Passive identification of subjective preferences toward individual items using eye-tracking in a virtual reality environment

The usage of virtual reality (VR) has been growing in many fields of research and therapy thanks to its immersive and gamified nature. Detection of the users’ subjective experience is thus essential for the effective personalization of content. Eye-tracking (ET) data and specifically gaze, in two-dimensional tasks, has been linked to value-based choices and emotional states. Therefore, here we aimed to develop a method for passive identification of subjective preferences based on ET data collected during a VR experience. For this purpose, we developed a naturalistic dynamic VR task where participants searched and looked at complex objects of pets and their control shapes that appeared in pre-defined locations in random order. At the end of the task, participants ranked their preference, valence, and arousal of the items they saw during the task. ET data was recorded using a built-in binocular eye-tracker within the VR headset. We found that the gaze behavior features of the median distance of gaze from the center of objects and the median gaze scan speed showed a significant interaction with object type (pets/shapes), as well as a significant positive relation to preference and valence rankings of pets. Our results suggest that these gaze behavior features could be used as passive biomarkers for detecting individual preferences and pleasantness, and in the future may enable successful personalization of VR content in real-time for various applications such as optimization of psychiatric diagnosis and treatment sessions.

In recent years, personalization has become popular in various fields of research from medicine (Schleidgen et al. 2013;Schork 2015), to education (Hwang et al. 2012;Ravyse et al. 2017;Xiao et al. 2017;Bernacki and Walkington 2018), through marketing (Ng and Wakenshaw 2017), to genome-and microbiome-based nutrition (Bashiardes et al. 2018). Computer-based applications benefit from personalization. For example, developing an educational computer game that is personalized based on learning style improved learning achievements and increased learning 1 3 motivation (Hwang et al. 2012). Likewise, in their review, (Ng and Wakenshaw 2017) thoroughly discussed utilizing the Internet of Things (IoT) data in marketing for personalizing in products offering to meet the consumers' geographic location and predicted needs, as well as for research of the influence of context on costumers behavior and decisions. In a similar manner to 2D applications, the widely spreading usage of VR discussed above will also benefit from content personalization. A natural direction is for real-time adaptation gameplay according to the player's mood, performance, and behavior (Soler-Dominguez et al. 2017). In addition, it was suggested that outcomes of treatment of mental health disorders could be enhanced with VR (for a detailed discussion see (Freeman et al. 2017)). More specifically, a recent study showed a significant reduction in agoraphobic avoidance and distress using automated VR therapy with a virtual coach that provided feedback to personalize the therapy progression relative to the usual care group in patients with psychosis (Freeman et al. 2022). In the field of education, a recent meta-analysis (Wu et al. 2020) suggested that headmounted displays (HMDs) can improve knowledge, skill development, and learning maintenance over time. These may be enhanced even more with personalized adaptivity and feedback as was reviewed for serious games (Ravyse et al. 2017) and could be applied to VR as well. Thus VR, being more immersive, providing more behavioral measures can both obtain better personalization with greater impact.
Thus, to allow for effective content personalization, the challenge is to be able to identify the subjective experience of users at each point in time during a VR task. To address this challenge, attempts have been made to record physiological signals. For example, to elicit and recognize affective states (valence and arousal), electroencephalography (EEG) and electrocardiography (ECG) signals were recorded during participants' immersion in four virtual rooms via HMD (Marín-Morales et al. 2018). In another study, respiration, ECG, electromyography (EMG), and electrodermal activity (EDA) signals were recorded to develop a biofeedback system for emotion awareness and regulation strategies in VR (Bermudez I Badia et al. 2019). However, the acquisition of these physiological signals required expertise for their operation and designated equipment beyond the mere access to a VR headset. This highlights the need for subjective experience detection in VR that is within reach for common users and that does not disrupt the common usage of HMDs.
Recent VR headsets have been pre-installed with eyetracking (ET) devices (for example, HTC-Vive https:// www. vive. com/ eu/ produ ct/ vive-pro-eye/ overv iew/, the Pico Neo https:// www. picoxr. com/ us/ neo3. html, the Varjo headset https:// varjo. com/ produ cts/ aero/, and the Meta Quest Pro https:// www. meta. com/ quest/ quest-pro/). Thus, ET is becoming accessible in advanced HMDs with the potential to benefit various research questions and applications. One of the dominant advantages of ET in VR is experimental control of content, while mobility is preserved (Meißner et al. 2017) and due to its non-intrusive mechanism (Soler-Dominguez et al. 2017) and its potential to facilitate personalization in VR was suggested to benefit different fields as education (Mikhailenko et al. 2022), gaming (Soler-Dominguez et al. 2017), retail (Pfeiffer et al. 2020), as well as for emotion recognition (Lim et al. 2020).
Although relatively new in 3D VR, the use of ET has been a promising approach in 2D studies to continuously measure subjective experiences in human-computer interfaces (Jacob and Karn 2003). It has been shown that it is possible to infer the individuals' emotional arousal (Lanatà et al. 2013;Aracena et al. 2015) and valence (R-Tavakoli et al. 2015) and different combinations of their values (Raudonis et al. 2013) by using ET while participants were observing 2D images. Gaze measures have also been linked to value-based decision-making and preferences in 2D (Shimojo et al. 2003;Krajbich et al. 2010;Dixson et al. 2011;Krajbich and Rangel 2011;Graham and Jeffery 2012;Van Loo et al. 2015;Noland et al. 2017). Most 2D ET studies focused on gaze measures derived from fixations, during which gaze is maintained on a single location; and saccades, which are ballistic eye movements of gaze from one fixation to the other (Rayner 2009). The accumulated amount of time gazed on an object, also referred to as dwell time, is the primer gaze measure that was found to be related to preferences and value-based choice (Shimojo et al. 2003;Krajbich et al. 2010;Krajbich and Rangel 2011;Graham and Jeffery 2012;Kim et al. 2013;Noland et al. 2017). In addition, the gazed item identity and the dwell time of the first and last time points were correlated with choice and item liking rating (Krajbich et al. 2010;Krajbich and Rangel 2011). When fixations order, and their respective dwell time were studied in a trinary choice task (Krajbich and Rangel 2011), a complex relationship of fixation patterns was found to choice selection. Additional ET measures that were tested for their relation to preferences of urban landscape images in addition to viewing time were the number of fixations, the time to first fixation and the fixation region category (Kim et al. 2013;Noland et al. 2017;Wu et al. 2021) as well as the fixation duration, saccade counts, scan paths, pupil size, and heat map (Kim et al. 2013). In another study, a combination of various time and space-related fixation and saccade event statistics including count, duration, direction, position, and time was fed into a random forest algorithm to classify the preferred image out of a viewed pair (Sugano et al. 2014).
In the field of emotion or internal states classification, the circumplex model of emotion is most commonly used (Russell 1980). The model includes a four-quadrant combination of valence and arousal dimensions with varying emotions in each quadrant that correspond to the high or low arousal and negative or positive valence. Several studies aimed at classifying the emotional state elicited by the displayed content with ET data. To detect arousal levels of emotional images extracted from the International Affective Picture System (IAPS), a wireless ET system was used by (Lanatà et al. 2013). Raw samples were considered as fixations, and several spatial and temporal features were computed including fixations' position along diagonal lines, their length, maximum dwell time per pixel or area, gaze path length, and pupil size. The features were fed into a multilayer perceptron to classify neutral vs. maximum level of arousal images with the lowest valence and obtained preliminary promising results. Another study used a selection of IAPS images (Aracena et al. 2015) and modified them to get a similar level of luminosity. Neural networks and decision trees were used to classify the images negative, neutral, and positive arousal levels according to pupil diameter and gaze position. The contribution of eye movement measures to valence classification of IAPS images was addressed by (R-Tavakoli et al. 2015). The histogram representations for features extracted from saccades and fixations were saccade orientation, slope, length, duration, velocity, and fixation duration, as well as most visited coordinates, and saliency map. The features' classification performance was assessed by a support vector machine (SVM) with several combination schemes. It was found that the most contributing features were the histograms of fixation, fixation duration, and saccade slope, and the saliency map. An artificial neural network (ANN) was trained on the pupil's size, position, and motion speed, acquired by an ET device fixed on ordinary eyeglasses (Raudonis et al. 2013). Classification of four categories of emotional images: neutral, disgusted, funny, and interested was achieved with approximately 10% deviation using data of 2 s. In a comparison of EEG and pupillary response including blinks features to classify videos of three classes of arousal and valence, classification based on ET performed better than based on EEG (Soleymani et al. 2012).
