Humans-as-a-sensor for buildings: Intensive longitudinal indoor comfort models

Prageeth Jayathissa, Matias Quintana, Mahmoud Abdelrahman, and Clayton Miller ∗

Building and Urban Data Science (BUDS) Lab, National University of Singapore (NUS), Singapore

∗Corresponding Author: clayton@nus.edus.sg, +65 81602452

Abstract

Evaluating and optimising human comfort within the built environment is challenging due to the large number of physiological,psychological and environmental variables that affect occupant comfort preference. Human perception could be helpful to capturethese disparate phenomena and interpreting their impact; the challenge is collecting spatially and temporally diverse subjectivefeedback in a scalable way. This paper presents a methodology to collect intensive longitudinal subjective feedback of comfort-based preference using micro ecological momentary assessments on a smartwatch platform. An experiment with 30 occupants overtwo weeks produced 4,378 field-based surveys for thermal, noise, and acoustic preference. The occupants and the spaces in whichthey left feedback were then clustered according to these preference tendencies. These groups were used to create different featuresets with combinations of environmental and physiological variables, for use in a multi-class classification task. These classificationmodels were trained on a feature set that was developed from time-series attributes, environmental and near-body sensors, heartrate, and the historical preferences of both the individual and the comfort group assigned. The most accurate model had multi-classclassification F1 micro scores of 64%, 80% and 86% for thermal, light, and noise preference, respectively. The discussion outlineshow these models can enhance comfort preference prediction when supplementing data from installed sensors. The approachpresented prompts reflection on how the building analysis community evaluates, controls, and designs indoor environments throughbalancing the measurement of variables with strategically asking for occupant preferences in an intensive longitudinal way.

Keywords:Indoor environmental quality, Thermal comfort models, Personalised comfort model, Machine learning, Ecological momentaryassessment, Occupant-centric, Occupant behaviour

1. Introduction

Many office workers are familiar with the battle of the ther-mostat, or that co-worker who talks loudly on the phone. Manyresearchers in indoor comfort are also aware of the high ratesof discomfort amongst office workers [1, 2]. Vast global effortshave been undertaken to evaluate this discomfort, and with thatknowledge, build models that can be used for the design andcontrol of buildings. In the realm of thermal comfort, for exam-ple, two dominant models are in use. The first is the PredictedMean Vote (PMV) that models comfort based on heat transfercharacteristics between the human and their surrounding envi-ronment [3]. The other, more modern version, is the AdaptiveComfort model that includes the human adaptability to climate,drawing a linear relationship between the indoor and outdoorenvironments [4].

The underlying issue with modelling human comfort is thesheer number of variables present and the difficulty in accu-rately measuring them. Figure 1 highlights this issue by de-tailing a list of studied physiological, psychological, and en-vironmental variables that influence thermal, visual, and auralcomfort. While the empirical models in the academic literatureare capable of incorporating a handful of these variables, theexclusion of the rest can cause significant errors. One reasonis that the interrelationship between different indoor environ-

mental parameters is not well-known [5]. It was shown in arecent study that the lowest indoor environmental satisfactionfactor drives the overall satisfaction [6]. For example, whileone can measure the temperature and humidity of a room; thetype of meal a person ate, and even the spices present in themeal, can put the human body in a different state of thermalperception [7, 8]. Furthermore, most of the studies that dependon measuring environmental variables using mobile carts withmounted sensors [9, 10, 11] or low-cost continuous sensing sen-sors [12] face problems related to the accuracy and calibration[13]. While being comprehensive in capturing most comfort-related factors, Figure 1 excludes literature about physical andmental ailments, which further adds variance to the models.

It is, therefore, not surprising to find that preference pre-diction models with only a handful of the factors in Figure 1have low accuracy. For example, the previously mentionedPMV model uses personal and environmental parameterssuch as temperature, humidity, mean radiant temperature,air movement, and clothing and metabolism levels to predictthermal comfort. A recent analysis showed that this modelis only accurate 34% of the time [50]. In the control of realbuildings, these models are further simplified, and it is usuallythe only variables of temperature, illuminance, and noise levelsthat are used to evaluate thermal, visual, and aural comfort,

Preprint submitted to Buildings August 26, 2020

arX

iv:2

007.

0201

4v2

[cs

.HC

] 2

5 A

ug 2

020

Air Temperature

Solar Radiation

Environmental Long Wave Radiation

Clothing Level

Air Flow

Near Body HeatSources

Perspiration

Thermal Regulation

Diffuse

Air Humidity

Adaptation toOutdoor

Environment

Base Metabolic Rate

Diet-Induced Thermogenesis

Posture

Age

Vascular AnatomySubcutaneous Fat Thickness

Circadian Clock

Physical Activity

Direct

Thermal Comfort Visual Comfort

Daylight Utilisation

Glare

Light - Spectrum- Uniformity- Illuminance- Flicker

Aural Comfort

Presenceof Others

Constant Noise- Amplitude- Frequency

Variable Noise- Amplitude- Frequency of Perception- Annoyance*

ControlabilityTask

ViewSubjective Sensitivity

Sound Absorption

Aliesthesia

Sound Privacy

Seniority in the CompanyCircadian Reset

MigraineSusceptibilityDaylight Perception

Biophilia

Gender

Figure 1: Graphical review of physiological, psychological and environmental factors influencing human comfort. Thermal - clockwise from top left: adaptation tooutdoor environment [4], air flow [14], solar radiation [15], circadian rhythm [16], daylight perception [17], environmental long wave radiation [18], perspiration[19], diet-induced thermo-genesis [8], subcutaneous fat thickness [20], posture [21], temperature/humidity [3] [22], physical activity [23], clothing level[24], vascular anatomy [25], near body heat sources [26], gender [27], age [28], basal metabolic rate [29], thermal regulation [30] [31], alliesthesia [32].Visual - clockwise from top left: circadian calibration [33], daylight [34], view [35], spectrum [36], uni-formity [37], illuminance [38], flicker [39], susceptibility to migraines [40], biophilia [41], glare [42].Aural - clockwise from top: seniority in a company [43], subjective sensitivity [44], sound privacy [45], sound absorption [46], controllability [47], task[48], variable and constant noise [49].

respectively.

1.1. Can longitudinal human perception feedback supplementsensors?

The human nervous system detects sensation and converts itinto thoughts and feelings, which are the very foundation of theword comfort. What if occupants in buildings were asked abouttheir subjective preference in spaces, instead of only measur-ing environmental variables and using them to infer comfort?Collecting enough comfort preference feedback from a singleperson over days or weeks would take advantage of a human’sability to evaluate dozens of variables simultaneously, includ-ing those that are difficult to measure. How can this type ofmethodology be accomplished in a scalable way without an-noying occupants too much or inducing survey fatigue? Canthis approach provide insight into comfort problem areas thatcontemporary sensors are too expensive or problematic in im-plementation?

The goal of this paper was to test the ability of an intensivelongitudinal method to capture numerous environmental feed-back data from experimental participants in a field setting. Thisstudy uses Micro Ecological Momentary Assessments (EMA) asa subjective feedback methodology that overcomes many of thechallenges presented by traditional methods [51]. Micro-EMAis a method of using a smartwatch interface to prompt andcollect momentary, right-here-right-now subjective feedbackfrom a single person over several weeks [52]. Receiving a largeamount of feedback from a single person in a diversity of spacesand comfort exposures provided the ability to understand thecomfort preference tendencies of a person. It is proposed thatthese behavioural tendencies can be used to segment peopleinto groups related to how they perceive their environment.

Grouping people with similar comfort preferences could,therefore, increase the accuracy of predicting where a personwill be comfortable and what the system can to respond withoutadditional sensors. Additionally, collecting large amounts ofsubjective preference data from numerous people in a particularspace can characterise the comfort-related attributes of thatspace to supplement data being collected from the sensorsinstalled. If technically scalable and not too disruptive to anoccupant, using humans-as-a-sensor in buildings could changethe way post-occupancy evaluations, building and systemdesign, and controls and automation are done. There wouldbe opportunities for people to provide feedback for short-termepisodic uses (days or weeks), for building commissioning orlong-term (months or years), and for continuous system controland management. This work complements the momentumfrom other disciplines focused on the use of humans as sensorsfor applications in detection of events using social media data[53], for detecting emergencies [54], and for cybersecurity [55].

