinner, Science and human behavior. Simon and Schuster, 1953, learning toolbox,” arXiv preprint arXiv:1506.01709, 2015.
no. 92904. [71] C.-C. Chang and C.-J. Lin, “Libsvm: a library for support vector
[47] R. M. Ryan and J. P. Connell, “Perceived locus of causality and machines,” Transactions on Intelligent Systems and Technology (TIST),
internalization: Examining reasons for acting in two domains.” Journal vol. 2, no. 3, p. 27, 2011.
of Personality and Social Psychology, vol. 57, no. 5, p. 749, 1989. [72] V. Vapnik, “Chapter 5 constructing learning algorithms,” The Nature of
[48] E. L. Deci, R. Koestner, and R. M. Ryan, “A meta-analytic review Statistical Learning Theory, pp. 119–157, 1995.
of experiments examining the effects of extrinsic rewards on intrinsic [73] C. Croux and C. Dehon, “Influence functions of the spearman and
motivation.” Psychological Bulletin, vol. 125, no. 6, p. 627, 1999. kendall correlation measures,” Statistical Methods & Applications,
[49] M. Lombard and T. Ditton, “At the heart of it all: The concept of vol. 19, no. 4, pp. 497–515, 2010.
presence,” Journal of Computer-Mediated Communication, vol. 3, no. 2, [74] H. P. Martı́nez and G. N. Yannakakis, “Deep multimodal fusion:
1997. Combining discrete events and continuous signals,” in Proceedings of
[50] G. Calleja, In-game: From immersion to incorporation. MIT Press, the International Conference on Multimodal Interaction. ACM, 2014,
2011. pp. 34–41.
[51] R. M. Ryan and E. L. Deci, “Intrinsic and extrinsic motivations: Classic [75] H. P. Martinez, Y. Bengio, and G. N. Yannakakis, “Learning deep
definitions and new directions,” Contemporary Educational Psychology, physiological models of affect,” Computational Intelligence Magazine,
vol. 25, no. 1, pp. 54–67, 2000. vol. 8, no. 2, pp. 20–33, 2013.
[52] J. H. Brockmyer, C. M. Fox, K. A. Curtiss, E. McBroom, K. M.
Burkhart, and J. N. Pidruzny, “The development of the game engage-
ment questionnaire: A measure of engagement in video game-playing,”
Journal of Experimental Social Psychology, vol. 45, no. 4, pp. 624–634,
2009.
[53] L. E. Nacke, C. Bateman, and R. L. Mandryk, “Brainhex: A neurobio-
logical gamer typology survey,” Entertainment Computing, vol. 5, no. 1,
pp. 55–62, 2014.
[54] B. Chen, M. Vansteenkiste, W. Beyers, L. Boone, E. L. Deci, J. Van der
Kaap-Deeder, B. Duriez, W. Lens, L. Matos, A. Mouratidis et al., “Basic
psychological need satisfaction, need frustration, and need strength
across four cultures,” Motivation and Emotion, vol. 39, no. 2, pp. 216–
236, 2015.
[55] G. N. Yannakakis, R. Cowie, and C. Busso, “The ordinal nature of
emotions,” in Proceedings of the International Conference on Affective
Computing and Intelligent Interaction (ACII). IEEE, 2017, pp. 248–
255.
[56] A. R. Damasio, Descartes error: Emotion, rationality and the human
brain. New York: Putnam, 1994.
[57] B. Seymour and S. M. McClure, “Anchors, scales and the relative coding
of value in the brain,” Current Opinion in Neurobiology, vol. 18, no. 2,
pp. 173–178, 2008.You can also read