As the availability of ET in HMD VR increased only in the very recent few years, to the best of our knowledge, at the time this study was set to motion, no other studies were published that aimed at preference or affect detection with or without personalization of content. Therefore, there were not standard practices in the field. A few related studies have been published, while our study was conducted. One study was based on the gaze 'cascade effect' (Shimojo et al. 2003), where it was found that due to the evaluation process itself, the gazed time varies and may alter the preference and vice versa between two images that were displayed side by side in an HMD choice task (Sato et al. 2022). A recurrent neural network (RNN) classifier trained by eye movement data achieved approximately 91 percent success when the eye gaze data were longer than 3 s. In another study, 360° videos with four combinations of arousal and valence levels that represent the respective emotions were presented in an HMD VR headset (Zheng et al. 2020). Each of the four emotions was classified with either SVM, k-nearest neighbor (KNN), and random forest (RF) based on pupil diameter solely. A more recent study by this group (Zheng et al. 2021) found that the classification performance based on fixation position was better compared to pupil diameter. An additional study that presented 360° videos with four combinations of arousal and valence levels to trigger emotions captured ET, ECG, and galvanic skin response (GSR) (Tabbaa et al. 2021). The ET features extracted were basic statistics of fixations (number per second, first, duration), saccades and microsaccades (number per second, direction, duration, peak velocity, and blinks (number per second, duration). The classification results of an SVM were compared and the ET features substantially outperformed the ECG features for arousal and the GSR features in valence for four-class classification. To allow the personalization of VR interfaces for language and concept learning, ET metrics were extracted to predict the difficulty level of presented words in HMD VR (Orlosky et al. 2019). An SVM classifier was built based on features including fixation duration, saccade frequency, pupil size, blink frequency, eye movement per second, and head movements. They found that the most significant metric was the average eye movement per second while gazing at the word.
Other studies that included ET analysis targeted at personalization applications in a virtual environment were not performed with a built-in ET in an HMD VR. With the purpose of encouraging pulses choice and consumption in France, gaze behavior toward pulses in a virtual supermarket in different shelf compositions was studied (Melendrez-Ruiz et al. 2021). The point at the center of the screen which was obtained using a Gear VR headset was considered a proxy to gaze tracking of objects (products) that were identified as in the region of interest. To allow for the future personalization of information systems, in a comparison study of a cave automatic virtual environment (CAVE) vs. a real physical supermarket, mobile ET was used to classify one of the two search tasks: goal-directed and exploratory search (Pfeiffer et al. 2020). An SVM classifier trained on a wide range of fixations and saccades derived features obtained promising results for both virtual and physical environments. And with the purpose to develop future intervention methods for trainees, visual scan paths were studied in a virtual simulated environment of the Deepwater Horizon operation that was displayed on a large 2D screen (Kang et al. 2020).
Prior work relating gaze measures to preferences and affect mostly examined fixations, saccades, and their derivatives. These two ocular events characterize eye movement patterns during the majority of sedentary tasks and those that consist of viewing stationary stimuli (Lappi 2016;Kim et al. 2020). However, the eye movements that represent viewing a moving object are named smooth pursuits for which examining raw samples data was recommended (Kim et al. 2020;Carter and Luke 2020). Considering that VR offers a naturalistic experience mimicking daily life viewing tasks, these usually include gazing at a moving object due to the movement of the observer or the viewed object. Previous work with ET in VR utilized eye movements of smooth pursuit for interaction with moving objects, where different types and features of interactions were examined (Piumsomboon et al. 2017;Khamis et al. 2018). Furthermore, in 2D screens, ET samples and fixations can be obtained in specific (x,y) coordinates, or pixels according to the screen resolution. However, in 3D HMD VR, the ET samples and fixations are obtained in slightly different (x,y,z) coordinates on the object due to different perspectives and points of view that are derived from the free personal movement and position of the head. These factors imply that naturalistic gaze behavior should be studied in 3D VR while observing moving objects, allowing participant movement, and examining global gaze pattern behavior that captures spatial dynamics of eye scan of an object, rather than specific fixations' location. Furthermore, the fact that HMDs include eye trackers very close to the eyes offer a better technological solution that allows for better eye gaze signal.
The focus of previous studies was mainly on fixations and their dwell time, hotspot, counts, and order mainly between objects, and some within objects in the field of preferences and value. In the emotion detection literature, even though ET studies were performed while observing single items, mostly the combinations of fixations, saccades, and pupillometry features were fed altogether into complex analyses of machine and deep learning algorithms to detect affective states, and very few studies analyzed the relation of individual gaze pattern features to affect. Previous analyses lacked an interpretable description of the gaze behavior within a single object and its relation to preferences or affect that is necessary for enabling personalization of content for various tasks and applications.
Subjective preferences were formerly studied either including an explicit ranking of single items (Graham and Jeffery 2012;Noland et al. 2017) or an assessment of ranking via choosing between two or three objects in the display (Shimojo et al. 2003;Krajbich et al. 2010;Krajbich and Rangel 2011;Van Loo et al. 2015), or by simultaneously presenting and ranking four objects per trial (Wu et al. 2021). These experimental designs involved an active assessment, evaluation, or choice, which means that participants were viewing the items presented while actively involved in one of these valuation tasks. Such tasks induced gaze behavior that was derived from some kind of active viewing task. In a study that compared two active scene viewing tasks of visual search and memorization (Castelhano et al. 2009), several eye movement measures were found to be influenced by the viewing task. The effect of the viewing task was additionally examined on classifiers' performance to predict preferences of two natural images based on gaze data (Sugano et al. 2014). Lower accuracy was obtained for the classifier that used the free viewing condition data than for the one that used the labeling task data, suggesting that eye movements were influenced by the activity of preference or choice itself. Differences in eye movement measures during active comparison between two or more objects, or ranking of single objects, might be caused by the involvement of the gaze 'cascade-effect' (Shimojo et al. 2003), for which the gazed time both affects and is affected by the preference. It was suggested that even in a free viewing task (without labeling) with two images simultaneously displayed, the comparison of the images can induce a gaze bias (Sugano et al. 2014), and the importance of understanding the relationship between subjective preference and gaze behavior for single items was deduced. Likewise, in the studies of emotion or affect detection, either pre-classified labeling or ranking of arousal, valence or other emotions were done by another group of people (Raudonis et al. 2013;Lanatà et al. 2013;Aracena et al. 2015), or the participants themselves performed the ranking during the task for each stimulus (R-Tavakoli et al. 2015), or both conditions were compared (Soleymani et al. 2012;Tabbaa et al. 2021). Hence, in those, the participants were either aware of the emotion evaluation process itself while observing the stimuli, as was done in the value-based choice literature, or were not necessarily aligned to the pre-defined labeling. Therefore, studying the relation of gaze measures to subjective preference measures of single items in a passive (e.g., non-active) viewing task, that does not include an overt value or affect estimation is novel and would shed light on this unexplored field. Considering the natural process of value representation being covertly manifested spontaneously as we passively gaze at objects even without a defined purpose, a passive task design is especially appropriate in the endeavor for ET-based content personalization. In addition, ET data gathering on objects, not necessarily requiring a subsequent choice or action may allow generalization to interaction with content that is not task dependent.
Here, we aimed to develop an objective and quantitative method for the identification of subjective experience measures using ET during an HMD VR experience. Based on the above-mentioned findings mostly in 2D research, relating ET to subjective preferences and affect (measured by valence and arousal), we aimed to test whether ET measures could passively reflect individual experience in 3D VR. We set to investigate passive gaze time series patterns of individual items to detect subjective experience measures without choice or valuation elicitation. We developed a novel dynamic VR task, during which participants explored a scene using head movements and gaze. The scene included pets and control shapes that moved in their location on a stationary background and participants were requested to observe the jumping stimuli. Following the viewing task, in a subsequent part of the experiment participants ranked their preference, valence, and arousal of the pets and control shapes.
We found a significant positive correlation with preference and valence rankings only for the pets category with two gaze measures: median distance of gaze from the center of the object, and the median scan speed of gaze. Both were significant while accounting for object size and percent time gazed (to account for mere exposure effect, MEE (Zajonc 1968)). These novel findings shed new light on gaze behavior patterns of different object types during natural 3D viewing and demonstrate the potential use of eye-tracking features as passive biomarkers for identifying and enhancing subjective preference and pleasantness that could be used for real-time VR content personalization. Taking into consideration the increasing availability of ET in VR HMDs, our findings have the potential to enhance the optimization of different tasks in VR, such as learning, psychiatric diagnosis and treatment sessions.