1.2. Paper overviewThis paper presents how high-frequency micro-EMA, com-

bined with sensor data time-series analysis, can enable the eval-uation, control, and rethinking of the design of indoor environ-ments. Section 2 first gives a more detailed overview of founda-tional work in indoor preference capture and modelling and thenovelty being proposed. Section 3 provides a comprehensiveexplanation of the design and deployment of a smart watch-based subjective preference data collection and environmentalvariable measurement system. Section 4 details the results froma field-based implementation at [removed for anonymization]and the testing of various preference models based on intensivelongitudinal data. Finally, Section 5 and 6 discusses integration

2

Home Screen

ThermalFeedback

Aural Feedback

VisualFeedback

Figure 2: The cozie watch-face, built on the Fitbit smartwatch platform was used to collect subjective feedback. The phone that is paired with the Fitbit can be usedto set up additional questions.

methods in buildings, limitations, future work, and details onhow to reproduce the study using open data and code.

2. Background and novelty

This work builds upon previous literature focused on themeasurement of factors that may influence thermal, visual,and aural comfort in the built environment. These modellingtechniques are converged with an intensive longitudinal expe-rience sampling technique that is common in the medical andpsychological communities, but only emerging in the analysisof buildings. This section covers previous work in the buildingcontext using intensive longitudinal data and an overviewof the novelty of the work in this paper as compared to theliterature.

2.1. Indoor environmental comfort variables and modelsThere are generally two models types used in the literature

for indoor comfort assessment: 1) objective-subjective, and 2)objective-criteria [56]. Which method to use is decided basedon the aim of the evaluation. On the one hand, the objective-subjective model combines the indoor environmental measure-ments from sensors with the subjective feedback from users,mostly in the form of post-occupancy evaluation (POE) sur-veys [57, 58, 59, 60]. On the other hand, the objective-criteriamodel is used in ranking or rating a building by comparing theindoor measurements from IEQ sensors with building perfor-mance measurement protocols such as LEED or WELL certifi-cations [11]. Both of the methods have drawbacks both in mea-suring the environmental data as well as surveying occupants[56].

In terms of environmental measurements, work has beendone that used accurate sensors that were mounted on movablecarts [61, 11]. However, these sensors were not affordable toall building operations scenarios [56]. The affordability chal-lenge was met using low-cost continuous sensing sensors thatrequired frequent calibration [12]. Nevertheless, the locationof these sensors in buildings, and interpolation of the readingsstill represented a challenge in the literature, given the fact thatindoor spaces are heterogeneous [62]. On the other side of the

spectrum, surveys pose some problems related to questions, e.g.what to ask, whom to ask, and how to interpret the results [56].Additionally, Porter et al. [63] discussed the term survey fatiguein which users feel overwhelmed by questions that may lead toa misrepresentation in responses and reduced response rates.

A related area of recent focus is the use of wearable andinfrared radiation sensors to capture near-body physiologicaldata that define the environmental conditions close to or atthe skin surface of an occupant. A recent study focused oncreating personalised comfort models from these data in thecontext of field-based deployment on 14 subjects [64]. Thisdeployment and the models produced used wrist and ankle skintemperature from several sensors placed on the participantsand a smartphone application to collect surveys. Further workin the indoor context showed that both wearable sensors andinfrared radiation cameras led to a 3-4% increase in accuracyof thermal comfort sensation prediction, marginally justifyingthe cost of implementation in a field setting [65].

2.2. Ecological momentary assessments (EMA)

The next area of background focuses on the challenge ofcollecting large amounts of longitudinal data from a person.Many fields of study have relied upon the ecological momen-tary assessment [51] methodology to meet this challenge. Thismethod is a type of intensive longitudinal experience samplingmost often utilised in studying human behaviour. The wordecological describes that fact that the measurement is takenin the subjects’ natural environment without impacting theirtask at hand. The word momentary pertains to the fact thatfeedback is requested at the moment of experience, as opposedto asking a subject to recall a past experience. And finally,the assessments are not static one-off outcomes but occurover time, thus accounting for temporal dynamics. Traditionalmodels found in literature such as surveys are insufficient astheir sampling rates are low, require the occupant to completelystop their task at hand to focus on the survey, and in manycases, ask for a recollection of past experiences. There is thefurther issue of survey fatigue [63] and even when willingto participate, there is a concern about how accurate theirresponses are [66]. The use of a smartwatch for data collection,

3

coined micro ecological momentary assessments, has beenshown to be so user-friendly that it does not significantlydisrupt any ongoing activity [52]. Furthermore, an eight-foldincrease in sampling frequency can be obtained, in comparisonto smartphone use, without burdening the user. Recent workhas used ecological momentary assessments to assess the builtenvironment through the use of smartphones [67]. While suchapplications are a step in the right direction, they were onlyable to collect eight feedback points per occupant, which isinsufficient for time-series analysis.

2.3. Similar work in intensive longitudinal data collection inthe built environment

Intensive longitudinal methodologies have begun to emergeas a way to characterise occupants for various built environ-ment objectives. In the urban context, several studies havedeployed sensors on people to understand their experiencesacross their daily lives. A large study based in Singaporeused thousands of wearable sensors in populations of studentsto discover travel patterns [68], collect information aboutthermal parameters [69], and even infer the impact of publicspaces on happiness [70]. Work has been done in a controlledoutdoor field study to understand the impact of the urbancontext on various emotions and physiological responses ofhuman [71]. In the indoor setting, targeted work on collectinglongitudinal data for more specific purposes has also emerged.The previously mentioned wearable study focusing on thermalcomfort collected numerous data from the 14 participants overthe 2-4 week study [64]. Another recent study that deployeda cyber-physical system to collect longitudinal data in officesfocused on occupant concentration [72]. The work in thispaper is most directly related to previous work in collectinglongitudinal comfort feedback from smartphone interfaces forthe allocation of activity-based workspaces [73] and through asustainability tour in a university campus building [74].

2.4. Novelty of proposed approachDespite the momentum in field-based intensive longitudinal

methodologies, there are still several barriers to their implemen-tation in real-world settings. Not the least is the challenge ofgetting human occupants to give data for comfort surveys, in-stall applications, or wear devices. Working from this knowl-edge, the authors developed cozie, as seen in Figure 2, an open-source, smartwatch clock-face designed to conduct micro-EMAsurveys for high-frequency data collection [75]. The applica-tion is open-sourced and free to download and use on the Fitbitgallery 1.

The innovations outlined in this work as compared to the pre-viously mentioned studies are:

• The hardware and software deployment methodology havea focus on practicality in scalable, field-based implementa-tions. Experimental participants were only asked to wear a

1https://cozie.app/

single smartwatch device and answer survey questions thatutilise a relatively small amount of time. The focus was ontesting a configuration that was easily applied in a real-world context. The modelling methodology was designedto capture as much signal as possible in the field settingwithout the ability to control and verify sensor proximityand accuracy consistently.

• A series of pre-processing steps were developed to convertintensive longitudinal data into model input features thatcharacterise the tendency of groups of people to have simi-lar comfort preferences, sometimes independent of the ob-jective environmental factors such as temperature. A sim-ple example of this concept is the commonly discussed, yetoften anecdotal, person who seems always to need morecooling, even when the temperature is already low rela-tive to the comfort zone. In this study, clustering was usedto group people into comfort preference types as an inputfeature to a preference prediction model.

• This paper introduces and tests a simple form of a coldstart variant to the preference models that could be usedto predict an occupant’s preference with limited or no dataabout their preference history in a particular space or ac-cording to particular objective measurements such as tem-perature, humidity, or other factors. This model enablesthe deployment of the cozie data collection methodologyby a set of participants in a building and then the creationof prediction models that could accommodate future occu-pants regardless of whether they have worn a smartwatchin those spaces.

• The process seeks to show that comfort-based preferenceprediction can be accurate even in the absence of environ-mental sensors if enough intensive longitudinal data hasbeen collected from enough occupants. The context of thisexperiment was in relatively uncontrolled, field-based set-tings as opposed to laboratory conditions.

