Daily Stress Recognition from Mobile Phone Data, Weather Conditions and Individual Traits

Dail y Stress Recognition fr om Mobile Phone Data, W eather Conditions and Individual T raits Andrey Bogomolov University of T rento , SKIL T elecom Italia Lab Via Sommar ive , 5 I-38123 Po vo - T rento, Italy andrey .bogomolov@unitn.it Bruno Lepri F ondazione Br uno Kessler Via Sommar ive , 18 I-38123 Po vo - T rento, Italy lepri@fbk.eu Michela Ferron F ondazione Br uno Kessler via Sommar ive , 18 I-38123 Po vo - T rento, Italy f erron@fbk.eu F abio Pianesi F ondazione Br uno Kessler Via Sommar ive , 18 I-38123 Po vo - T rento, Italy pianesi@fbk.eu Ale x (Sandy) P entland MIT Media Lab 20 Ames Street Cambridge, MA, USA pentland@mit.edu ABSTRA CT Researc h has prov en that stress reduces quality of life and causes many diseases. F or this reason, several researc hers devised stress detection systems based on physiolog ical pa- rameters. Ho w ever, these systems require that obtrusiv e sensors are contin uously carried by the user. In our pa- per, we propose an alternativ e approach providing evidence that daily stress can b e reliably recognized based on b e- ha vioral metrics, derived from the user’s mobile phone ac- tivit y and from additional indicators, such as the weather conditions (data p ertaining to transitory prop erties of the en vironmen t) and the p ersonalit y traits (data concerning permanent dispositions of individuals). Our multifac torial statistical model, which is p erson-independent, obtains the accuracy score of 72.28% for a 2-class daily stress recog- nition problem. The model is efficient to implement for most of multimedia applications due to highly reduced lo w- dimensional feature space (32d). Moreo v er, we identify and discuss the indicators which hav e strong predictive pow er. Categories and Subject Descriptors I.5 [ Computing Metho dologies ]: P A TTERN RECOG- NITION— Mo dels, Design Metho dolo gy, Implementation ; J.4 [ Computer Applicatio ns ]: SOCIAL AND BEHA VIORAL SCIENCES— So ciolo gy, Psycholo gy General T erms Algorithms; Exp erimen tation; Measurement; Theory Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full cita- tion on the first page. Copyrights for components of this work owned by others than A CM must be honored. Abstracting with credit is permitted. T o copy otherwise, or re- publish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. MM’14, November 3–7, 2014, Orlando, Florida, USA. Copyright 2014 A CM 978-1-4503-3063-3/14/11 ...$15.00. http://dx.doi.org/10.1145/2647868.2654933. K eywords stress recognition; mobile sensing; p erv asiv e computing 1. INTR ODUCTION No w ada ys, the num ber of mobile phones in use worldwide is ab out 5 billion, with millions of new subscrib ers everyda y 1 . Mobile phones allow for unobtrusiv e and cost-efficient ac- cess to h uge streams of previously inaccessible data related to daily so cial b eha vior [29]. These devices are able to sense a w ealth of b eha vioral data such as (i) lo cation, (ii) other devices in physical proximit y through Blueto oth scanning, (iii) communication data, including b oth metadata (logs of who, when, and duration) of phone calls and text messages (sms), etc. Corresp ondingly , the av ailability is contin uously gro wing of huge streams of p ersonal data related to activ- ities, routines and social interactions [11, 29] whic h repre- sen t a nov el opp ortunit y to address fundamental problems of our so cieties in differen t fields, such as mobilit y and ur- ban planning [19], finance [46], healthy livin g and sub jective w ell-being [31, 33]. In this work, we fo cus on one of the most widespread and debilitating problem of our sub jective well-being and our society: stress. Stress is a well-kno wn condition in mo dern life and research has sho wn that the amoun t of cum ulativ e stress pla ys a role in a broad range of physical, psyc holog- ical and behavioural conditions, such as anxiety , low self- esteem, depression, so cial isolation, cognitive impairmen ts, sleep and immunol ogical disorders, neurodegenerative dis- eases and other medical conditions [8], while also signif- ican tly contributing to healthcare costs. Hence, measur- ing stress in daily life situations has b ecome an imp ortan t c hallenge [39]. T o day , the a v ailability of huge and diverse streams of p erv asiv e data pro duced by and about people allo ws for automatic, unobtrusive, and fast recognition of daily stress levels. An early prediction of stress symptoms can indeed help to preven t situations that are risky for h u- man life [23]. Sev eral s tudies hav e pro duced in teresting results that sup- port the feasibility of detecting stress levels through phys- 1 h ttp://www.ericsson.com/ericsson-mobilit y-rep ort iological sensors (see [23], [26]). How ev er, the