Participants
The participants included in this study were healthy, with intact or corrected vision between the ages of 18 and 40. The experiment exclusion criteria that were pre-defined were: participants with dizziness, pregnant women, participants with eyeglasses only for near vision correction (not contacts lenses; according to eye-tracker manufacture instructions), and participants who regularly consume anti-depressants or stimulants. Participants gave their informed consent to participate in the experiment and received monetary compensation for their time of 40 NIS per hour (~ 12$). The study was approved by the ethics committee of Tel-Aviv University. We did not collect previous VR experience. Studies with other more complex focuses such as spatial learning (for example, (Maneuvrier et al. 2020)) usually report these measures. The experimenter did ensure that participants could read the instructions and that the eye-tracker was calibrated prior to the beginning of the task.
A total of eighty-six (86) participants were recruited, out of which six participants were excluded from any analyses due to the following reasons: four participants declared eventually that they used antidepressant medication and thus met one of the exclusion criteria. One participant did not complete the VR task due to his report of dizziness and requested to discontinue the experiment. The data of one participant were not recorded due to technical issues. Additional exclusion criteria were predefined before data collection completion and before data analysis which was related to the data validity of each ranking session as detailed in the Ranking section below. One participant was excluded from all statistical analyses due to the ranking exclusion criteria of all ranking sessions. This resulted in a final study sample of 79 participants, (42 females) in the age of (M = 25.28, SD = 4.17). The sample size rationale is detailed in the Preregistration note section.

Experimental setup
The VR task was performed using a commercial HMD VR system (HTC-Vive), with a screen resolution of 1080 × 1200 pixels per eye and a 90-Hz refresh rate. The VR headset was adapted with a built-in binocular ET device (Tobii-installed) that provided pupil diameter data, eye gaze and head movement tracking while participants were engaged in VR. During the VR task, ET samples were recorded at a frequency of 120 Hz and a manufacturer-estimated accuracy of 0.5°. The VR task was programmed and developed in C# in Unity 3D (version 2018.3) environment and ran on a dedicated computer with a Core i7-7700 K CPU at 4.20 GHz, 32.0 GB RAM, an NVIDIA GeForce GTX 1080 TI GPU, and a 64-bit Windows10 operating system. The ET data acquisition was integrated into the task code and environment using Tobii's software development kit (SDK) for Unity and edited to fit the task design and specifications.
The VR scene was based on a purchased scene from the Unity asset store and consisted of a pastoral house (which was downloaded from:https:// asset store. unity. com/ 3d? categ ory= 3d& versi on= 2018&q= ArchV izPRO% 20Int erior% 20Vol. 4& order By=6) and was imported to Unity and modified to fit the purpose of the task. Some of the task stimuli were also downloaded and extracted out of packages available at the Unity asset store (https:// asset store. unity. com/ packa ges/ 3d/ chara cters/ anima ls/ mamma ls/ cute-pet-96976; https:// asset store. unity. com/ packa ges/ 3d/ chara cters/ anima ls/ fish/ cute-whale-115646; https:// asset store. unity. com/ packa ges/ 3d/ envir onmen ts/ propr ototy pe-colle ction-46952? categ ory= 3d& versi on= 2018&q= Primi tiveP lus& order By=1), and modified to fit the task purpose. Some stimuli were uniquely modeled in Blender and imported to and further designed in Unity. The participants' control over the embedded task instructions, as well as the ranking selection during the ranking sessions, was performed using a trigger press on the Vive controller.

Experimental procedure
The general task flow comprised two main stages beginning with a dynamic task and concluding with a ranking session (see Fig. 1).

Dynamic viewing task
The placement of the HMD on participants was followed by the first ET calibration, to ensure ET signals were recorded properly while participants were engaged in the dynamic viewing task. The first part of the task was carried out while participants were standing and instructed to explore a pastoral house using body posture and head movements within the VR environment. Participants were requested to search and look at task stimuli consisting of animals (pets) and control objects (shapes) that appeared in pre-defined locations in the scene (see Fig. 2). To make the task closer to real-life viewing conditions and to enhance participants' engagement and feeling of presence, the stimuli jumped in place up and down and therefore were distinguishable compared to the scene furniture that was fixed. Once the objects were recognized, the participants were asked to look at them until they disappeared and thereafter to continue and search and look at the next object, and so forth. The task included two consecutive runs, during which a total of 48 stimuli (24 pairs of animals and shapes) appeared in each run in a random order (see Fig. 2). Each pet and shape pair appeared in the same location in the scene and had resembling color and size (see Fig. 2). This design directed at the possible confounds of the objects' color, size, and position in the scene, that itself accounted both for the distance between the participant and the object and the effect of the entire scenery view at that location. Each object remained visible in the scene for 2 s from the moment it was first identified by the eyetracker. This resulted in equal exposure times for all objects to prevent the possible confound of the mere exposure effect (Zajonc 1968(Zajonc , 2001Zajonc and Markus 1982). The intertrial interval (ITI) was at least 2 s, which was the time gap between the appearance of the next object and after the previous object disappeared. In case, the next object was not seen by the participants after 2 s from its appearance, a beep sound was played to inform the participants that an object is somewhere in the scene outside of their current field of view and that they needed to look for it. This was meant to enhance task engagement. The task began with 5 s of scene adaptation before the first stimulus was presented.

