Towards Automated Decision Making in Dating Apps Through Pupillary
Responses
Jan Ehlers
1 a
, Sebastian Laverde Alfonso
2
and Arup Mazumder
3
1
Department of Computer Science, Bauhaus-Universit
¨
at Weimar, Schwanseestr. 143, Weimar, Germany
2
Unstructured.io, 7580 Horseshoe Bar Rd., Loomis, CA, U.S.A.
3
University of Rhode Island, 45 Upper College Rd, Kingston, RI, U.S.A.
Keywords:
Cognitive Pupillometry, Affective Processing, Arousal, Decision Making, Automation.
Abstract:
Decision making is a multi-stage process that involves a series of rational evaluations. Recently, bodily arousal
has been identified as a factor that mediates individual decisions, particularly during partner selection. The
current study investigates pupil size changes in response to facial images of the opposite sex from controlled
eye-tracking data (Experiment 1) and by reading out signals from front-facing smartphone cameras in noisy
environments (Experiment 2). The aim is to enable automated decision-making in dating apps using arousal-
based information. The rating results showed a tendency towards moderate evaluations when coping with
facial attractiveness, while pupil diameter did not clearly discriminate between all four rating categories. How-
ever, a ROCKET model was trained on the pupil data from Experiment 1 with a prediction accuracy of 77%
for binary classification of clearly preferred and non-preferred images. Ambiguous responses will therefore
continue to pose a problem for cognition-aware systems. Capturing pupil diameter from mobile phone cam-
eras resulted in a high proportion of inadequate recordings, probably due to a lack of experimental control.
However, an overly systematic approach should run contrary to the intended scenario of lifelike mobile dating
app usage.
1 INTRODUCTION
Over the past decades, online dating has become in-
creasingly popular with more than 200 million ac-
tive users by the end of 2019 (Castro and Barrada,
2020). In recent years, however, regular use has
shifted to mobile dating apps that feature real-time
location based services and typically require users to
decide for or against a candidate’s photos by swiping
left (rejection) or right (acceptation) across the screen
(Sawyer et al., 2018; Wu and Trottier, 2022). This
image-based approach is based on the view that phys-
ical appearance exerts a decisive influence on part-
ner selection (Berscheid and Walster, 1974). Tally-
ing with this, evolutionary psychology suggests that
phenotypic features of the face (more than body char-
acteristics) are particularly well suited to predicting
overall attraction (Fink and Penton-Voak, 2002). And
although the relevance of facial cues appears to differ
between the sexes (Buss, 2016), both men and women
will pay more attention to attractive faces compared to
a
https://orcid.org/0000-0002-4475-2349
unattractive ones (Dai et al., 2010; Hahn and Perrett,
2014). Thus, today’s dating apps meet our needs dur-
ing partner search with the result that they are used
frequently and in almost any environment. The bi-
nary selection mechanism (reject vs accept) features a
straightforward design, but manual input can become
tiring during long-term usage and even prone to in-
put errors while on the go or in case of one-handed
operation.
Decision making - in dating apps but also in gen-
eral - constitutes a multistage process that involves a
series of (rational) judgments. However, when the
stakes are high and solutions have long-term conse-
quences, individual choices are strongly influenced
by visceral factors (due to anxiety) or drive states
(e.g. sexual desire) (Damasio, 1996). Especially
sexual arousal has recently gained recognition as a
situational factor mediating individual mate choice
(Skakoon-Sparling et al., 2016). The current work
explores arousal-based changes in response to facial
images of the opposite sex. To quantify arousal we
aim for a peripheral measure that can be administered
remotely and is applicable in different settings. As
522
Ehlers, J., Alfonso, S. and Mazumder, A.
Towards Automated Decision Making in Dating Apps Through Pupillary Responses.
DOI: 10.5220/0012471100003660
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 19th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2024) - Volume 1: GRAPP, HUCAPP
and IVAPP, pages 522-529
ISBN: 978-989-758-679-8; ISSN: 2184-4321
Proceedings Copyright © 2024 by SCITEPRESS Science and Technology Publications, Lda.
indicated above, visual cognition is ubiquitous during
partner selection, suggesting eye-based data to con-
stitute a promising channel for incorporating bodily
activation.