3. Methodology

To collect intensive longitudinal data in a field setting, thecozie platform was built on the Fitbit smart watch2 and varioustime-series database technologies. In this section, the details ofthis technology stack are explained in the context of a deploy-ment on 30 test participants in buildings at the National Uni-versity of Singapore (NUS) in the School of Design and En-vironment (SDE). The definition of an occupant in this studywas a test participant who wore the smartwatch, and a manageras the person who coordinated the study. Thirty participantswere recruited via an online form, were compared to the inclu-sion criteria for the study, and were on-boarded according toan approved ethics review application. Priority was given toparticipants who work full time in the SDE-related buildings

2https://www.fitbit.com/

4

Tier 2: smart-watch + localisation

Tier 3: smart-watch + localisation + environmental sensors

Tier1b: smart-watch + near body sensor

Bluetooth BasedIndoor Localisation

Steerpath Beacons

- Thermal Feedback- Visual Feedback- Noise Feedback- GPS- Heart Rate

Near BodyTemperature

SENSing Air Sense Environmental Sensors- Temperature- Humidity- Noise Level- Illuminance

Influx Time-SeriesCloud Database

Tier 1: smart-watch

Fitbit

mbientsensor

Physical Proximity

Bluetooth

Bluetooth

Bluet

ooth

WiFi

WiF

i /4G

Fitbit App

Yak App

Metabase App

Figure 3: Overview of the experimental deployment in the SDE buildings in four distinct tiers: Tier 1 is the base methodology which is production-ready forreal-building deployment. It requires a smart watch with the cozie clock-face installed. Tier 1b, is an extension to the base methodology by adding a temperaturesensor to the watch. Tier 2 includes building-wide indoor localisation. In this experiment, Steerpath Bluetooth beacons were used, which communicate with theoccupant’s smartphone to determine the occupant’s location. Tier 3 merges the localised feedback points with environmental sensors in the same comfort zone asthe occupant.

on campus, and they were selected to maintain an even genderdistribution.

The technology used in the deployment of this study can besub-sectioned into individual tiers as described in Figure 3, witheach level requiring additional resources to implement. For theexperiment in SDE4, all tiers were incorporated.

3.1. Tier 1: Smartwatch for micro-EMATier 1 is the core methodology presented in this paper, which

uses the cozie clock-face, as shown in Figure 2. The occupantswere asked to wear a Fitbit Versa smartwatch during daytimehours while on the NUS campus at the very least but were alsowelcome to wear the device for the entire duration of the study.Participants were asked to leave momentary assessment feed-back on their comfort preferences at different points throughoutthe day on the watch face of the Fitbit device. Each time theyresponded to the survey, they were asked about their thermal,visual and aural preference using the options found in Figure2. Comfort preference was chosen as the feedback most ap-plicable to the methodology due to a three-point scale that ismost appropriate for frequent watch-based surveys. Preferencesurveys also provide more meaningful information by indicat-ing how the occupant would want the environment to changeas opposed to satisfaction or sensation survey types that onlycapture how the occupant feels. The participants were asked toanswer the questions when they moved from one environmentto another, which amounted to approximately 5-15 assessments

per day. The smartwatch also prompted the occupants with asmall vibration that requested feedback from them at differenttimed points in the day. This prompt only occurred during day-time hours when the subject was active. The momentary as-sessment took less than 15 seconds to complete. Throughoutthe experiment, the cumulative amount of time spent answeringthe momentary assessment was approximately 20-40 minutes.

Detailed documentation for using cozie, along with thesource code to an open-source repository, can be found on aGitHub repository3. The platform also can collect sensation,satisfaction and objective feedback such as clothing and activ-ity levels. These features were added after the experiments out-lined in this paper and were not used in this study.

3.2. Tier 2: Indoor localisation

Tier 1 is likely sufficient for experiments conducted in smalloffice spaces. If only a few different zones exist, then an occu-pant’s location could be quickly determined through a supple-mentary question in the question flow of the survey. However,in a large building, such as the SDE4 building where the out-lined experiment was conducted, a more sophisticated indoorlocalisation system was required. The SDE4 building has sixdifferent floors, a gross floor area of around 8,500 square me-ters, and a large variety of different indoor environments. To

3https://github.com/buds-lab/cozie

5

determine an occupant’s location in a building, 100 Bluetoothbeacons and the Steerpath4 platform were installed throughoutthe building. These beacons communicated with a custom-builtsmartphone application, called the Yak App [76], to determinetheir location with a one-meter precision. The location data wasthen used to geo-fence the occupant within various zones of thebuilding and was merged with the subjective preference feed-back data in the cloud.

3.3. Tier 3: Preference data convergence with environmentalsensors

Tier 3 included the deployment of 45 indoor and outdoorenvironmental quality (IEQ) sensors in the experimental con-text. This data collection tier was used to compare the resultsof the subjective feedback, with existing environmental mod-els. The IEQ sensors were WiFi-connected and were deployedby the company SenSING5 as part of an installation of sensorscampus-wide. These sensor kits measured temperature, humid-ity, noise level, and illuminance. At least one sensor device wasinstalled in each zone of the building, and the data was pulledfrom an API and merged with the subjective preference data inthe cloud.

3.4. Tier 1b: Strap-mounted sensor kit

Tier 3b included a temperature sensor was used from mbientlabs6, which was attached to the watch through a custom threedimensional (3D) printed case. The design file for this casecan be found online7. The mbient device logged data locally,which was transferred to the cloud database at the end of theexperiment.

3.5. Occupant and room preference clustering

This analysis included the hypothesis that the feedback ofone of the occupants in such groups could be used to char-acterise the preferences of all group members for a particularspace or set of conditions. In this step, the preference historyof occupants was used to do a simple clustering-based segmen-tation step to group occupants according to their raw feedbackpreference tendencies. For example, occupants who more fre-quently indicated prefer cooler as compared to a no changewould be grouped. This strategy was a simplified version ofthis type of clustering as it neglects other context-based vari-ables (environmental and physiological measurements). Thischoice was made to keep the method feasible even in situationsin which other measurements are not available.

Given its widespread usage in related literature, occupant androom clustering is calculated using the k-means clustering algo-rithm with Euclidean distance, using the scikit-learn package8.The features used for clustering were the ratio of votes of each

4https://steerpath.com/5https://sensing.online/6https://mbientlab.com/7https://myhub.autodesk360.com/ue29ab3ac/g/shares/

SH919a0QTf3c32634dcfe0a71457c47296998https://scikit-learn.org/stable/

feedback class value for each subject. For example, the ratioof prefer cooler for a given participant, or room, would be cal-culated as follows: #prefer cooler votes

#total votes . This calculation is repeatedfor all types of feedback responses for thermal, light, and auralfeedback. Then, the number of clusters was chosen to matchthe number of possible responses per type of feedback, this ledto initially k = 9, but given that there were no data points withprefer louder responses, the clusters were merged into eight.

3.6. Occupant comfort preference predictionThe metric of comparison in this study was the prediction

improvement of a machine learning model using added fea-ture sets extracted from the intensive longitudinal preferencedata. This structure matches implementation-based environ-mental comfort studies outlined in the literature that showedthe predictive improvement of additional data [64, 65]. Thisapproach can be compared to more controlled, lab-based meth-ods that seek to isolate variables and individually test their in-fluence.

The prediction problem translates to predicting the right classvalue or, in this case, the preference feedback response, at thegiven feature values. A random forest classifier from the scikitlearn package was chosen to handle this comfort prediction.Random forest classifiers have been proven to have the highestaccuracy at predicting personal comfort in one previous study[77] and is one of the best performing of other recent studies[64, 65, 78]. The decision was made to focus on the implemen-tation of a single model type that has been proven effective andis straightforward to use based on documentation and ease-of-tuning. With this in mind, we fixed the hyper-parameters for therandom forest classifier to 1000 number of trees, Gini criterionfor node splitting, and two minimum samples per split.

Additionally, the prediction problem was divided into an in-dividual and a grouped prediction task. The former refers to amodel developed specifically for a given occupant, using onlyparts of its data to train a model and test it on its remainingdata. On the other hand, the latter approach consists of combin-ing all occupants’ training data subsets with training a modeland testing it on all the occupants’ remaining data.

The data of each occupant was split into a 60:40 train testset based on time. That is, the first 60% of votes from eachoccupant was used in their training set, and the remaining 40%was used for testing. The sets were split by time to prevent thescenario of future data being used to predict the past. For thegrouped model, all the occupants’ training sets (60% of eachoccupant’s data) was used as one training set, and the remaining40% of each occupant’s data was combined with being used asone test set.

A primary component of the method was to test the abilityfor various feature sets to influence the prediction power of therandom forest model. The method used six combinations ofthese feature sets to test the influence each has in the predictivecapability of the overall model. The following is an overviewof these feature categories developed for testing:

• Time was created through feature engineering the timestamp of when an occupant gave feedback. This feature

6

was a cyclical representation of the hour of the day andday of the week. This simple feature type detects if cer-tain cyclical habits or components have a role in prefer-ence prediction and was included in all scenarios.