Ranking
Following the dynamic task, the participants ranked their preference for the stimuli they saw, and the valence and arousal levels they felt while they observed the stimuli during the task. The ranking was performed on a pseudo-continuous  Fig. 1 Task flow building blocks. a The dynamic viewing task was conducted while standing and freely moving the head and body (no walking). The task began with an ET calibration that was followed by two runs in which the 48 stimuli appeared in random order. b The ranking sessions were performed while seated and freely moving the head and hands. Following additional ET calibration, all participants ranked their preference for the stimuli. Thereafter, valence and arousal ranking sessions were performed in a counter-balanced order across participants scale that contained unbeknownst to participants 37 bins (see Fig. 3). Therefore, the ranking score for each stimulus and each ranking session consisted of numeric integer values in the range of 0 to 36. This part of the experiment was conducted while participants were seated and after the second ET calibration to ensure proper ET readout in the updated positioning. In the preference ranking session, the participants ranked how much they liked each object on a scale ranging from "do not like" to "like very much" (see Fig. 3). In the valence ranking session, the participants ranked how pleasantly they felt when they observed each object on a scale ranging from "very unpleasant" to "very pleasant". In the arousal ranking session, the participants ranked the arousal level they felt when they observed each object on a scale ranging from "very unaroused" to "very aroused". Since preference was the main subjective measurement that we aimed to test, it was the first to be ranked for all participants. Following that, the ranking order of arousal and valence was counterbalanced across participants. During each ranking session, each stimulus was presented and ranked solely with the same neutral pastoral background, in a sequential random order with an ITI of 2 s. On each ranking trial, the stimulus first appeared on the screen (see Fig. 3a). Two seconds after the eye-tracker identified that the participant's gaze hit the object then a ranking scale appeared on the screen below the item (see Fig. 3b). This resulted in identical exposure time to the items to prevent the possible confound of the mere exposure effect (Zajonc 1968(Zajonc , 2001Zajonc and Markus 1982) also in the ranking phase. The selection of the rank by participants was conducted in an engaging fashion, as the controller in the participants' hand radiated a green "laser beam", that colored the scale location pointed at in green (see Fig. 3b). Finally, the ranking was chosen using a trigger press while the scale region of interest was pointed at and marked in green.

Variables
We aimed to study eye movements in a 3D ecological environment while viewing moving objects and thus used raw samples for the analysis of gaze behavior (Carter and Luke 2020). Each ET data sample included measurements of gazed ray origin (x,y,z), gazed ray direction (x,y,z), headset position origin (x,y,z), headset position direction (x,y,z), pupil diameter (mm), sample validity defined by the eye-tracker, and an arbitrary time stamp in millisecond precision. We also collected a specific time stamp readout in microseconds precision that allowed synchronization of the data with the task timeline and multiple object variables. These included: gazed object name; gazed object center position (x,y,z), for which we considered the center of the object bounding box (see Fig. 4b, c); and the gazed ray hit point (x,y,z), which was the location of collision of the gaze ray with the gazed object (see Fig. 4a). The majority of these measures were used for feature extraction as detailed in the Feature extraction section below. Samples validity and time stamps were used for trial validity analysis as detailed in the eye-tracking data pre-processing section below. Fig. 2 Dynamic viewing VR task setup. Pairs of pets (middle) and shapes (bottom) appeared at the same location in the scene on different trials, where the scene's default design is shown at the top panel. a A cat and a cube. b A chick and a sphere. *Similar shape objects were used in the task. **Images captured from the scene are published with consent from the asset designers. (https:// asset store. unity. com/ 3d? categ ory= 3d& versi on= 2018&q= ArchV izPRO% 20Int erior% 20Vol. 4& order By=6, https:// asset store. unity. com/ packa ges/ 3d/ chara cters/ anima ls/ mamma ls/ cutepet-96976

Eye-tracking data pre-processing
Pre-processing of eye-tracking signals was programmed with Python 3.6 software (https:// www. python. org/). ET samples with Nan values or non-valid samples, as defined by Tobii's SDK, were treated as blinks. We additionally considered as blinks all samples in a 150 ms time window before each blink onset and after each blink offset, as was done previously for 2D data in our laboratory (Salomon et al. 2020). All blink samples were scrubbed before the analysis. We included in the analysis only trials during which at least 50% of ET samples were valid and not Nan values, and that the stimulus gazed time was at least 50% of the 2 s during which it was presented from the first gaze time. Participants with less than 50% of valid trials according to these criteria were excluded from further ET analysis as in previous studies from our laboratory using a 2D eye-tracker (Salomon et al. 2020).

Ranking analysis
Valid ranking data for each of the ranking sessions was defined as having a minimal standard deviation (SD) of 4 (bins) for the pets (since the power analysis registered was focused on a model for pets, and as the ranking SD of the shapes was expected to be lower than that of the pets due to their similarity to each other and relatively simple appearance). This threshold was based on two criteria that were detailed in the registration page (https:// osf. io/ h9bpq).

Feature extraction
We aimed to study the link between gaze patterns on moving objects and individual experience measures. We calculated several eye-tracking features for each trial after data preprocessing. Calculated features included: percent time the object was gazed at out of the time it was presented (referred to as GazedTime), right and left pupil diameter percentiles (10th, 50th, 95th percentiles) of samples while gazing the object continuously for at least 0.5 s (referred to as Per10_ PupilR/L, Per50_PupilR/L, and Per95_PupilR/L, respectively); gaze shift distance percentiles (10, 50, 95) that was defined as the x,y,z Euclidian distance between the gazed hit point to the centroid of the gazed object (referred to as Min_Per_Dist, Med_Dist, and Max_Per_Dist respectively), gaze shift distance percentiles (10, 50, 95) normalized by the Euclidian distance between the gazed ray origin and the gazed object centroid (referred to as Min_Per_Norm_Obj_ Dist, Med_Norm_Obj_Dist, and Max_Per_Norm_Obj_Dist, respectively); blink measures included: number of blinks, total blinks duration, minimum, maximum and mean of blinks length based on raw data and according to the previous suggested constraints (Pietrock et al. 2019). We also calculated the following speed measures: the difference in time of (1) the gaze shift distance and (2) the Euclidian distance between the participant's head position and the gazed object position (for each, the percentiles (10, 50, 95) were calculated and referred to as Min_Per_Diff_Dist, Med_diff_ Dist, Max_Per_Diff_Dist, Min_Per_Diff_Obj_Dist, Med_ Diff_Obj_Dist, and Max_Per_Diff_Obj_Dist, respectively). For each feature, the mean value for the two runs was used for further analysis. We further extracted features that represented the scan speed, as saccade velocity was found to be related to negative emotion expression (Susskind et al. 2008). Scan speed percentiles (10, 50, 95) were defined as the gaze scan Euclidian distance per time unit (referred to as Min_Per_Scan_Speed, Med_Scan_Speed, and Max_Per_ Scan_Speed, respectively).
An additional measure for object size was tested as a possible confound for the extracted ET features. This measure was the circumference of the bounding box that surrounded each object. This measure was referred to as DIM_SUM.