Pupil diameter primarily reflects the nervous sys-
tem response to ambient luminance (’pupillary light
reflex’) (Kardon, 1995). However, in controlled light-
ing environments, pupil size changes are supposed to
reflect processes underlying cognitive functions such
as attention (Einh
¨
auser et al., 2008) or working mem-
ory (Kucewicz et al., 2018; Grimmer et al., 2021).
Affective pupillometry relies on innervations from the
autonomic nervous system and indicates that pupil di-
ameter increases with general bodily arousal (Ehlers
et al., 2016; Ehlers et al., 2018). Liao et al. (Liao
et al., 2021) as well as Pronk et al. (Pronk et al., 2021)
report an increase in pupil diameter when viewing at-
tractively rated faces. Laeng and Falkenberg (Laeng
and Falkenberg, 2007) observe strong dilations in fe-
male participants to pictures of their sexual partners.
Tallying with this, Rieger and Savin-Williams (Rieger
and Savin-Williams, 2012) identify pupil diameter as
an indicator of sexual orientation, with heterosexual
men showing greater dilations to images of women
than vice versa. Bernick et al. (Bernick et al., 1971)
report pupil diameter to increase with sexual arousal
and conclude that pupil size changes can be applied
to distinguish between sexual arousal and more gen-
eral states of activation. However, studies on affective
pupillometry tend to apply cumulative results over a
period of time after stimulus-onset and to ignore time-
dependent changes in pupil diameter. And although
research has focused almost exclusively on laboratory
experiments, results are often confounded with stim-
ulus luminance or contrast (Liao et al., 2021).
The present study builds on this research. The
overall aim is to enable arousal-based decision mak-
ing via pupil responses in order to automate the use
of dating apps. Therefore, we investigate pupil size
changes in response to facial images of the oppo-
site sex in a fully controlled laboratory experiment
(Experiment 1). A state-of-the-art machine learning
model is applied to predict individual choices by clas-
sifying pupil responses. Experiment 2 is a field study
that adopts the previous design but reads out pupil di-
ameter from front-facing smartphone cameras. The
results will show the extent to which the laboratory
findings can be transferred to real-life conditions.
2 EXPERIMENT 1
Experiment 1 investigates changes in pupil diameter
when viewing facial stimuli of the opposite sex. En-
vironmental factors are kept constant in order to de-
termine influences from cognitive and affective pro-
cessing while coping with facial attractiveness.
2.1 Methods
The following subsections provide information on the
experimental design and specify details on data pro-
cessing and sample characteristics.
2.1.1 Design and Procedure
Experiment 1 was carried out in a laboratory room
featuring a constant illumination of 300 lux. Partici-
pants received an introduction to the research subject
and were then seated in an upright position in front
of a computer monitor. A stationary Eye-tracker was
attached to the lower edge of the screen at a distance
of 90cm.
We applied a within-subject design by providing
a set of 60 high-quality AI-generated images that was
randomly taken from a pool of 400 stimuli (Fig. 1A)
(Karras et al., 2019). Each image showed a human
face that looked directly into the camera with no other
object in the foreground or background. Our volun-
teers identified as heterosexual and were asked to pas-
sively view 30 images of the opposite sex. Each im-
age was presented for five seconds against a light grey
background and preceded by a one-second baseline
recording. Baseline acquisition involved the presen-
tation of a pixel scrambled version of the subsequent
image to control for differences in luminance and con-
trast (Fig.1B).
Figure 1: A: AI-generated images. B: Trial procedure dur-
ing the image viewing task.
Subsequent to each image presentation, volunteers
were asked to evaluate the corresponding face on a
Towards Automated Decision Making in Dating Apps Through Pupillary Responses
523
four-point scale ranging from (1) “Do not like at all”,
(2) “Rather do not like” to (3) “Rather like” and (4)
“Totally like”. After the individual assessment par-
ticipants proceeded to the next trial via button press.
Pupil diameter was recorded throughout the entire ex-
periment (approx. five minutes).