use of ph ys- iological sensors is limited by several shortcomings. Stress detection systems based on physiological measurement such as heart-rate v ariabilit y or skin conductance are intrusiv e and need to b e easily wearable to b e exploited in natural settings; the data they pro duce can b e confounded by daily life activities such as sp eaking or drinking; they exhibit im- portant b etw een-p erson differences [39]. Recen tly , social psyc hologist Miller wrote “The Smart- phone Psyc hology Manifesto” in whic h he argued that the smartphones should b e seriously considered as new research tools for psychology . In his opinion, these to ols could rev olu- tionize all fields of psyc hology and other b eha vioral sciences making these disciplines more p ow erful, sophisticated, and grounded in real-w orld b ehavior [36] and [30]. Indeed, sev- eral w orks hav e started to use smartphone activity data in order to detect and predict p ersonalit y traits [6, 9, 37, 48], moo d states [31], and daily happiness [38]. Stopczynski et al. [49] describ ed the Cop enhagen Netw orks Study , a large- scale study designed to measure human interactions span- ning multiple years. Smartphones data can b e used to detect stress levels as w ell. Indeed, stress levels are associated with the type of ac- tivities people engage in, including those executed at/through their smartphone (for instance, a high n um b er of phone calls and/or e-mails from many different p eople could be associ- ated with higher stress levels). W eather conditions – an en vironmen tal transitory prop ert y – in turn, hav e b een ar- gued [24], [41] to b e often asso ciated with stress, acting ei- ther directly (as stressors) or indirectly (by affecting in divid- ual sensitivity to stressors). Finally , the impact of all these transitory factors – (smartphone) activities and weather con- ditions – on stress induction can b e exp ected to b e modu- lated by p ersonal ch aracteristics and differences [50], [52]. F or example, a neurotic p erson could react with higher lev- els of stress to a high num ber of interac tions (call, sms or pro ximit y interactions) than an emotionally stable person; an extrov ert or agreeable person, in turn, might well find him/herself at ease with a high n um b er of interaction s. In this paper, we approach the automatic recognition of daily stress as a 2-class classification problem (non-stressed vs stressed) based on information concerning different types of data: a) people activities, a s detected through their smart- phones; b) w eather conditions; c) personality traits. The information ab out p eople activities is represented by fea- tures extracted from call and sms logs and from Bluetooth hits, able to capture (i) the amoun t of calls, of sms and of pro ximit y interactions; (ii) the diversit y of calls, of sms, and of proximit y interactions; and (iii) regularity in user b eha v- iors. In addition, we use weather conditions (environmen tal and transitory factors) along with personality traits (internal and stable factors); the latter are mediating factors that can modulate people resp onses to stressors (e.g., weather, daily activit y). This multifactorial approach will b e compared to approac hes based only on a family of features (p ersonalit y , w eather conditions, mobile phone features) or simpler com- binations of families of features (personality and weather conditions; personality and mobile phone features; w eather conditions and mobile phone features). Classification experiments are performed using a v ariet y of approac hes and the b est solution for our classification prob- lem was found using an ensem ble of tree classifiers based on a Random F orest algorithm. Our multifac torial approac h obtains an accuracy score of 72.28 % for a 2-class daily stress recognition problem, pro viding evidence that individual daily stress can b e reliably predicted from the combination of smartphone usage data, weather conditions and individual dispositions (personality traits). Interestingly , if one of these information sources is dropp ed, the recognition performances decrease drastically . In sum, the main con tributions of this pa p er are as follows: 1. W e prop ose a multi-factorial data-driven approach to the prediction of individual daily stress; 2. W e v alidate our approac h with a seven-mon ths dataset collected from 111 sub jects; 3. W e provide a comprehensive analysis of the predic- tiv e pow er of the proposed approac h and a co mparison with approaches based only on single families of fea- tures (p ersonalit y , weather conditions, mobile phone features) or pairwise combinations thereof (p ersonal- it y and w eather conditions, personality and mobile phone features, w eather conditions and mobile phone features). 