Like very much Medium like Do not like
Prefere nce a b Fig. 3 Ranking session. An example of a preference ranking trial. a The ranked object, a cat in this trial, first appeared by itself in the "yard" for two seconds. b After two seconds since it was first viewed, the ranking scale appeared, along with its bounds and definition beneath it (the text boxes that were used in the task were in Hebrew). When the participants' controller was pointed at a chosen region on the ranking scale it changed its color to green

Feature selection
The relation of all the above-detailed extracted ET features to subjective experience measures was assessed using the results of a mixed linear regression model. The ranking value was the dependent variable, and the independent variables were the ET features adjusted for the object size and interacted by the viewed object type (pets vs. shapes), where participants were considered as random effects.
Since preferences were our main interest, we focused on the preference ranking regression models' results for feature selection. To include the possibilities of both similar and different effects according to object type, the selection criteria of features accounted for the fixed effect as well as the interaction terms in the mixed linear models. This was done by choosing the features with the smallest p-value in each of these effects.
During the experiment, the stimuli were presented in random order. This resulted with diversity in participants' body and head movements prior to and during their gaze of stimuli. Consequently, the gaze ray origin and direction differed slightly from one participant to the other which affected stimuli brightness accordingly. It had been previously shown that brightness (Mathôt 2018) affects pupil diameter measures and thus we removed pupil diameter features from further analysis (nevertheless, their relation to rankings is presented in Tables 4, 5, 6 in the Supplementary  Material). Indeed, technically it is possible to control for the brightness of all objects and pixels in the scene and allow for the inclusion of pupil features in the analysis as well.
However, these adjustments are likely to decrease the sense of depth and the ecological nature of the task. Since these naturalistic components (depth and ecological relevance) were important characteristics in the current task design, we chose not to adopt the technical scene definitions and thus could not use the pupillometry measures. Nevertheless, as an assessment of the possible effect of brightness on our results, we calculated the average brightness of screenshot images of the stimuli in the scene as viewed in the task. We performed an additional exploratory analysis of the same feature selection calculated similarly as described above, while adding the brightness as an additional explanatory variable to the ranking mixed linear models. We found it did not alter the feature selection process (for details see section S7 in Supplementary Material).

Selected features construction
When we ordered the extracted features in an ascending order of p-value of fixed effect term, the Med_Scan_Speed was the first from the top (i.e., the lowest p-value), the Med_ Dist feature was the second from the top (i.e., the 2nd lowest p-value), and Min_Per_Scan_Speed was the third from top (i.e., the 3rd lowest p-value). When we ordered the extracted features in an ascending order of p-value of the interaction term, the Med_Scan_Speed was the first from the top, the Max_Per_Dist was the second from the top, and Med_Dist was third from the top. The union of the top three according to these two criteria included Med_Scan_Speed, Med_Dist, Min_Per_Scan_Speed, and Max_Per_Dist. However, since  Fig. 4 Eye-tracking features construction scheme. a The location of eye gaze hit points on the cat are represented by red dots with the notation h i , h i+1 , and h i+2 for the i th, i+1, and i+2, three timeconsecutive ET data samples respectively. b The center of the cat's bounding box denoted as o is marked by a yellow circle with a black cross in the middle, and the 3D Euclidean distances between that center and each of the gaze hit points h i , h i+1 , and h i+2 , are marked with black arrows and notated by d i , d i+1 , and d i+2 , respectively (see Eq. 1 in Feature extraction). The Med_Dist feature is the median of these measures (see Eq. 2 in Feature extraction). c The 3D Euclidean distances between pairs of each two time consecutive gazed hit points as h i and h i+1 for the i th and the i +1 time points, and h i+1 and h i+2 for the i +1 and the i +2 time points, are the gaze distances per time unit, and represent the scan speed in the transition from h i to h i+1 and from h i+1 to h i+2 , are marked by blue arrows with the notation s i and s i+1 , respectively (see Eq. 3 in Feature extraction). The Med_Scan_Speed feature is the median of these measures (see Eq. 4 in Feature extraction). **Images captured from the task are published with consent from the assets' designers. (https:// asset store. unity. com/ 3d? categ ory= 3d& versi on= 2018&q= ArchV izPRO% 20Int erior% 20Vol. 4& order By=6, https:// asset store. unity. com/ packa ges/ 3d/ chara cters/ anima ls/ mamma ls/ cute-pet-96976) Min_Per_Scan_Speed is from the same group of features as Med_Scan_Speed, but less significant, we focused on Med_Scan_Speed out of these two. Similarly, since Max_ Per_Dist is from the same group of features as Med_Dist but less significant, we focused on Med_Dist out of these two. Therefore, these two features were the focus of continued analyses according to the feature selection criteria.

Med_Dist
The Med_Dist feature was calculated based on the gaze shift distance, noted as d i for the i th time point, which we defined as the 3D Euclidian distance between the gazed hit point h i for the i th time point and the centroid of the gazed object o (see Fig. 4b), where: We considered the center of the objects bounding box as their 3D centroids denoted as o . The 50 th percentile of the gaze shift distances obtained in each trial was denoted as the median gaze distance, and shortly as Med_Dist:

Med_Scan_Speed
Scan speed was defined as the gaze scan distance per time unit. Therefore, this feature was based on the distance, noted as s i , of the location of two time-consecutive gazed hit points h i and h i+1 for the i th and the i +1 time points (see Fig. 4c), such that: Since s i was calculated per each two time-consecutive ET data samples, the corresponding time difference for all time points equaled 8.3 ms based on the 120 Hz ET sampling frequency, resulting in a scaled measure for speed. The 50 th percentile of the scan speed per trial is referred to as Med_Scan_Speed: Time samples for which d i was greater than the maximal Euclidean distance obtainable for the object were considered measurement errors of the eye-tracker and excluded from further analysis. (1) (4) Med_Scan_Speed trial = Median s i i∈trial

Gazed time
In addition to these gaze behavior features, we used the GazedTime feature to examine trials and participants' data validity as described in the Eye-tracking data pre-processing section, as well as for considering the possibility of the mere exposure effect on preferences (Zajonc 1968(Zajonc , 2001Zajonc and Markus 1982).
The GazedTime was defined as the percentage of time the object was gazed at out of the total time it was presented (2 s).

Statistical analyses
Statistical analysis was performed with lme4 (Bates et al. 2015) and visualizations with ggplot2 (Wickham 2016) R software (R Core Team 2022) packages. All ET features were standardized before the regression models were performed, and values greater than or smaller than 3 standard deviations from the mean were treated as outliers and were removed from the analysis. We used the standard p-value of 0.05 criterion for assessing the significance of each effect. For each feature, the mean value for the two runs was calculated and used for all analyses.

Pre-registration note
This study was initiated with a pilot sample of 23 valid participants. The pilot results were used to calculate the required sample size to achieve 80% power, a significant effect with p-value < = 0.05, for the Med_Dist in the model for preference ranking. Power analysis resulted in a target sample of 53 participants with valid preference ranking, which was noted in the preregistration: https:// osf. io/ h9bpq. During the main study sample collection (of the n = 53), three participants were excluded from the preference ranking analysis according to the ranking session exclusion criterion. Therefore, we finished the data collection with 56 participants out of which 53 with valid preference ranking.
Following the completion of the main study data collection as planned and described in the pre-registration, we discovered a miscalculation that affected our eyetracking features, including the Med_Dist feature, which we originally hypothesized upon and calculated the sample size based on.
After the calculation correction, we discovered that even though the hypothesized effect was significant in the study sample (n = 53), it was not obtained in the pilot sample (n = 23, see Table 1 in the Supplementary Material). In light of these findings, in retrospect, the Med_Dist would not have been chosen as the measure to calculate the power analysis on, which determined the study sample size (details on feature selection in the Feature selection section). However, since both samples have been collected by that point in time and we had already peeked at the data, we combined them both for reassessment and all further analyses referred to in the following sections.