• Environmental Sensors were features extracted frommeasurement data from lighting (lux level), noise (dBlevel), temperature (deg. Celsius), and relative humidity(RH%) measurement. These variables were collected fromthe IEQ sensors that were closest spatially and temporallyto an occupant when they gave feedback.

• Near Body Temperature was a feature created from thetemperature sensor mounted on the smartwatch strap thathad temporal proximity to the time-stamp of when the oc-cupant gave feedback.

• Heart Rate was collected from the Fitbit smartwatch de-vice as an instantaneous value collected when the occupantgave feedback.

• Room was a feature that was encoded to a numerical pref-erence type based on the history of feedback in the roomin which the survey was taken. This feature was designedto increase the prediction accuracy by complimenting datafrom rooms of similar comfort profiles. For example, if anoccupant only works from their office, the model will stillbe able to accurately predict how that occupant may feelin other rooms that have a similar comfort profile to theiroffice.

• Preference History features are similar to the Room fea-tures. These features use the ratio of responses of eachtype (thermal, visual, and aural) that were calculated foreach user. This ratio was only calculated for the responsesof prefer cooler, prefer warmer, prefer dimmer, preferbrighter, prefer quieter, and prefer louder. E.g., the ratioof response of prefer cooler responses of a given occupantis calculated the following way: #prefer cooler votes

#total votes .

Model classification results were calculated using the F1-micro scores (as shown in Equation 1) which were equivalentto accuracy in the a multi-class classification problem by cal-culating precision and recall averaged across all classes,i.e., subjective thermal comfort response value. As the objec-tive was to provide a comparison among different feature setswith a standard metric, F1-micro was chosen due to its usagefor benchmarking different aspects of the modelling pipeline inthermal comfort datasets [78].

F1 = 2 ·precision · recallprecision + recall

(1)

4. Results

The results presented in this section are complemented withan interactive web application9 and interactive code10 which

9https://sde4demo.herokuapp.com/10https://github.com/buds-lab/humans-as-a-sensor-for-buildings

ThermalPreference

LightPreference

SoundPreference

Intensive Longitudinal Preference Collection

4 participants had relatively warmer preferences

9 participants had relatively cooler preferences

Prefer CoolerPrefer No ChangePrefer Warmer

Prefer DimmerPrefer No ChangePrefer Brighter

Prefer QuieterPrefer No ChangePrefer Louder

Figure 4: Overview of the intensive longitudinal data collected from the occu-pants according to the three categories: Prefer warmer (red) and Prefer cooler(blue) for thermal, Prefer Brighter (orange) and Prefer Dimmer (purple) forlighting, and Prefer Quieter (green) for acoustics. Each row is an occupant andeach box in that row shows that occupant’s feedback answers collected sequen-tially. The visualisation is diagrammatic in that vertical alignment of the boxesbetween different occupants does not imply identical time stamps.

enables the reader to regenerate all the plots. During a twoweek collection time of 30 participants, 4,378 comfort prefer-ence votes were collected, which is 146 data feedback pointsper person on average. From this set, 1,474 data points weresuccessfully localised to building environmental sensors. To al-low for comparison with those data, this subset was used foranalysis and machine learning in the following sections.

4.1. Grouping comfort preference tendencies

Figure 4 illustrates an overview of the intensive longitudinalpreference history data for each person according to the threepreference categories. These feedback responses were onlythose collected in the SDE4 building, and a maximum num-ber of 75 votes is shown. A simple clustering step was appliedin this figure to represent the segmentation according to eachpreference category on its own. This visualisation shows howthis simplified clustering step captured the tendency of an occu-pant to lean more towards one feedback response over the oth-ers. This segmentation was independent of the environmentalparameters of the spaces to maintain the simplicity of the ap-proach. The subsequent modelling steps were designed to testthe effectiveness of doing this type of simplified segmentation.

Figure 5a is an aggregated representation of the segmenta-tion process for each occupant, this time with all three prefer-

7

Occ

upan

t

PreferWarmer

PreferCooler

PreferQuieter

PreferLouder

PreferDimmer

PreferBrighter

Thermal AuralVisual

Per

cent

age

Vote

60%

45%

30%

15%

0%

No Preference

Prefer CoolerPrefer Cooler

Prefer Cooler

QuieterDimmer

Prefer Brighter

Quieter

Prefer Cooler Brighter Quieter

Prefer Brighter Quieter

Prefer Cooler

Prefer Warmer

Dimmer Quieter

Indoor Comfort Personality Tendencies:

(a) Occupant-based Clustering

a

PreferWarmer

PreferCooler

PreferQuieter

PreferLouder

PreferDimmer

PreferBrighter

Thermal AuralVisual

Naturally Ventilated Work Space

Naturally Ventilated Cafe

Naturally Ventilated Circulation Space

Lecture Rooms

Per

cent

age

Vote

60%

45%

30%

15%

0%

No Preference

Cooler Quieter requests

Cooler requests

Warmer requests

Relatively higher percentage of:

Brighter

requestsBrighter

Cooler Quieter requests

Quieter requests

Work Spaces / Offices

Roo

m

(b) Room-based Clustering

Figure 5: K-means clustering of preference tendencies quantified by average number of votes by occupant (a), and by room (b) within the test building. Each rowpresents the percentage of votes that fell into a respective preference. Dark colours in cells indicate higher preference.

8

Figure 6: Interactive visualisation of the data collection in the SDE4 building that highlights the spatial distribution of subjective preference data for thermal comfortin three dimensions. In terms of preference feedback, the blue dots indicate prefer cooler responses, the yellow dots are no change, and the red dots are preferwarmer.

ence categories being used in the clustering process. This fig-ure summarises each occupant as a row of data, and the colourof the box represents the percentage of votes given to a par-ticular preference category, where dark colours indicate higherpreference. These clusters provided segmentation of the usersaccording to their preference tendency types that were used inthe preference models. Even in a sample size of 30 occupants,there were varying comfort tendencies present, which comple-mented the concept of a personal comfort model tested by Kimet al. [79]. This clustering step provided the foundation for thecreation of the individual versus grouped models used in theprediction step.

4.2. Tagging the spatial context with preference feedback

While the subjective feedback highlighted varying comforttendencies within a building, localisation also enabled the char-acterisation of preference tendencies in certain zones. Figure 5bpresents each room as a row, where the colour of each cell rep-resents the percentage of a preference vote given for a particularroom. The utilisation of k-means clustering once again enabledthe splitting and labelling of these zones, this time by the ten-dency for different comfort preferences to be left by occupantsin those spaces. This result firstly served as an overview forfacility managers to understand the office spaces they manage,and take action to improve upon the comfort. A visualisation ofthe subjective thermal preference data can be found in Figure 6and online11.

11https://sde4demo.herokuapp.com/

4.3. Correlation with indoor environmental quality variables

One standard aspect of environmental comfort studies is thecomparison of feedback to objective environmental measure-ments. For the data collected in this study, standard distribu-tion plots of the environmental sensor data are summarised inFigure 7. Intuitive insight in the data can be observed, suchas the absence of prefer brighter votes after an illuminancethreshold of 250 lux. Nevertheless, there was a significant over-lap between classes for each of the environmental parameters,which were likely attributed to the numerous unmeasured vari-ables described by Figure 1, and the varying comfort tendenciesshown in Figure 5. This result reinforces the evidence that en-vironmental measurements are not descriptive enough to char-acterise a person’s preferences, which results in poor predictionas found in previous studies [50].

4.4. Predicting field-based indoor preference using intensivelongitudinal data

In this section, the time-series feedback was used to predictcomfort satisfaction. Figure 8 shows a comparison of the var-ious models built with the feature sets and process outlined inSection 3. The individual comfort model uses the occupant’straining data for prediction, while the grouped comfort modeluses the input data for the groupings outlined in Figure 5. Thetop of Figure 8 shows a table in which each row represents thefeature set that was used to train the model in that column.