2. RELA TED WORK A large bo dy of research on stress detection focused on ph ysiological measuremen ts to infer stress levels (see [23], [34], [39]). Heart-rate v ariability , galv anic skin response, respiration, muscl e activity and temp erature are among the most relev an t features. How ev er, despite providing reliable insigh ts on stress levels, this approac h has ma jor limitations because it comprises wearable sensors that need to b e car- ried at all times to allow for con tin uous monitoring. Among the differen t changes in physiological parameters that happ en during stressful situations, v ariation in sp eec h production has inspired a num ber of studies using acous- tic sensing on smartphones. Researc h on stress detection based on voice analysis considered different sp eec h ch ar- acteristics such as pitch, glottal pulse, sp ectral slope and phonetic v ariations. F or example, Lu and colleagues [32] proposed StressSense, an Android application for stress de- tection from h uman v oice in real-life conv ersation, and they ac hiev ed 81% and 76% accuracy for indo or and outdo or en- vironmen ts. Ho w ever, these metho ds depend on sound qu alit y , whic h is not granted in natural settings (e.g., crowded public places, noisy outdoor), and the correlation b et ween sp eec h and emo- tion is sub jected to large individual differences [43]. Hence, our p erformance of 72.28% is a go od and reliable alterna- tiv e to stress detection. Other studies fo cused on the video analysis of behaviou ral correlates of psyc hological stress [18]. These systems, despite pro viding an unobtrusiv e method for stress monitoring, cannot b e employ ed in a large v ariet y of real world and mobile en vironmen ts and p ose priv acy con- cerns related to the recording of people’s b ehavio ur. A promising approach that can o vercome the ma jor short- comings of stress detection based on physiological measures and on audio/video analysis is activit y recognition from smart- phone usage patterns. Studies in this field ha ve been mainly focused on the understanding of relational dynamics of in- dividuals [14]. Recently studies ha ve started to in vestigate ho w smartphone usage habits can provide insights in to users’ affectiv e state [31] and stress levels [1]. LiKamW a and col- leagues [31] proposed Mo odScop e, a mobile softw are system that recognizes the users’ mo od, but not stress states, from smartphone usage analysis. They collected usage data and self-reported moo d in a tw o months longitudinal study and used them to train moo d mo dels. Smartphone usage data consisted in phone calls, SMSes, e-mail messages, applica- tion use, web bro wsing histories and lo cation changes, while self-reported mo od was collected from users’ input at least four times a day . Moo dScop e reached a 66% accuracy of participan ts’ daily-av erage mo o d, with phone calls and cat- egorized applications as the most useful features for moo d discrimination. Bauer and Luko wicz [1] fo cused on mid-term stress detec- tion, monitoring 7 students during a tw o week exam session follo w ed by tw o weeks of non-stressful p eriod. The recorded data were related to participants’ mobilit y patterns and so- cial interactions, and included users’ lo cation, Blueto oth pro ximit y , phone calls and SMSes. These features allow ed to detect an av erage b eha viour mo dification of 53% for each user during the exam session. A limitation of this study is the small num ber of sub jects. Our multifactorial approach outperforms the approach prop osed by [1] although a direct comparison ma y be not adequate giv en the differen t fo cus: our approach tend to daily classify people as ” not stressed” or ” stressed” , while Bauer and Luko wicz try to detect stress- ful situations. In 2013, Sano and Picard [42] reported an accuracy per- formance in stress recognition of 75% using a combination of features obtained from mobile phones and w earable sen- sors. Ho w ev er, the limited num b er of sub jects used in their experiments (18) and the limited n umber of da ys (5) mak e preliminary the results of this study . 3. D A T A COLLECTION F rom Nov ember 12, 2010 to Ma y 21, 2011, we collected a dataset capturing the lives of 117 sub jects living in a married graduate student residency of a ma jor US universit y . Our sample of sub jects has a large v ariety in terms of pro venance and cultural background : we ha v e sub jects from 16 countries suc h as USA, China, Israel, India, Iran, Russia, etc. During this perio d, eac h partic ipan t w as equipp ed with an Android- based cellula r phone incorp orating a sensing softw are explic - itly designed for collecting mobile data. Such soft w are runs in a passiv e manner and does not interfere with the ev ery da y usage of the phone. The data collected consisted of: (a) call logs, (b) sms logs, (c) pro