Valid data
For the final sample of 79 valid participants, 8.2% of the trials were excluded due to less than 50% of the objects gazed time, out of which 0.7% of the trials were excluded due to less than 50% valid or non-Nan values. All participants had more than 50% valid trials and hence were valid for ET data analysis.

Reaction time
During the dynamic viewing task, the duration it took participants to find the animated shapes since they first appeared in the scene was (Mean = 1.8 s, SD = 1.1 s), which was slightly but significantly longer (in 0.1 s) than the time it took to find their paired pets (Mean = 1.7 s, SD = 1.2 s) across all participants (t [23] = − 2.8, p = 0.01, linear mixed model). The time to identify the objects was defined as the participants' reaction time (RT), and its distribution according to object type is shown in Table 1 and Fig. 5a.

Gazed time
Participants were instructed to look at the appearing stimuli from the moment they were first observed until they disappeared. Nevertheless, we found that the percent gazed time was significantly longer by 5.4% for pets (M = 87.25%, SD = 12.38%) than for their paired control shapes (M = 82.37%, SD = 15.19%) across all participants (t[29.9] = 4.9, p = 2.5e-5, linear mixed model, Table 1 and Fig. 5b). Therefore, to account for possible mere exposure effect (MEE, (Zajonc 1968(Zajonc , 2001Zajonc and Markus 1982)) we added the percent of the gazed time as an additional independent variable to all subsequent ranking models.

Ranking
The ranking sessions validity criterion yielded 75, 77, and 66 participants with valid ranking data for the preference, valence, and arousal rankings respectively. As expected, the preference, valence, and arousal rankings consisted of significantly higher values for pets than for their paired control shapes (Table 1 and Fig. 5c). After Bonferroni correction for 3 tests in 3 mixed linear regressions, we found a significant difference of 24% of the scale range in the pets preference ranking compared to shapes (t[55.8] = 6.37, p = 1.14e-07); a significant difference of 14% of the scale range in the pets valence ranking compared to shapes (t[35.3] = 4, p = 8.47e−4); and a significant difference of 35% of the scale range in the pets arousal ranking compared to shapes (t[76.5] = 12.87, p = 2.3e−20). The gazed time we added to the model did not have a significant effect on any of the rankings (p > 0.4 in all models for the effect of gazed time).

Feature selection
As explained above in the Feature selection section the features were sorted in ascending order according to the fixed effect p-value and interaction p-value. Table 2, is sorted according to the fixed effect parameter and shows the Med_Scan_Speed was the first, and the Med_Dist was the second from the top.

Selected features
The median distance of gaze, referred to as Med_Dist, was significantly higher for pets relative to their paired control shapes across all participants (t[22] = 2.67, p = 0.028, mixed linear regression, adjusted for gazed time and object size, Fig. 6). This feature represents a measure for how far the participants scan the stimuli away from their center. Therefore, this difference between the pets and shapes could reflect or be influenced by the greater complexity and points of interest the pets have in regions proximal to bounds relative to the shapes. The gazed time in this model had a trending effect on the Med_Dist values (t[1774] = − 2, p = 0.08, mixed linear regression), but still, Med_Dist is significant above and beyond gazed time.
A similar relation was found for the median scan speed feature, referred to as Med_Scan_Speed, where significantly higher values for pets were obtained relative to their paired control shapes across all participants (t[54.2] = 5.47, p = 2.3e−06, mixed linear regression, adjusted for gazed time and objects size, Fig. 6). Unlike for Med_Dist, the paired difference in Med_Scan_Speed was significantly negatively influenced by the gazed time (t[21.4] = − 4.1, p = 8.6e−4, mixed linear regression). This could result from an attempt to compensate with higher scan speed for less time of viewing an object. These ET results were Bonferroni corrected for the 2 features.

Med_Dist relation to subjective rankings
We tested the relation of Med_Dist to ranking using a mixed linear model adjusted for object size and gazed time to control for the possible mere exposure effect (Zajonc 1968(Zajonc , 2001Zajonc and Markus 1982). The model included object type interaction, and the participants were modeled as random effects (Bonferroni corrected for 3*32 tests, i.e., 96). A significant interaction of Med_Dist and type was found for preference (t[106.4] = − 4.9, p = 2.6e-4, Fig. 1a in Supplementary Material) and valence and (t[222.2] = − 6.2, p = 2e−7, Fig. 1b in Supplementary Material). In addition, we found a significant positive relation between Med_Dist and preference of pets (beta = 0.18, t[70.4] = 4.6, p = 0.0016) and valence rankings of pets (beta = 0.25, t[105] = 7.4, p = 3.3e−9). This implies that the furthest away from the center of a pet the participants fixated upon, the more they liked it, and the more pleasant they felt while looking at it. This effect was not obtained for the shapes (p > 0.87 for both preference and valence). We additionally tested the Med_Dist relation to arousal rankings. We did not find any significant interactions (t[606] = 1.6, p = 1). However, a significant main effect of Med_Dist was found in relation to arousal rankings (beta = − 0.13, t[581] = − 5.5, p = 5.3e−6, Fig. 1c in Supplementary Material) suggesting that across all items the furthest away participants looked from the center the less they were aroused.
The gazed time had no effect on the rankings in these models (p = 1, for both main effect and interaction with type in all corresponding models).

Med_Scan_Speed relation to subjective rankings
The relation between Med_Scan_Speed and subjective rankings was tested using linear mixed models adjusted for object size and gazed time, including object type interactions, where participants were considered as random effects (Bonferroni corrected for 3*32 tests). A significant interaction of Med_Scan_Speed and type was found for preference (t[83] = − 5.6, p = 2.88e−5, Fig. 7a) and valence (t[75.5] = −6.6, p = 5.3e−7, Fig. 7b). In addition, we found a significant positive relation between Fig. 7 Med_Scan_Speed relation to rankings. Scatter plots of the subjective rankings as a function of Med_Scan_Speed per participant per object, and the linear models predicted values of rankings as a function of Med_Scan_Speed for each object type (pets in red and shapes in blue; fitted lines are surrounded by gray polygons of the 95% confidence interval). a Preference rankings as a function of Med_Scan_Speed. A significant interaction of Type was obtained (beta = − 0.2, t[83] = − 5.6, p = 2.88e−5), CI = (− 0.28, − 0.14)) as described by a significant positive slope only for the pets in the plot. b Valence ranking as a function of Med_Scan_Speed. A significant interaction of Type was obtained (beta = − 0.3, t[75.5] = − 6.6, p = 5.3e−7, CI = (−0.39, − 0.21)) as described by a significant positive slope only for the pets in the plot. c Arousal ranking as a function of Med_Scan_Speed. Type interaction was not found for arousal ranking (beta = 0.05, t[53.3] = 1.2, p = 1, CI = (− 0.03,0.14)) as described with two non-significant slopes for both types in the plot. (All variables are z-scored, and significance is noted for the Med_ Scan_Speed estimate in a mixed multilinear model adjusted for objects dimension and percent time gaze for each type, where p < 0.001 '***', p < 0.01 '**', p < 0.05 '*', p < 0.1 '.'; all tests are Bonferroni corrected for 3*32, i.e., ranking times the number of features = 96 tests) Med_Scan_Speed and preference of pets (beta = 0.19, t[68.3] = 6.2, p = 4.1e−6) and valence of pets (beta = 0.28, t[66.7] = 6.9, p = 2.1e−7), while no relation was obtained for the shapes type (p = 1 for both). These results suggest that the faster participants scanned the pets with their gaze, the more they liked them, and the more pleasant they felt while looking at them.
The gazed time had no effect on the rankings in these models (p = 1 for both main effect and interaction with type in all these models).