2.1.2 Data Processing
Eye-tracking data was obtained binocular with a sam-
pling rate of 60 Hz. According to (Mantiuk et al.,
2012), data from the left eye were taken to deter-
mine changes in pupil diameter. Following (Leys
et al., 2013), we applied the median absolute devia-
tion (MAD) to correct for outliers. Recordings mainly
included blink artefacts, i.e. segments of missing data
as a result of eyelid closure. Exemplary raw data tri-
als are depicted in Figure 2. Outliers were removed
from the data set and replaced via linear interpolation.
For purposes of presentation, eye-tracking data were
smoothed using a moving window of 25 data points.
Figure 2: Exemplary raw data trials featuring pupil size
changes during the image viewing task. Vertical lines in-
dicate blink artefacts (segments of missing data).
2.1.3 Apparatus
Pupil size changes were measured with the Tobii Pro
Nano Eye-tracker, Tobii Pro Lab Software (v. 1.123)
was applied for data recording and storage. The Eye-
tracker was attached to a 22-inch Dell P2213 monitor
with a resolution of 1680 x 1050 pixels. Statistical
analyses were carried out using the open-source soft-
ware JASP (v.0.12.2) (Love et al., 2019).
2.1.4 Particiants
16 volunteers (five female, M: 28 years (SD: 3)), all
of them students of computer science at Bauhaus-
Universit
¨
at, participated in Experiment 1. Informa-
tion on regular medication was not collected, how-
ever, volunteers reported no history of head injury and
no neurological or psychiatric disorders. Written in-
formed consent was obtained prior to the experiment.
All measurements were performed in accordance with
the Declaration of Helsinki and approved by a local
ethics committee (Ulm University).
2.2 Results
The following subsections summarize our findings on
rating decisions, pupil size changes and data mod-
elling.
2.2.1 Ratings
Participants were asked to evaluate 30 facial images
of the opposite sex on basis of a four-point scale.
16% of all cases were rated as (1) “Do not like at all”
whereas the majority of faces were either assessed as
(2) “Rather do not like” (35%) or (3) “Rather like”
(34%). 14 % were classified as (4) “Totally like”.
Four participants allocated images to categories 2 and
3 only.
2.2.2 Pupillometry
Absolute pupil diameter differs between participants.
To make individual responses comparable, we aver-
aged across the last 30 data points of baseline record-
ing and subtracted the result from each value during
the five seconds of image viewing (Grimmer et al.,
2021). P-values of the Shapiro-Wilk test indicated
that pupillary data was well modelled by a normal dis-
tribution.
The first half second after stimulus-onset involved
a pupil constriction that appeared comparably strong
while viewing disliked facial images (category 1).
Hereafter, pupil diameter increased and diverged ac-
cording to feelings of affection after approx. two sec-
onds. Averaged responses to preferred images (cate-
gory 3 and 4) continued to increase whereas pupil di-
ameter yielded a temporary constriction while view-
ing (rather) disliked images (category 1 and 2) (Fig.
3, left). Figure 4 depicts exemplary responses (indi-
vidual averages) to images that were rated as ”Do not
like at all” (category 1, left) and ”Totally like” (cate-
gory 4, right). As can be seen, variability between in-
dividuals increased with increasing viewing duration,
featuring similar courses only during the first second
after stimulus-onset.
Descriptive statistics suggested a tendency to-
wards larger diameters in response to attractively
rated faces (Fig. 3, right). Facial images of cat-
egory 1 (“Do not like at all”) involved an increase
of 0.09mm (SE: 0.03) compared to baseline mean
whereas category 2 stimuli (“Rather do not like”) in-
duced diameter changes of 0.11mm (SE: 0.02). Av-
eraged values during image viewing from category 3
(“Rather like”) proved to be similar (M: 0.11mm, SE:
0.02) whereas facial images from category 4 (“To-
tally like”) were associated with the strongest increase
in pupil diameter (M: 0.13mm, SE: 0.02). However,
HUCAPP 2024 - 8th International Conference on Human Computer Interaction Theory and Applications
524
Figure 3: Total averages (n=16) of pupil diameter while viewing images of the opposite sex. Values depict differences from
baseline mean and are grouped according to individual preferences. Left: Zero on the x-axis indicates stimulus-onset. For
purposes of presentation, signal course changes are smoothed using a moving window of 25 data points.