Several insights were evident from this modelling analysis.Firstly, there were only small differences in the F1 scores be-tween the different feature sets for the visual and aural pref-erence models. These models, in general, had higher F1 scores

9

Environmental Temperature (C°) Environmental Humidity (RH%)Near Body Temperature (C°)

Environmental Light (lux)Heart Rate (beats per minute)

Prefer CoolerPrefer No ChangePrefer Warmer

Prefer DimmerPrefer No ChangePrefer Brighter

Prefer QuieterPrefer No ChangePrefer Louder

Environmental Noise (dB)

Figure 7: Distribution of sensor data by preference vote. While trends can be observed many feedback votes overlap for the same environmental or physiologicalmeasurement. This was possibly due to the different comfort tendencies as shown in Figure 5 or numerous other variables described in Figure 1 that are notaccounted for. Near body temperature and noise appear to have the most distinct differentiation.

than thermal preference prediction. Aural preference predictionhad the highest F1 score, which was intuitive since it became abinary classification challenge due to the lack of prefer louderfeedback responses.

Thermal preference prediction had more diversity across thefeature sets tested as compared to the other preference cate-gories. Merely using the conventional time-series and environ-mental sensor features had the lowest F1 score. Adding thephysiological attributes of heart rate and near body temperatureprovided marginal improvements. The best thermal preferencemodel used the physiological, room, and preference history fea-tures while excluding the environmental sensor data.

For all three preference categories, the grouped comfortmodel performed better than the individual version. Partici-pants with similar comfort preferences became clustered to-gether, thus increasing the training dataset for that particularoccupant type. This result showed the impact that assigning a

variety of peer groups can have on preference prediction.

4.5. Cold-start comfort preference prediction

The success of the preference models using grouping allowedfor models that can predict an occupant’s preferences withouttheir own personal data present. This scenario was labelled asa cold-start situation as it emulates when an occupant doesn’twear a watch to collect data in a particular building, but relieson crowdsourced data from peers that have a similar comfortpreference. The line graphs to the right of Figure 8 show theresults of this type of analysis. They illustrate the number ofoccupants required to sufficiently crowdsource the data for anaverage occupant for each of the preference categories. Theorange line represents an ordinary person who doesn’t wear asmartwatch, whereas the blue line is a smartwatch owner whois regularly giving feedback. In this study, nine and five userswere sufficient on average to crowdsource the thermal and vi-

10

sual comfort prediction respectively to the same accuracy as auser wearing a watch.

4.6. Predicting continuous comfort preference without sensors

Since the preference feedback in this methodology was at amuch higher high-frequency than a typical survey or occupantsacting on the thermostat, this study had preference data withrelatively high temporal and spatial diversity. The random for-est classifier was used to predict comfort preference based ona time-stamp input for each zone to create a continuous pre-diction over time. This approach emulates the concept of us-ing human feedback as a type of sensor. Figure 9 illustratesthe prediction of two different zones, an office and an outdoorspace, for a typical week using this model output. First, one cansee that the office was generally a comfortable space, while theoutdoor seating had an overall higher preference for cooling.Time-dependent fluctuations show how the model was able topredict comfort preference for different parts of the day or daysof the week. The office had a peak of warmer preference aroundmid-day. Finally, it was observed how the model, often inaccu-rately, tried to predict comfort at times where no data is present.The square peaks in the office for aural and visual predictionbetween the hours of 22:00 and 7:00 were due to an absence ofdata to accurately predict during these times.

5. Discussion

The results of this implementation showed the potential ofcollecting intensive longitudinal feedback from occupants inthe built environment. This approach revealed that the deploy-ment and implementation of such a methodology were effec-tive, and comfort models for visual, aural, and thermal comfortcan have similar performance to sensor measurements. The keyfocus in this section is to discuss the practical uses and limita-tions of the proposed methodology.

5.1. Practical Application of Intensive Longitudinal Data inIndustry

At the foundation of the method, the creation of more sig-nificant amounts of occupant feedback information in the formof preferences was successful. The utilisation of these type ofhigh-frequency subjective feedback data has potential for build-ing evaluation and occupant comfort optimisation. It changesthe paradigm in which facility managers operate a building.For example, instead of saying light levels are below the com-fort threshold in Office-1, the new conclusions could state that ahigher frequency of prefer brighter votes are recorded in Office-1. Furthermore, due to the high-frequency sampling rate pro-vided by the micro ecological momentary assessment method-ology, these periods of discomfort can also be mapped to partic-ular times, and certain groups of people. The time series com-fort profiles could also serve as input data for occupant-centric-control efforts of building systems which can then optimise forhuman comfort and energy optimisation.

5.1.1. Post-occupancy evaluation, commissioning, and sensorcalibration

A focus on post-occupancy evaluation could be a key tar-get for this type of data collection. In this scenario, a particu-lar sample of non-transient occupants of a recently constructedbuilding would be given smartwatches and asked to wear themfor 2-4 weeks. These data could then be used to supplement thesystems installed to characterise whether there are blind spotsin terms of sensors not picking up comfort-influencing phe-nomenon that is not being measured. For example, it’s rare tomeasure mean radiant temperature in most buildings; thereforehot spots might exist that are undetectable by thermostats andmight be a result of inadequate shading or control of shadingsystems. To adapt the presented methodology to this context isstraightforward as the two-week time frame of the experimentis similar. The current method involved asking each participantto wear the smartwatch until 100 data points were recorded.In a real-world setting, a similar approach could be deployed.At that point, perhaps the occupant could choose to return thedevice and rely on the data of co-workers for comfort predic-tion or continue to wear it and help crowdsource the data forothers. It is strongly recommended that these deployments useautomated indoor localisation to put the feedback in the spatialcontext without user intervention.

5.1.2. Potential for spatial recommendation systems and im-pact on activity-based workspace design

A less typical application for intensive longitudinal datamight be the development of spatial recommendation enginesfor occupants in activity-based workspaces. In these spaces, anoccupant doesn’t have a constant workspace but instead findsa space that matches their immediate needs. This paradigmcould prove to be an integral part of future working style, es-pecially in light of social distancing due to global pandemicssuch as COVID-19 that forces a less conventional spatial work-ing arrangement. This recommendation engine might work ina way that an occupant’s comfort tendencies could be matchedwith the comfort zone of the building. For example, those thatprefer warmer spaces can be recommended to work in areasthat have a higher percentage of prefer cooler votes. Previouswork in this direction showed progress using a platform knownas Spacematch [73]. Smart watch-based longitudinal feedbackcould enhance the model development process for this type ofplatform.

The aspect of testing group-based models in this study is es-sential for this context as building owners can’t expect all oc-cupants to be willing to wear or use devices. And those thatdo agree will likely have a limited amount of patience for giv-ing feedback over long periods. This paper tested the ability tocluster occupants such that it was not necessary for everyone inan office to wear a smartwatch. The only requirement for thistype of system to work is that each new employee would weara smartwatch during a two-week data collection phase, whichis sufficient to build their comfort preference tendency historyas described in Section 4.1. The experiment also showed thatnot everyone in an office space needed to be using the smart-watch application all the time. In this particular experiment,

11

Aur

al

Gro

uped

Indi

vidu

alG

roup

edIn

divi

dual

Gro

uped

Indi

vidu

al

Visu

alTh

erm

al

TimeFeature Sets

Environmental SensorsNear Body TemperatureHeart RateRoomPreference History

Without test subject’s data (occupant without a watch)

With test subject’s data (occupant with a watch)

0.64

0.72

0.80

0.88

0.96

Number of users added to the training set

Occupant without a watch has similar accuracy as an occupant with a watch

mean

upper quartile

lower quartile

Occupant without a watch has similar accuracy as an occupant with a watch

0.660.72

0.740.79

0.84 0.87

Best performingfeature set

Comfort prediction using the best performingfeature set:

Model has single occupant data:Individual comfort model

Model has all occupant data:Grouped comfort model

σ: σ: σ: σ: σ: σ:

σ: σ: σ: σ: σ: σ:

σ: σ: σ: σ: σ: σ:

σ: Standard Deviation

Figure 8: Left: Comparison of prediction F1-micro-score between grouped and individual comfort models using data from different feature sets. The feature set thatexcluded environmental sensor data for the thermal model had the highest F1-score, while minimal differences in F1-score were noted between feature sets of thevisual and aural models. Right: The accuracy in predicting the comfort of an individual as further participants are added to the training set. The blue line includesthe test participants training set in the training data, and the orange line excludes the test the participants training data meaning that it depends on crowdsourcedfeedback from other occupants.