The relation of ET to the paired ranking difference
Our assumption in pairing a control shape to each pet was that they will be ranked differently, despite their similar color, size, and position in the scene. As detailed above, we found a significant positive relation between Med_Dist and Med_Scan_Speed to pets' preference and valence rankings accompanied by a significant interaction with shapes, and a significant paired difference of these rankings and ET measures. Therefore, even though these ET features had no effect on the shapes' valence and preference rankings, we further tested whether the paired difference in the ET measures is related to the paired ranking difference for each pair of pet and its control shape. We tested this with a mixed linear regression fitted to the paired ranking difference, this time in a combined model with both the paired difference in Med_Dist and Med_Scan_Speed as independent variables adjusted for the paired difference in size and gazed time. After Bonferroni correction for the two rankings of preference and valence, we found that the paired difference in the Med_Scan_Speed was positively related to the paired difference in preference ranking (beta = 0.12, t[15.3] = 3.6, p = 0.005) and valence rankings (beta = 0.14, t[15.4] = 3.8, p = 0.003). Nevertheless, the paired difference in Med_Dist was not correlated with the paired difference in ranking (t[16.7] = −1.2, p = 0.5) and (t[17.5] = −1.3, p = 0.4) for preference and valence ranking, respectively. No relation was found for both paired differences in gazed time and objects object size to the paired ranking difference (p > 0.27 for all). These results suggest that the higher the difference in the characteristic Med_Scan_Speed of pets relative to their control shapes, the greater the subjective difference in their preference and in the valence that participants felt while looking at them.