Figure 4: Individual averages (n=11) of pupil diameter while viewing images of the opposite sex that were rated as ”Do not
like at all” (category 1, left) and ”Totally like” (category 4, right). Values depict differences from baseline mean; zero on the
x-axis indicates stimulus-onset. For purposes of presentation, signal course changes are smoothed using a moving window of
25 data points.
a one-way repeated measures ANOVA (Greenhouse-
Geisser corrected) indicated no considerable differ-
ences (F(2.541, 27.950)= 1.149, p= 0.34).
In a further step, we compared pupillary responses
that involved clear preferences (category 1 vs cate-
gory 4) via a paired samples t-test and observed a
strong trend towards larger dilations while viewing
attractively rated faces (t(11)= -1.536, p= .08, d= -
0.44). However, this excludes trials that entail some
degree of uncertainty (category 2 and 3). Follow-
ing the selection technique from mobile dating apps,
we regrouped pupillary data into binary samples by
merging category 1 and 2 as well as category 3 and
4 and computed weighted averages due to different
number of data sets. In a subsequent t-test, the effect
of category was no longer visible (t(11)= -0.184, p=
.43). Four participants were excluded from inferential
statistics as they allocated images only to categories 2
and 3.
2.2.3 Time Series Classifciation
Machine learning (ML) methods enable to identify
complex correlations in time series data by figur-
ing out patterns and extracting useful features (Ruiz
et al., 2021). We implemented an ML-model that was
trained on the pupillary data from Experiment 1. To
reduce ambiguity, we addressed data from only those
trials that indicated clear preferences (category 1 and
4). Along with the outlier correction this resulted in
a total of 79% data reduction. However, data quality
was considerably increased and should be particularly
suited for the model to learn from it. The remaining
data were standardized by re-scaling the distribution
(M:0, SD:1), series lengths were normalized to adjust
the number of decimals.
We applied the state-of-the-art ROCKET model
(Random Convolutional Kernel Transform) that fea-
tures a high classification accuracy by transforming
Towards Automated Decision Making in Dating Apps Through Pupillary Responses
525
time series on basis of random convolutional kernels
and training a linear classifier (Dempster et al., 2020).
Data was split using Stratisfied Kfold (k=5) to ensure
a balanced proportion of both labels resulting in 102
samples to train ROCKET and 26 samples to test it.
Given the small data sets we selected the evaluation
strategy LOO (LeaveOneOut) built in a Ridge Classi-
fier. Training was carried out five times using differ-
ent random seeds on top of Stratisfied Kfold to cap-
ture the best model (25 models trained in total). Table
1 indicates relevant metrics for the best obtained uni-
variate ROCKET model featuring a prediction accu-
racy of at least 77% during binary classification (cat-
egory 1 vs 4).
2.3 Discussion
Volunteers classified 30 facial images into four cate-
gories according to their individual preferences. Re-
sults suggest a tendency towards moderate ratings
when dealing with facial attractiveness. Dislikes (cat-
egory 1) as well as strong likes (category 4) account
for only a small proportion of cases (around 30%),
suggesting that ambiguous responses and a degree of
uncertainty pose an ongoing challenge for cognition-
aware systems. It could be argued that our five female
volunteers saw different faces due to the presentation
of images of the opposite-sex. This is a valid objec-
tion and further research may balance the sex ratio
or focus on one sex only. On the other hand, recent
studies have found no differences in the pupillary re-
sponses of heterosexual men and women to images of
the opposite sex (Attard-Johnson et al., 2021).