12

Preference:

Mon

day

Tues

day

Wed

nesd

ay

Thur

sday

Frid

ay

Sat

urda

y

Sun

day

Mon

day

Tues

day

Wed

nesd

ay

Thur

sday

Frid

ay

Sat

urda

y

Sun

day

CoolerNo ChangeWarmer

DimmerNo ChangeBrighter

QuieterNo ChangeLouder

Zone 1 - Outdoor Seating Zone 2 - Office

Figure 9: Comfort prediction for two zones for an average occupant over a week. The grey circles indicate votes that were given for each category, and the shaded-out sections indicate times where no data were present. These time-series predictions can be used to detect anomalies, such as the mid-day peak for a warmerpreference for the office, or the general discomfort in the outdoor seating area. Note that there was an absence of data in these zones between the hours of 22:00 -7:00, and on the weekend. This lack of data caused inaccurate predictions as seen in the square-shaped peaks in the office.

six occupants were sufficient, on average, to crowdsource theprediction for the remaining 24. This value is not generalisableamongst all buildings and would change depending on the dif-ferent comfort tendencies the building occupants might present.The higher the variation in comfort preferences, the greater thenumber of the occupants needed to crowdsource the data.

In terms of office space design, the collection of intensivelongitudinal preference data could facilitate floor plan designdecisions. Understanding the breakdown of comfort needs ac-cording to the tendencies of the occupants would enable archi-tects to design or retrofit buildings with different comfort zonesto match the different types of people. For example, if the zoneswith more cooling were popular and being used to their capac-ity, then the floor plans or systems control could respond to bycreating more spaces of that type to increase the probability thata person feels comfortable.

5.1.3. Integration into building control systemsIntensive longitudinal data has the opportunity to influence

the control and automation systems of buildings through the useof preference feedback data in the control logic. Most buildingcontrol systems rely on optimising a set-point temperature thatis considered comfortable for the average occupant or comfortstandard [80]. In that scenario, discomfort instead of comfortis evaluated as the current difference of the environment ther-mostat and the HVAC system set-temperature [81], occupancydensity estimation, or via more traditional ways such as PMV

[82]. While some of these approaches have dealt with single-occupant offices or Personal Comfort Systems (PCS), there isa distinction between controlling the actual HVAC system andallowing the occupant to control their immediate space. PCSsystems are those that locally condition the occupant indepen-dent of the centralised HVAC system [83]. The intensive longi-tudinal data and the models developed in this study could helpthe controls field take the next step forward in occupant-centredbuilding controls through the use of reinforcement learning[84]. The feedback mechanism in reinforcement control is gen-erally the standard occupant-building interface such as switchor thermostat [85]. Intensive longitudinal data could be usedto enhance that interaction by focusing on finding the motiva-tions of those control actions. This work is a strong focus ofthe occupant-centric building operations in the IEA Annex 79project [86].

5.2. Prediction models are only as good as the training dataThe primary limitation of the presented approach was that

it would only work where data were present. As seen in Fig-ure 9, there were errors in the prediction when data were ab-sent, i.e., no historical data collected at similar time windowssuch as the middle of the night. Furthermore, since there wasa reliance on other subjects’ historical preferences, i.e., crowd-sourcing preferences, to evaluate environments, an office spacethat was rarely used would have a poor prediction of occupantcomfort. Classical comfort models based on sensor data do not

13

have this issue as spaces that are not used can still be charac-terised by the measured data. Furthermore, this particular studywas conducted in Singapore, which doesn’t have seasons andhas minimal variability in temperature. For seasonal countries,the day of the year would be an added feature that may take upto a year worth of data to train. Further work could investigatethe opportunity of using sensor data to characterise a space andthen continuously refine the comfort prediction by crowdsourc-ing the occupants’ preference on said space.

6. Conclusions

This paper presents how micro ecological momentary assess-ments of subjective comfort can generate sufficiently large in-tensive longitudinal data for occupant comfort prediction andenhancement that can supplement objective environmental sen-sor data, and empirical comfort models. Results of an imple-mentation of the platform on 30 occupants showed the segmen-tation and variation of indoor occupant comfort tendencies andhighlighted the shortcomings of one-size-fits-all comfort mod-els that are commonly applied in real buildings. Furthermore,the use of a smartwatch enabled data collection at a sufficientfrequency to build time-series models of indoor spaces. Thesemodels could be used to detect building anomalies, serve asbuilding data for subjective driven building control, or be usedto recommend spaces that best match the comfort preferencetendencies of each occupant. The optimum technological set-up uses a smartwatch for subjective data collection, combinedwith a method for localising an occupant in the building. Thislocalisation may be achieved by asking the occupant directlythrough the smartwatch, or through Bluetooth or WiFi signals.

Author contribution statement

PJ: hardware, software, infrastructure development, experi-mental design, implementation and lead author of the paper;MQ: software, infrastructure development, data analysis, ma-chine learning lead and author of the paper; MA: software, in-frastructure development, and author of the paper; CM: fund-ing, project leadership, experimental design, the correspondingauthor of the paper.

Funding

The Singapore Ministry of Education (MOE)(R296000181133 and R296000214114) and the NationalUniversity of Singapore (R296000158646) provided supportfor the development and implementation of this research.

Acknowledgements

This research contributes to the body of work for the Interna-tional Energy Agency (IEA) Energy in Building and Communi-ties (EBC) Annex 79 - Occupant-Centric Building Design andOperation.

Data availability statement

Segments of the raw data and analysis code used for thisstudy are available in an open-access Github repository that in-cludes further documentation12.

References

[1] M. Frontczak, S. Schiavon, J. Goins, E. Arens, H. Zhang, P. Wargocki,Quantitative relationships between occupant satisfaction and satisfactionaspects of indoor environmental quality and building design, Indoor air22 (2012) 119–131.

[2] I. Sakellaris, D. Saraga, C. Mandin, C. Roda, S. Fossati, Y. De Kluizenaar,P. Carrer, S. Dimitroulopoulou, V. Mihucz, T. Szigeti, et al., Perceived in-door environment and occupants’ comfort in european “modern” officebuildings: The officair study, International journal of environmental re-search and public health 13 (2016) 444.

[3] P. O. Fanger, et al., Thermal comfort. analysis and applications in envi-ronmental engineering., Thermal comfort. Analysis and applications inenvironmental engineering. (1970).

[4] R. De Dear, G. S. Brager, Developing an adaptive model of thermal com-fort and preference, ASHRAE Transactions, Vol 104 (1998).

[5] I. Asadi, I. Hussein, K. Palanisamy, W. Conference, M. Comfort, R. Cum-berland, E. Use, C. Environmental, M. Engineering, V. Mesiano, R. Su-laiman, S. N. Kamaruzzaman, T. Ramos, S. Dedesko, J. A. Siegel, J. A.Gilbert, B. Stephens, I. Asadi, N. Mahyuddin, P. Shafigh, E. M. d. A.Vieira, L. B. da Silva, E. L. de Souza, A survey of evaluation methodsused for holistic comfort assessment, PLoS ONE 953-954 (2016) 1513–1519.

[6] H. Tang, Y. Ding, B. Singer, Interactions and comprehensive effect ofindoor environmental quality factors on occupant satisfaction, Buildingand Environment 167 (2020).

[7] C. Henry, B. Emery, Effect of spiced food on metabolic rate., Humannutrition. Clinical nutrition 40 (1986) 165–168.

[8] R. Swaminathan, R. King, J. Holmfield, R. Siwek, M. Baker, J. Wales,Thermic effect of feeding carbohydrate, fat, protein and mixed meal inlean and obese subjects, The American journal of clinical nutrition 42(1985) 177–181.

[9] T. Webster, F. Bauman, G. Anwar, CBE Portable Wireless MonitoringSystem (PWMS): UFAD Systems Commissioning Cart Design Specifi-cations and Operating Manual, Internal Report, Center for the Built En-vironment, UC Berkeley (2007) 4.

[10] J. H. Choi, V. Loftness, A. Aziz, Post-occupancy evaluation of 20 officebuildings as basis for future IEQ standards and guidelines, in: Energy andBuildings, volume 46, pp. 167–175.

[11] H. Kim, Methodology for Rating a Building’s Overall Performance basedon the ASHRAE/CIBSE/USGBC Performance Measurement Protocolsfor Commercial Buildings, Journal of Chemical Information and Model-ing 53 (2013) 1689–1699.

[12] T. Parkinson, A. Parkinson, R. de Dear, Continuous IEQ monitoring sys-tem: Context and development, Building and Environment 149 (2019)15–25.