Discussion
Virtual reality has been gaining popularity in research (Blascovich et al. 2002;Peck et al. 2013;Ossmy and Mukamel 2017;Freeman et al. 2017Freeman et al. , 2019Reggente et al. 2018;Areces et al. 2019;Huang et al. 2021) and therapeutics (Dahlquist et al. 2010;Cesa et al. 2013;Herrero et al. 2014;Jeffs et al. 2014;Chirico et al. 2016;Maples-Keller et al. 2017;Sacks and Axelrod 2020) due to its immersive gamified nature. Simple personalization of presented content is highly needed to increase its effectiveness. In this study, we aimed to address this personalization challenge by using eyetracking of moving individual objects. We devised a novel task whereby we presented participants with individual moving objects and analyzed eye gaze data. We found that the larger the scanned distance from the center of an object (named Med_Dist) and the scan speed (Med_Scan_Speed) of that object, the greater the participants' preference for this object. The gazing on individual items and ranking of items were performed in separate phases of the experiment. This suggests that gaze tracking while passively viewing items can serve as a marker for preference without the need to probe it directly. We compared two stimuli categories of pets and shapes, controlling for color, size, and scene position of the pets compared to shapes. We tested multiple gaze behavior features, and following a feature selection process, we focused on two features, median distance and median scan speed. We found a positive correlation between these ET features and the subjective rankings of preferences as well as the valence of viewed pets. These findings describe a general pattern of gaze behavior within an object in a 3D HMD VR environment. These results suggest that ET measures extracted during passive viewing of individual moving objects during a VR experience could be used as biomarkers of subjective preference and pleasantness. Moreover, when we performed these analyses while considering a reduced amount of gazed time of 1 s and even 0.5 s (for details see section S6 in Supplementary Material), significant results were perceived for the selected features of Med_Dist and Med_Scan_Speed. Keeping in mind that fixation duration is between 180 and 500 ms (Rayner 2009;Balcombe et al. 2015) this suggests that only 1-2 fixations might be required for preference detection. This threshold seems reasonably obtainable even in the case of simultaneous presentation of multiple objects of interest in the scene, in a more general design than performed in this study. Our results call for follow-up studies including multiple objects.
Findings relating gaze and preferences have been studied in tasks with 2D images of faces and shapes (Shimojo et al. 2003), snacks (Krajbich et al. 2010;Krajbich and Rangel 2011) or food items (Graham and Jeffery 2012). During these tasks participants were requested to choose the one they preferred out of two (Shimojo et al. 2003;Krajbich et al. 2010) or three (Krajbich and Rangel 2011) presented items. Alternatively, they were requested to indicate if they would buy an item in a simulated food purchase (Graham and Jeffery 2012). The common result in these studies was that the accumulated amount of time gazed, also referred to as dwell time, was positively related to preference or choice (Shimojo et al. 2003;Krajbich et al. 2010;Krajbich and Rangel 2011;Graham and Jeffery 2012). In addition, with the development of an extended version of the drift diffusion model (DDM), a more complex relationship was found between fixation patterns and choices, as well as several fixation-driven decision biases in binary (Krajbich et al. 2010) and trinary (Krajbich and Rangel 2011) value-based choice tasks. Additional ET features in these studies were the item's identity and the dwell time at the first and last time points (Krajbich et al. 2010;Krajbich and Rangel 2011). Thus, while prior work examined gaze and preferences during active choice, in the current study, we set out to investigate passive gaze time series patterns of individual items with the purpose of detecting subjective experience without choice or valuation elicitation. Therefore, our task was designed to study passive viewing we often encounter in daily life, which includes data gathering on objects which are gazed at one at a time, without necessarily applying a subsequent choice or action.
Previous studies employing 2D stimuli found that the amount of time an object was gazed at influenced preference (the mere exposure effect, MEE, (Zajonc 1968(Zajonc , 2001Zajonc and Markus 1982;Shimojo et al. 2003). In the attempt to control for this potential confound, all participants were instructed to look at each stimulus from the moment it was first noticed until it disappeared. Still, a small but significant difference was obtained between the percent time the pets were gazed at, relative to their control shapes. To account for MEE, we added this variable to our models and found it had no relation to preference or any of the other ranking measures.
The positive correlation between both Med_Dist and Med_Scan_Speed features to subjective rankings was demonstrated from multiple perspectives. These two ET features obtained higher values for the pets relative to their control shapes. Similarly, the pets were also ranked higher than shapes for the measures of preference, valence, and arousal. A possible factor accounting for the general difference in ET features between the two object types could have been influenced by the greater complexity and salient details of the pets such as eyes, mouth, ears, and tail, whereas the shapes lacked these properties. Participants could have been attracted to observe these features that are close to the pets' perimeter, and thus derive their greater values of Med_Dist. Furthermore, a possible attempt to capture the additional details in the pets could have led to their obtained higher Med_Scan_Speed. This idea is in line with the negative relation found between Med_Scan_Speed and percent gazed time, which could point to an attempt to compensate for the shorter time gazed while collecting the same amount of data on the viewed object. We additionally found a positive relation between the difference in Med_Scan_Speed of pairs of pets and their control shapes and the corresponding difference in their preference and valence rankings. The percent gazed time had no effect on this paired difference in valence and preference rankings. Taken together, these results suggest that the gazed time relation to Med_Scan_ Speed describes a difference between the categories that is unrelated to the difference in their liking and pleasantness.
A positive relation of these ET features to preference and valence rankings was found within the pets' category alone, suggesting a characteristic difference in gaze patterns when looking at shapes vs. pets. A possible explanation for this interaction with the object category is the greater number of salient details in pets, that could have contributed to a greater diversity in gaze behavior in this category compared to their control simple shapes. In a study that classified images by modeling scan paths with variational hidden Markov models (HMMs), a positive correlation was obtained between the classification accuracy and the amount of salient objects in the image (Coutrot et al. 2018). This suggests that objects' complexity allows for better inference of their content based on gaze behavior. Taken together with the findings in our study, it seems that the greater the complexity or salient items 2D or 3D objects have, the greater the ability to deduce their content, preference, and possibly other traits, based on ET data.
Object size was previously tested but not found to be related to affective preferences when it was examined in a controlled fashion (Seamon et al. 1997). Other studies found that preference for different stimuli categories was affected by different visual attributes, although size was not tested . Therefore, since both the gaze distance (Med_Dist) from the center of the object and the Med_Scan_Speed of the object could have been influenced by the object size, we included the objects' circumference as a measure for its size to all our mixed linear models for rankings prediction. We found that both ET features of Med_Dist and Med_Scan_Speed were positively related to pets' preference and valence ranking adjusted for stimuli size. This result suggests that extended gaze at peripheral points and higher scan speed characterize the exploration pattern gaze at pets, regardless of their size. Furthermore, this gaze pattern indicates higher subjective preference or valence. The pets in this study may be considered as an example of complex objects relative to their rather simple control shapes. This suggests that our results could be extrapolated to more complex and emotional objects such as other animals, or human characters compared to other simple objects such as simple furniture. Future studies will allow exploring such generalization of our results among others.
In contrary to the positive relation found for the Med_Dist feature to pets' preference and valence rankings, a negative relation was obtained for both the pets and shapes arousal rankings. This result was initially surprising. Therefore, we examined in depth the identity of the objects that were mostly ranked to induce high arousal levels. We found that they were all also ranked with low valence (and low preference for the pets in this group). Hence, high arousal stimuli generated the opposite trend for valence, and preference in the pets group (for details see section S5 in Supplementary Material). Related work, aimed to investigate the relation between ET features and emotional states, was done on 2D images using the International Affective Picture System (IAPS) (Lanatà et al. 2013;R-Tavakoli et al. 2015;Aracena et al. 2015), which are standardized emotion inducing images (Lang et al. 1997). Those studies aimed to classify one dimension of affect as arousal levels (Lanatà et al. 2013), valence (R-Tavakoli et al. 2015), or positivity level (Aracena et al. 2015) and used a subsample of the IAPS images accordingly. In contrast, we studied the relation of ET features to two affect dimensions on 3D objects. The affect rankings in studies that used the IAPS over the years are robust compared to introducing and ranking stimuli for the first time as done in our study, which is one of its drawbacks. We used novice stimuli as publicly available affective stimuli for 3D VR are yet rather scarce. We found only two datasets of emotional 360° videos Tabbaa et al. 2021) which are not necessarily suitable for being embedded in many task designs. It seems that standardizing 3D emotional objects to be used in VR studies of affect could significantly contribute to the field as the IAPS has been for 2D images. Our study included only two object types of pets and shapes as our design aimed to keep the task at a reasonable time frame for wearing the VR headset. Thus, future work will be required to generalize our findings to other types of objects, environments, and scenarios, including stimuli with various affective values including human characters.
The pastoral scene that was chosen for the task was aimed to enable a spectrum of preferences and affective states of valence and arousal. Since we did not set to study fear, threat, anxiety, or other stressful states, we utilized the gamified immersive nature VR has and induced a general pleasant experience for the participants. We indeed noticed that in total the participants had a pleasant experience. It seems that the neutrality we aimed for was achieved considering that the stimuli obtained wide distributions for all ranking measures (except for the valence of several shapes, see Figs. 2,3,4,5,6,7 in the Supplementary Material). Some stimuli received ranks in the lower half of the scale, suggesting we accomplished elicitation of a spectrum of experience measures. Even though the neutrality of this environment showed its suitability for various stimuli, future studies will allow generalization of our results to other experiences, scenarios, and stimuli.
Most ET studies thus far were done on sedentary tasks and in front of a 2D screen (Lappi 2016) sometimes with the addition of a chinrest, to ensure all participants look at the screen from the same point of view. However, in many natural gazing tasks we encounter in our daily lives, we look at objects from different points of view and even during movement. In our study, even though all stimuli appeared at the same locations for all participants since they appeared in a random and therefore different order they were observed from possibly slightly different angles, points of view and perspectives due to participants' bodies, heads, and eye movements before the stimuli being observed. As a result, participants might have not all looked at the same coordinates on all stimuli. Thus, extracting global features for spatial dynamics of the eye scan of objects as we performed, is compatible and relevant to real-world viewing conditions more than a coordinate-wise fixation location analysis usually done in 2D studies. This naturally adds variance to the data, it also allows experimenting in a naturalistic fashion. Therefore, this current study provides novel analysis and insights into gaze behavior in a naturalistic engaging environment. Hence, it could also apply to studies with eye-tracker built-in glasses as it allows free head and body movements as in VR systems.
Passively identifying the individual preference or pleasantness of various objects or figures in a computerized scene, simulation, or training session, will allow real-time updating of the content presented in the direction of the preferred content and will increase the participant enjoyment and engagement. These were previously shown to enhance learning in serious games (Ravyse et al. 2017). Therefore, using the methodology presented in the current work may contribute to the optimization of computerized learning or training tasks to enhance learning. In an unpublished study, we established a closed-loop system for this task in our laboratory that identifies the individual preference of stimuli based on the selected ET features and updated the task in real-time accordingly. In one application, the size of the following presented stimulus was changed according to the preference of its previous one. In another application, a group of stimuli was presented according to the preference inference of an initial group. When the entire experience was ranked, greater enjoyment was found for participants who were presented with stimuli with higher preference, than those with lower (data not shown). This may apply to different regimes such as elementary or high school and other education systems; skills training tasks as for surgeons virtual training; factory employee machinery training; different private training software for language, programming, talking in front of an audience; or any knowledge transferred via software in a 2D screen or VR headset format. These principles may also be adopted for gaming, where enhancing individual enjoyment and engagement, may virtually contribute to improving the game. In addition, preference identification may be used for the personalization of in-game purchases, or even advertising during gaming or any other content viewing.
For therapeutic purposes, an example could be the attempt to achieve a target positive emotional state in treatment of depression, and anxiety, or for example, as a distraction when removing bandages after burn injuries (Jeffs et al. 2014). Personalization of the presented virtual content could be done according to the identification of the individual preference and feeling of pleasantness of the presented objects, surrounding people, scenery design, trees, etc. and increase the therapy's impact. In addition, user preference calculation in the framework of psychotherapy software may be used to evaluate the progress and effect of psychotherapy during the execution of the software, or over multiple different executions/sessions, by measuring the difference in user preference for the presented objects throughout the session(s). These are just a few examples of utilizing our findings in future real-time updated applications, that are likely to grow with the technological and scientific developments in the coming years.
To conclude, the concept of VR content updating based on subjective internal states has been studied using physiological signals (EEG and ECG (Marín-Morales et al. 2018), respiration, ECG, EMG, and EDA (Bermudez I Badia et al. 2019)). However, these necessitated additional equipment and expertise. Here, we show that using a built-in eye-tracker and unique analyses we could identify preferences and valence. Finally, the use of ET in human-computer interfaces has been noted as a promising approach either alone or combined with inputs from other sensors or devices (Jacob and Karn 2003). The utilization of the immersive naturalistic nature of VR is increasing in various research fields and applications. Therefore, our findings suggest that ET measures could passively reflect an individual internal preference and state. In turn, these could be used for content personalization, either with or without additional physiological signals, via real-time updating of VR tasks for research, learning, or therapeutic purposes.