Due to baseline referencing and controlled envi-
ronmental conditions, current changes in pupil size
cannot be explained through low-level factors such
as luminance or contrast. During early processing
stages (up to 1 second) pupil diameter was charac-
terized by low variability between individuals and
likely shaped by cognitive processes that underlying
attention (Math
ˆ
ot et al., 2014) or memory processes
(Naber et al., 2013) during face recognition. Two sec-
onds after stimulus onset, pupil responses diverged
according to attractiveness ratings. While pupil re-
actions to unfavourable images (category 1 and 2) de-
creased, there was a strong and sustained autonomous
response to favourable images (category 3 and 4).
These findings replicate previous studies, indicating
pupil diameter to increase when viewing preferred fa-
cial images of the opposite sex (Bernick et al., 1971;
Laeng and Falkenberg, 2007; Liao et al., 2021; Rieger
and Savin-Williams, 2012). However, arousal-related
changes in pupil size did not clearly distinguish be-
tween all four categories. Rather, it appeared that
pupil diameter was only sensitive to discriminating
physical states of clear rejection from those of clear
acceptance. More recently, Pronk et al. (Pronk et al.,
2021) conducted a similar study with a larger num-
ber of participants and confirmed a positive asso-
ciation between pupil diameter and the acceptance
of a hypothetical partner. Larger samples allow for
more precise estimation of statistical properties and
greater power; therefore it is reasonable to assume
that a larger sample would yield similar results in the
present case.
The ROCKET model yielded a promising perfor-
mance. Although it’s slightly biased towards the most
positive category (4), the model predicted results for
77% of unseen data correctly. For comparison only:
A dummy classifier randomly classifying samples or
constantly predicting the same class provided an ac-
curacy of no better than 57%.
3 EXPERIMENT 2
Experiment 2 adapts the design from Experiment 1.
However, the setting is transferred from fully con-
trolled laboratory conditions to various domestic loca-
tions. Moreover, video-based eye-tracking is replaced
by recordings from unmodified smartphone cameras.
3.1 Methods
The following subsections provide information on the
experimental design and specify details on technical
implementation, data processing and sample charac-
teristics.
Table 1: Classification metrics of the univariate ROCKET model for pupillary responses to facial images of category 1 (”Do
not like at all”) and category 4 (”Totally like”).
Precision Recall F1-score Number of samples
1 (category 4) 71% 92% 80% 13
0 (category 1) 89% 92% 73% 13
Accuracy 77% 26
Macro avg 80% 92% 76% 26
Weighted avg 80% 92% 76% 26
HUCAPP 2024 - 8th International Conference on Human Computer Interaction Theory and Applications
526
3.1.1 Design and Procedure
Subsequent to Experiment 1 we collected another
sample of volunteers in order to pilot a field study in
domestic environments. A new set of images (30 male
and 30 female faces) was selected from the same pool
of 400 AI-generated pictures (Karras et al., 2019). An
Android package (APK) including all relevant files
was emailed to the participants. They were asked to
install it on their smartphones and grant access to the
mobile storage media in order to acquire video data
from the front-facing camera. Experimental design
and task procedure corresponded to the setting de-
picted in 2.1.1. Volunteers were asked to pick a quiet
moment and sit comfortably while holding the phone
steady in front of their face. However, task process-
ing was not supervised and influencing factors like
viewing angle, distance from screen or ambient light-
ing could not be controlled. Upon completion of the
viewing task, all video files were uploaded to a server.
3.1.2 Technical Implementation
Pupillary data was captured from video files of the
front-facing camera of mobile phones. To obtain
the center of the pupil, we applied the deep learn-
ing framework GazeML which constitutes an estab-
lished method to recognize landmarks of the eyes
(Park et al., 2018). RGB values were employed for
pupil detection and comparison with the closest re-
gion’s color codes. To detect a similar colored pixel,
we contrasted nearby regions until color values ex-
ceeded a predetermined range of black shading (RGB
values 51,51,51).
3.1.3 Data Processing
Mobile devices featured varying frame rates and we
normalized sampling to a constant value of 25 frames
per second (fps). As during Experiment 1, the MAD
algorithm was applied to detect outliers (Leys et al.,
2013). However, while reading out pupillary data, rel-
evant features could not entirely be captured frame to
frame from video files which resulted in fragmentary
data sets (Fig. 5). As a consequence, seven partic-
ipants had to be removed from the data base due to
insufficient recordings.