[13] J. F. Nicol, M. Wilson, A critique of European Standard EN 15251:Strengths, weaknesses and lessons for future standards, Building Re-search and Information 39 (2011) 183–193.

[14] S. Schiavon, B. Yang, Y. Donner, V.-C. Chang, W. W. Nazaroff, Thermalcomfort, perceived air quality, and cognitive performance when person-ally controlled air movement is used by tropically acclimatized persons,Indoor air 27 (2017) 690–702.

[15] S. G. Hodder, K. Parsons, The effects of solar radiation on thermal com-fort, International journal of biometeorology 51 (2007) 233–250.

[16] K. Krauchi, How is the circadian rhythm of core body temperature regu-lated?, Clinical Autonomic Research 12 (2002) 147–149.

[17] G. Chinazzo, J. Wienold, M. Andersen, Daylight affects human thermalperception, Scientific reports 9 (2019) 1–15.

12https://github.com/buds-lab/humans-as-a-sensor-for-buildings

14

[18] E. Halawa, J. van Hoof, V. Soebarto, The impacts of the thermal radia-tion field on thermal comfort, energy consumption and control—a criticaloverview, Renewable and Sustainable Energy Reviews 37 (2014) 907–918.

[19] T. Fukazawa, G. Havenith, Differences in comfort perception in relationto local and whole body skin wettedness, European journal of appliedphysiology 106 (2009) 15–24.

[20] B. Kingma, A. Frijns, W. van Marken Lichtenbelt, The thermoneutralzone: implications for metabolic studies, Front Biosci (Elite Ed) 4 (2012)1975–1985.

[21] P. Tikuisis, M. B. Ducharme, The effect of postural changes on bodytemperatures and heat balance, European journal of applied physiologyand occupational physiology 72 (1996) 451–459.

[22] A. P. Gagge, J. Stolwijk, J. Hardy, Comfort and thermal sensations andassociated physiological responses at various ambient temperatures, En-vironmental research 1 (1967) 1–20.

[23] G. A. Gaesser, G. A. Brooks, Muscular efficiency during steady-rate ex-ercise: effects of speed and work rate, Journal of applied physiology 38(1975) 1132–1139.

[24] G. Havenith, K. Kuklane, J. Fan, S. Hodder, Y. Ouzzahra, K. Lundgren,Y. Au, D. L. Loveday, A database of static clothing thermal insulationand vapor permeability values of non-western ensembles for use in ashraestandard 55, iso 7730, and iso 9920 ch-15-018 (rp-1504) (2015).

[25] J. M. Johnson, D. L. Kellogg Jr, Thermoregulatory and thermal controlin the human cutaneous circulation, Front Biosci (Schol Ed) 2 (2010)825–853.

[26] H. Zhang, A. Hedge, Laptop heat and models of user thermal discomfort,in: Proceedings of the Human Factors and Ergonomics Society AnnualMeeting, volume 58, SAGE Publications Sage CA: Los Angeles, CA, pp.1456–1460.

[27] S. Karjalainen, Thermal comfort and gender: a literature review, Indoorair 22 (2012) 96–109.

[28] H.-W. Huang, W.-C. Wang, C.-C. K. Lin, Influence of age on thermalthresholds, thermal pain thresholds, and reaction time, Journal of ClinicalNeuroscience 17 (2010) 722–726.

[29] G. Havenith, I. Holmer, K. Parsons, Personal factors in thermal comfortassessment: clothing properties and metabolic heat production, Energyand buildings 34 (2002) 581–591.

[30] S. M. Frank, S. N. Raja, C. F. Bulcao, D. S. Goldstein, Relative contri-bution of core and cutaneous temperatures to thermal comfort and auto-nomic responses in humans, Journal of Applied physiology 86 (1999)1588–1593.

[31] J. A. Boulant, Role of the preoptic-anterior hypothalamus in thermoreg-ulation and fever, Clinical infectious diseases 31 (2000) S157–S161.

[32] M. Cabanac, Physiological role of pleasure, Science 173 (1971) 1103–1107.

[33] G. C. Brainard, J. P. Hanifin, J. M. Greeson, B. Byrne, G. Glickman,E. Gerner, M. D. Rollag, Action spectrum for melatonin regulation inhumans: evidence for a novel circadian photoreceptor, Journal of Neuro-science 21 (2001) 6405–6412.

[34] R. Perez, P. Ineichen, R. Seals, J. Michalsky, R. Stewart, Modeling day-light availability and irradiance components from direct and global irra-diance, Solar energy 44 (1990) 271–289.

[35] P. Leather, M. Pyrgas, D. Beale, C. Lawrence, Windows in the workplace:Sunlight, view, and occupational stress, Environment and behavior 30(1998) 739–762.

[36] Q. Dai, L. Hao, Y. Lin, Z. Cui, Spectral optimization simulation of whitelight based on the photopic eye-sensitivity curve, Journal of AppliedPhysics 119 (2016) 053103.

[37] A. I. Slater, P. R. Boyce, Illuminance uniformity on desks: Where is thelimit?, Lighting research & technology 22 (1990) 165–174.

[38] R. A. Baron, M. S. Rea, S. G. Daniels, Effects of indoor lighting (illu-minance and spectral distribution) on the performance of cognitive tasksand interpersonal behaviors: The potential mediating role of positive af-fect, Motivation and emotion 16 (1992) 1–33.

[39] A. J. Wilkins, I. Nimmo-Smith, A. I. Slater, L. Bedocs, Fluorescent light-ing, headaches and eyestrain, Lighting Research & Technology 21 (1989)11–18.

[40] A. Main, A. Dowson, M. Gross, Photophobia and phonophobia in mi-graineurs between attacks, Headache: The Journal of Head and Face Pain37 (1997) 492–495.

[41] J. Yin, S. Zhu, P. MacNaughton, J. G. Allen, J. D. Spengler, Physiologicaland cognitive performance of exposure to biophilic indoor environment,Building and Environment 132 (2018) 255–262.

[42] R. G. Hopkinson, Glare from daylighting in buildings, Applied er-gonomics 3 (1972) 206–215.

[43] M. Pierrette, E. Parizet, P. Chevret, J. Chatillon, Noise effect on com-fort in open-space offices: development of an assessment questionnaire,Ergonomics 58 (2015) 96–106.

[44] R. Job, Community response to noise: A review of factors influencingthe relationship between noise exposure and reaction, The Journal of theAcoustical Society of America 83 (1988) 991–1001.

[45] J. Kim, R. De Dear, Workspace satisfaction: The privacy-communicationtrade-off in open-plan offices, Journal of Environmental Psychology 36(2013) 18–26.

[46] D. Templeton, D. Saunders, Acoustic design, Elsevier, 2014.[47] S. Y. Lee, J. Brand, Can personal control over the physical environment

ease distractions in office workplaces?, Ergonomics 53 (2010) 324–335.[48] A. Kjellberg, B. Skoldstrom, Noise annoyance during the performance of

different nonauditory tasks, Perceptual and motor skills 73 (1991) 39–49.[49] S. P. Banbury, D. C. Berry, Office noise and employee concentration:

Identifying causes of disruption and potential improvements, Ergonomics48 (2005) 25–37.

[50] T. Cheung, S. Schiavon, T. Parkinson, P. Li, G. Brager, Analysis ofthe accuracy on pmv–ppd model using the ashrae global thermal comfortdatabase ii, Building and Environment (2019).

[51] A. A. Stone, S. Shiffman, A. A. Atienza, L. Nebeling, A. Stone, S. Shiff-man, A. Atienza, L. Nebeling, Historical roots and rationale of ecologi-cal momentary assessment (ema), The science of real-time data capture:Self-reports in health research (2007) 3–10.

[52] S. Intille, C. Haynes, D. Maniar, A. Ponnada, J. Manjourides, µema:Microinteraction-based ecological momentary assessment (ema) using asmartwatch, in: Proceedings of the 2016 ACM International Joint Con-ference on Pervasive and Ubiquitous Computing, ACM, pp. 1124–1128.

[53] D. Wang, M. T. Amin, S. Li, T. Abdelzaher, L. Kaplan, S. Gu, C. Pan,H. Liu, C. C. Aggarwal, R. Ganti, X. Wang, P. Mohapatra, B. Szymanski,H. Le, Using humans as sensors: an estimation-theoretic perspective, in:Proceedings of the 13th international symposium on Information process-ing in sensor networks, IPSN ’14, IEEE Press, 2014, pp. 35–46.

[54] M. Avvenuti, M. G. C. A. Cimino, S. Cresci, A. Marchetti, M. Tesconi, Aframework for detecting unfolding emergencies using humans as sensors,Springerplus 5 (2016) 43.

[55] M. Vielberth, F. Menges, G. Pernul, Human-as-a-security-sensor for har-vesting threat intelligence, Cybersecurity 2 (2019) 23.

[56] D. Heinzerling, S. Schiavon, T. Webster, E. Arens, Indoor environmentalquality assessment models: A literature review and a proposed weightingand classification scheme 70 (2013) 210–222.

[57] M. Ncube, S. Riffat, Developing an indoor environment quality tool forassessment of mechanically ventilated office buildings in the UK - A pre-liminary study, Building and Environment 53 (2012) 26–33.

[58] L. T. Wong, K. W. Mui, P. S. Hui, A multivariate-logistic model foracceptance of indoor environmental quality (IEQ) in offices, Buildingand Environment 43 (2008) 1–6.

[59] A. C. Lai, K. W. Mui, L. T. Wong, L. Y. Law, An evaluation model forindoor environmental quality (IEQ) acceptance in residential buildings,Energy and Buildings 41 (2009) 930–936.

[60] R. Cohen, M. Standeven, B. Bordass, A. Leaman, Assessing buildingperformance in use 1: The Probe process, Building Research and Infor-mation 29 (2001) 85–102.

[61] T. Webster, E. Arens, G. Anwar, J. Bonnell, F. Bauman, C. Brown, Ufadcommissioning cart: Design specifications and operating manual (2007).

[62] M. Jin, S. Liu, S. Schiavon, C. Spanos, Automated mobile sensing:Towards high-granularity agile indoor environmental quality monitoring,Building and Environment 127 (2018) 268–276.

[63] S. R. Porter, M. E. Whitcomb, W. H. Weitzer, Multiple surveys of stu-dents and survey fatigue, New Directions for Institutional Research 2004(2004) 63–73.

[64] S. Liu, S. Schiavon, H. P. Das, M. Jin, C. J. Spanos, Personal ther-mal comfort models with wearable sensors, Build. Environ. 162 (2019)106281.

[65] A. Aryal, B. Becerik-Gerber, A comparative study of predicting indi-vidual thermal sensation and satisfaction using wrist-worn temperature

15

sensor, thermal camera and ambient temperature sensor, Build. Environ.160 (2019) 106223.

[66] A. K. Clear, S. Mitchell Finnigan, P. Olivier, R. Comber, ThermoKiosk:Investigating Roles for Digital Surveys of Thermal Experience in Work-place Comfort Management, Proc. of CHI (2018) 1–12.

[67] L. Engelen, F. Held, Understanding the office: Using ecological momen-tary assessment to measure activities, posture, social interactions, mood,and work performance at the workplace, Buildings 9 (2019) 54.

[68] B. Monnot, E. Wilhelm, G. Piliouras, Y. Zhou, D. Dahlmeier, H. Y. Lu,W. Jin, Inferring activities and optimal trips: Lessons from singapore’snational science experiment, in: Complex Systems Design & Manage-ment Asia, Springer International Publishing, 2016, pp. 247–264.

[69] E. Wilhelm, Y. Zhou, N. Zhang, J. Kee, G. Loh, N. Tippenhauer, Sensg:Large-scale deployment of wearable sensors for trip and transport modelogging, Technical Report, 2016.

[70] F. Benita, G. Bansal, B. Tuncer, Public spaces and happiness: Evidencefrom a large-scale field experiment, Health Place 56 (2019) 9–18.

[71] V. K. Ojha, D. Griego, S. Kuliga, M. Bielik, P. Bus, C. Schaeben,L. Treyer, M. Standfest, S. Schneider, R. Konig, D. Donath, G. Schmitt,Machine learning approaches to understand the influence of urban envi-ronments on human’s physiological response, Inf. Sci. 474 (2019) 154–169.

[72] M. S. Rahaman, J. Liono, Y. Ren, J. Chan, S. Kudo, T. Rawling, F. D.Salim, An Ambient-Physical system to infer concentration in open-planworkplace, IEEE Internet of Things Journal (2020) 1–1.

[73] T. Sood, P. Janssen, C. Miller, Spacematch: Using environmental prefer-ences to match occupants to suitable activity-based workspaces, Frontiersin Built Environment 6 (2020) 113.

[74] T. Sood, M. Quintana, P. Jayathissa, M. AbdelRahman, C. Miller, TheSDE4 learning trail: Crowdsourcing occupant comfort feedback at a net-zero energy building, J. Phys. Conf. Ser. 1343 (2019) 012141.

[75] P. Jayathissa, M. Quintana, T. Sood, N. Narzarian, C. Miller, Is yourclock-face cozie? a smartwatch methodology for the in-situ collectionof occupant comfort data, CISBAT - Climate Resilient Cities - EnergyEfficiency and Renewables in the Digital Era (2019).

[76] M. M. Abdelrahman, P. Jayathissa, C. Miller, YAK: An Indoor Position-ing App for Spatial-Temporal Indoor Environmental Quality Research(2019).

[77] J. Kim, Y. Zhou, S. Schiavon, P. Raftery, G. Brager, Personal comfortmodels: predicting individuals’ thermal preference using occupant heat-ing and cooling behavior and machine learning, Building and Environ-ment 129 (2018) 96–106.

[78] M. Luo, J. Xie, Y. Yan, Z. Ke, P. Yu, Z. Wang, J. Zhang, Comparing ma-chine learning algorithms in predicting thermal sensation using ASHRAEComfort Database II, Energy and Buildings 210 (2020) 109776.

[79] J. Kim, S. Schiavon, G. Brager, Personal comfort models–a new paradigmin thermal comfort for occupant-centric environmental control, Buildingand Environment 132 (2018) 114–124.

[80] D. Enescu, A review of thermal comfort models and indicators for indoorenvironments, Renewable and Sustainable Energy Reviews 79 (2017)1353–1379.

[81] L. Barrios, W. Kleiminger, The Comfstat - Automatically sensing thermalcomfort for smart thermostats, 2017 IEEE International Conference onPervasive Computing and Communications, PerCom 2017 (2017) 257–266.

[82] J. Y. Park, Z. Nagy, Comprehensive analysis of the relationship betweenthermal comfort and building control research - A data-driven literaturereview, Renewable and Sustainable Energy Reviews 82 (2018) 2664–2679.

[83] H. Zhang, E. Arens, Y. Zhai, A review of the corrective power of per-sonal comfort systems in non-neutral ambient environments, Buildingand Environment 91 (2015) 15–41.

[84] J. Y. Park, M. M. Ouf, B. Gunay, Y. Peng, W. O’Brien, M. B. Kjærgaard,Z. Nagy, A critical review of field implementations of occupant-centricbuilding controls, Build. Environ. 165 (2019) 106351.

[85] J. Y. Park, T. Dougherty, H. Fritz, Z. Nagy, LightLearn: An adaptive andoccupant centered controller for lighting based on reinforcement learning,Build. Environ. 147 (2019) 397–414.

[86] W. O’Brien, A. Wagner, M. Schweiker, A. Mahdavi, J. Day, M. B. Kjær-gaard, S. Carlucci, B. Dong, F. Tahmasebi, D. Yan, T. Hong, H. B. Gunay,Z. Nagy, C. Miller, C. Berger, Introducing IEA EBC annex 79: Key chal-

lenges and opportunities in the field of occupant-centric building designand operation, Build. Environ. 178 (2020) 106738.

[87] C. Duarte Roa, S. Schiavon, T. Parkinson, Targeted occupant surveys : Anovel method to effectively relate occupant feedback with environmentalconditions, Preprint (2020).

[88] A. Lipczynska, S. Schiavon, L. T. Graham, Thermal comfort and self-reported productivity in an office with ceiling fans in the tropics, Buildingand Environment 135 (2018) 202–212.

[89] M. Quintana, C. Miller, Poster Abstract: Towards Class-Balancing Hu-man Comfort Datasets with GANs, in: BuildSys ’19 Proceedings of the6th ACM International Conference on Systems for Energy-Efficient BuiltEnvironments, New York, NY, USA.

[90] M. M. Abdelrahman, A. Chong, C. Miller, Build2Vec: Building Repre-sentation in Vector Space, in: SimAUD 2020, Online, pp. 101–104.

16