3.1.4 Apparaus
Pupil diameter was determined on basis of video files
from the front-facing camera of latest Samsung smart-
phones. Android Studio, the development environ-
ment of Google’s operating system, was used to de-
sign the front end of the mobile app. The Python flask
Figure 5: Exemplary raw data trials featuring pupil size
changes during the image viewing task. Data were captured
from video files of front-facing smartphone cameras. Verti-
cal lines indicate segments of missing data).
framework was applied to create the backends’ REST
API.
3.1.5 Participants
12 male volunteers (M: 28 years, SD: 2) partici-
pated in Experiment 2. Seven data sets had to be
excluded from pupillometric analyses due to unsuc-
cessful recordings. Information on regular medica-
tion was not collected, however, volunteers reported
no history of head injury and no neurological or psy-
chiatric disorder. Written informed consent was ob-
tained prior to the experiment. All measurements
were performed in accordance with the Declaration
of Helsinki.
3.2 Results
The following subsections summarize findings on rat-
ing decisions and pupil size changes.
3.2.1 Ratings
Participants were asked to evaluate 30 facial images
of the opposite sex on basis of a four-point scale. 14%
of all cases were rated as “Do not like at all” whereas
the majority of faces were either assessed as “Rather
do not like” (40%) or “Rather like” (29%). 17% of
cases were classified as “Totally like”.
3.2.2 Pupillometry
A large part of pupillary data from the video files was
not analysable. Due to the small number of successful
recordings (n=5), we forego the documentation of de-
scriptive statistics and focus a qualitative evaluation.
During the first second of task processing, averaged
pupil size changes roughly corresponded to the eye-
tracking data during Experiment 1, indicating pupil
diameter to increase while viewing attractively rated
faces (Fig. 6). Smartphone recordings then yielded
strong responses to facial images from category 3
(approx. 0.3mm above baseline mean), whereas di-
ameters fell below baseline level while viewing pic-
Towards Automated Decision Making in Dating Apps Through Pupillary Responses
527
tures from category 2. Pupillary reactions to non-
preferred images (category 1) indicated a clear dila-
tion whereas pictures from category 4 (”Totally like”)
induced rather unspecific dynamics.
Figure 6: Total averages (n=5) of pupil diameter while
viewing images of the opposite sex. Values depict differ-
ences from baseline mean and are grouped according to in-
dividual preferences; zero on the x-axis indicates stimulus-
onset. Data were captured from video files of the mobile
phone front-facing camera.
3.3 Discussion
We applied the viewing task from Experiment 1 in a
noisy setting and during unsupervised task process-
ing, while pupil data were captured by video files
from front-facing smartphone cameras. Rating results
from Experiment 1 were essentially replicated, indi-
cating a tendency towards moderate ratings during at-
tractiveness judgments. However, there was a large
proportion of insuffient data for pupillometry. In addi-
tion, successful recordings (n=5) yielded inconsistent
results, featuring both a sustained drop below baseline
level (category 2) and a particularly strong response
(category 3) when processing rather moderate stim-
uli.
Rafiqi et al. (Rafiqi et al., 2015) obtained pupil
size changes from unmodified smartphone cameras
in a fully controlled setting and reported high corre-
lations with control measures from a certified Eye-
tracker. Therefore, it can be assumed that our data
loss was mainly due to processing issues, including
low ambient lighting and unfavourable viewing dis-
tances or camera angles. In addition, the current
sample consisted of volunteers from an international
study program, including a large proportion of Indian
students with rather dark eyes. This presents an addi-
tional challenge to the pupil detection algorithm, es-
pecially under unfavourable lighting conditions. The
latter may be countered by environmental illumina-
tion compensation as recently introduced by Gollan
et al. (Gollan and Ferscha, 2016). However, cogni-
tive pupillometry based on digital cameras requires a
certain degree of environmental control and a rather
stationary camera position.
4 CONCLUSION
The present study takes a further step towards auto-
mated decision making in dating apps using arousal-
based information. We applied pupillometry as an
established method to map transient and long-lasting
changes in underlying cognitive and affective pro-
cessing. Under laboratory conditions, pupil diam-
eter discriminated between states of clear rejection
and clear acceptance. Accordingly, data modelling
yielded promising results to enable proper automa-
tion of binary classifications. However, a controlled
setting runs contrary to the intended scenario of life-
like mobile dating app usage. Furthermore, the ques-
tion of whether this type of automation is desirable
in principle remains unanswered. A perceived loss of
control during partner selection could lead to rejec-
tion of the application. Therefore, usability measures
and qualitative data analyses should accompany fur-
ther research to ensure a positive user experience.
REFERENCES
Attard-Johnson, J., Vasilev, M. R.,
´
O Ciardha, C., Binde-
mann, M., and Babchishin, K. M. (2021). Measure-
ment of sexual interests with pupillary responses: A
meta-analysis. Archives of sexual behavior, 50:3385–
3411.
Bernick, N., Kling, A., and Borowitz, G. (1971). Phys-
iologic differentiation of sexual arousal and anxiety.
Psychosomatic medicine, 33(4):341–352.
Berscheid, E. and Walster, E. (1974). Physical attractive-
ness. In Advances in experimental social psychology,
volume 7, pages 157–215. Elsevier.
Buss, D. M. (2016). The evolution of desire: Strategies of
human mating. Hachette UK.
Castro,
´
A. and Barrada, J. R. (2020). Dating apps and their
sociodemographic and psychosocial correlates: A sys-
tematic review. International Journal of Environmen-
tal Research and Public Health, 17(18):6500.
Dai, X., Brendl, C. M., and Ariely, D. (2010). Wanting, lik-
ing, and preference construction. Emotion, 10(3):324.
Damasio, A. R. (1996). The somatic marker hypothesis and
the possible functions of the prefrontal cortex. Philo-
sophical Transactions of the Royal Society of London.
Series B: Biological Sciences, 351(1346):1413–1420.
Dempster, A., Petitjean, F., and Webb, G. I. (2020). Rocket:
exceptionally fast and accurate time series classifica-
HUCAPP 2024 - 8th International Conference on Human Computer Interaction Theory and Applications
528
tion using random convolutional kernels. Data Mining
and Knowledge Discovery, 34(5):1454–1495.
Ehlers, J., Strauch, C., Georgi, J., and Huckauf, A. (2016).
Pupil size changes as an active information channel
for biofeedback applications. Applied psychophysiol-
ogy and biofeedback, 41:331–339.
Ehlers, J., Strauch, C., and Huckauf, A. (2018). A view
to a click: Pupil size changes as input command in
eyes-only human-computer interaction. International
journal of human-computer studies, 119:28–34.
Einh
¨
auser, W., Stout, J., Koch, C., and Carter, O.
(2008). Pupil dilation reflects perceptual selection
and predicts subsequent stability in perceptual rivalry.
Proceedings of the National Academy of Sciences,
105(5):1704–1709.
Fink, B. and Penton-Voak, I. (2002). Evolutionary psychol-
ogy of facial attractiveness. Current Directions in Psy-
chological Science, 11(5):154–158.
Gollan, B. and Ferscha, A. (2016). Modeling pupil dilation
as online input for estimation of cognitive load in non-
laboratory attention-aware systems. COGNITIVE.
Grimmer, J., Simon, L., and Ehlers, J. (2021). The cog-
nitive eye: Indexing oculomotor functions for mental
workload assessment in cognition-aware systems. In
Extended Abstracts of the 2021 CHI Conference on
Human Factors in Computing Systems, pages 1–6.
Hahn, A. C. and Perrett, D. I. (2014). Neural and behavioral
responses to attractiveness in adult and infant faces.
Neuroscience & Biobehavioral Reviews, 46:591–603.
Kardon, R. (1995). Pupillary light reflex. Current opinion
in ophthalmology, 6(6):20–26.
Karras, T., Laine, S., and Aila, T. (2019). A style-based
generator architecture for generative adversarial net-
works. In Proceedings of the IEEE/CVF conference
on computer vision and pattern recognition, pages
4401–4410.
Kucewicz, M. T., Dolezal, J., Kremen, V., Berry, B. M.,
Miller, L. R., Magee, A. L., Fabian, V., and Worrell,
G. A. (2018). Pupil size reflects successful encoding
and recall of memory in humans. Scientific reports,
8(1):4949.
Laeng, B. and Falkenberg, L. (2007). Women’s pupillary
responses to sexually significant others during the hor-
monal cycle. Hormones and behavior, 52(4):520–
530.
Leys, C., Ley, C., Klein, O., Bernard, P., and Licata, L.
(2013). Detecting outliers: Do not use standard devi-
ation around the mean, use absolute deviation around
the median. Journal of experimental social psychol-
ogy, 49(4):764–766.
Liao, H.-I., Kashino, M., and Shimojo, S. (2021). Attrac-
tiveness in the eyes: A possibility of positive loop be-
tween transient pupil constriction and facial attraction.
Journal of Cognitive Neuroscience, 33(2):315–340.
Love, J., Selker, R., Marsman, M., Jamil, T., Dropmann,
D., Verhagen, J., Ly, A., Gronau, Q. F.,
ˇ
Sm
´
ıra, M.,
Epskamp, S., et al. (2019). Jasp: Graphical statistical
software for common statistical designs. Journal of
Statistical Software, 88:1–17.
Mantiuk, R., Kowalik, M., Nowosielski, A., and Bazyluk,
B. (2012). Do-it-yourself eye tracker: Low-cost pupil-
based eye tracker for computer graphics applications.
In Advances in Multimedia Modeling: 18th Interna-
tional Conference, MMM 2012, Klagenfurt, Austria,
January 4-6, 2012. Proceedings 18, pages 115–125.
Springer.
Math
ˆ
ot, S., Dalmaijer, E., Grainger, J., and Van der
Stigchel, S. (2014). The pupillary light response re-
flects exogenous attention and inhibition of return.
Journal of vision, 14(14):7–7.
Naber, M., Fr
¨
assle, S., Rutishauser, U., and Einh
¨
auser, W.
(2013). Pupil size signals novelty and predicts later
retrieval success for declarative memories of natural
scenes. Journal of vision, 13(2):11–11.
Park, S., Zhang, X., Bulling, A., and Hilliges, O. (2018).
Learning to find eye region landmarks for remote gaze
estimation in unconstrained settings. In Proceedings
of the 2018 ACM symposium on eye tracking research
& applications, pages 1–10.
Pronk, T. M., Bogaers, R. I., Verheijen, M. S., and Sleegers,
W. W. (2021). Pupil size predicts partner choices in
online dating. Social Cognition, 39(6):773–786.
Rafiqi, S., Wangwiwattana, C., Fernandez, E., Nair, S.,
and Larson, E. (2015). Work-in-progress, pupilware-
m: Cognitive load estimation using unmodified smart-
phone cameras. In 2015 IEEE 12th International Con-
ference on Mobile Ad Hoc and Sensor Systems, pages
645–650. IEEE.
Rieger, G. and Savin-Williams, R. C. (2012). The eyes have
it: Sex and sexual orientation differences in pupil di-
lation patterns. PloS one, 7(8):e40256.
Ruiz, A. P., Flynn, M., Large, J., Middlehurst, M., and Bag-
nall, A. (2021). The great multivariate time series clas-
sification bake off: a review and experimental evalua-
tion of recent algorithmic advances. Data Mining and
Knowledge Discovery, 35(2):401–449.
Sawyer, A. N., Smith, E. R., and Benotsch, E. G.
(2018). Dating application use and sexual risk be-
havior among young adults. Sexuality Research and
Social Policy, 15:183–191.
Skakoon-Sparling, S., Cramer, K. M., and Shuper, P. A.
(2016). The impact of sexual arousal on sexual
risk-taking and decision-making in men and women.
Archives of sexual behavior, 45:33–42.
Wu, S. and Trottier, D. (2022). Dating apps: a literature
review. Annals of the International Communication
Association, 46(2):91–115.
Towards Automated Decision Making in Dating Apps Through Pupillary Responses
529