ximit y data, obtained by scanning near-b y phones and other Blueto oth devices every five min- utes, and (d) data from surv eys administered to participant s, whic h provided self-rep orted information about personality traits ( “Big Fiv e” ) and self rep orted information about daily stress. Pro ximit y intera ction data w ere derived from Blueto oth hits in a similar w a y as in previous realit y mining studies [13]. Blueto oth scans were p erformed ev ery 5 minutes in order to keep the battery from draining while ac hieving a high enough temp oral resolution. The Blueto oth log of a giv en smartphone w ere then used to extract the list of the other participants’ phones which w ere in proximit y . In total, the dataset consisted of 33497 phone calls, 22587 SMS, and 1460939 Blueto oth hits. 3.1 Stress data A t the ev ening, the participants were also asked to fill daily surveys ab out their daily self-p erceiv ed stress level. The stress information was rep orted by the participants fill- ing a seven items scale with 1 = “not stressed” , 4 = “neutral” and 7 = “extremely stressed” . In our experiments w e used the data only for the sub jects (111 sub jects) who had pro- vided at least 2 weeks of consecutiv e data. The distribution of daily stress is visualized in Fig. 1. W e see that it has a small negative skew – the density is mov ed to the higher region of stress score. The distribution has negativ e excess kurtosis, which in our case means that the sample reported a sp ecific daily stress score more often than the neutral. Fig. 2 shows that within-person daily stress v ariance is more spread than b et w een-person, but the den- sit y of b etw een-p erson v ariance is higher. H i s to g r a m o f r $ s tr e s s Stress Value Count 0 1 2 3 4 5 6 7 0 500 1000 1500 2000 2500 Figure 1: Recorded Stress Scores Densit y 0 1 2 3 4 5 6 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 density.default(x = r1[, 1]) V ar iance Density Within − P erson V ariance Between − P erson Variance Figure 2: Within- and b et ween-sub ject v ariance 3.2 Personality Data Sev eral studies in so cial psychology inv estigated the rela- tionships betw een personality traits and psyc hological stress. P ersonalities that tend to be more negative are usually as- sociated with greater distress, while outgoing and p ositiv e T able 1: Selected F eatures Ranked b y Mean Decrease in Accuracy Rank F eature 0 1 Mean Decrease in Accuracy Mean Decrease in Gini Index 1 p ersonalit y .Conscientiousness 13.65 18.04 23.35 159.96 2 p ersonalit y .Agreeableness 14.22 19.73 22.92 167.30 3 p ersonalit y .Neuroticism 15.96 21.04 22.56 183.87 4 p ersonalit y .Op enness 14.20 14.18 21.38 139.23 5 p ersonalit y .Extraversion 15.75 15.02 21.07 158.51 6 weather.MeanT emp erature 14.50 6.34 17.44 322.27 7 sms.RepliedEven ts.Latency .Median 8.83 13.85 15.63 48.74 8 weather.Humidit y 15.33 2.10 15.45 298.13 9 sms.AllEven tsPerDay 8.61 0.56 10.50 42.91 10 bluetooth.Q95TimeF orWhichIdSeen 4.99 6.05 9.94 32.47 11 bluetooth.MaxTimeF orWhichIdSeen 6.24 7.23 9.47 32.12 12 sms.IncomingAndOutgoingP erDay 7.45 1.26 9.38 41.59 13 weather.Visibilit y 9.94 1.26 9.22 251.27 14 weather.WindSpeed 8.77 1.30 8.67 282.10 15 bluetooth.Q90TimeF orWhichIdSeen 4.24 6.75 8.64 28.41 16 bluetooth.T otalEntropySha nnon 5.04 3.51 8.56 31.37 17 call.Entrop yMillerMadowOutgoingT otal 4.25 4.10 8.54 27.49 18 call.Entrop yShannonOutgoingAndIncomingT otal 4.23 4.86 8.53 26.28 19 bluetooth.T otalEntropyMill erMadow 5.06 4.22 8.50 32.09 20 bluetooth.IdsMoreThan09TimeSlotsSeen 6.11 5.85 8.43 27.88 21 bluetooth.IdsMoreThan04TimeSlotsSeen 6.34 4.59 8.04 24.64 22 call.Entrop yShannonMissedOutgoingT otal 3.13 4.92 7.85 24.34 23 bluetooth.IdsMoreThan19TimeSlotsSeen 2.97 5.16 7.78 20.87 24 call.Entrop yShannonOutgoingT otal 3.10 6.45 7.78 24.79 25 bluetooth.Q75TimeF orWhichIdSeen 5.16 4.70 7.76 22.07 26 call.Entrop yMillerMadowMissedOutgoingT otal 4.09 5.45 7.55 24.64 27 call.Entrop yMillerMadowOutgoingAndIncomingT otal 3.87 6.29 7.51 28.63 28 sms.OutgoingAndIncomingT otalEn tropyMi llerMadow 4.68 3.84 7.19 17.63 29 sms.OutgoingT otalEntrop yMillerMadow 5.22 1.49 7.19 18.88 30 bluetooth.Q50TimeF orWhichIdSeen 1.53 7.29 7.08 18.91 31 bluetooth.Q68TimeF orWhichIdSeen 2.36 5.96 6.68 19.05 32 sms.OutgoingT otalEntrop yShannon 2.53 2.77 5.13 17.59 personalities generally exp erience less distress [50], [52]. The ma jority of the studies that hav e examined the relationship betw een p ersonalit y and distress fo cused on the Big Five traits [25], a p ersonalit y model owing its name to the five traits it tak es as a constitutiv e of p eople’s p ersonalit y: Ex- tra v ersion, Neuroticism, Agreeableness, Conscientiousn ess, and Openness to experience. Research ers show ed significant associations betw een psychological stress, on the one hand, and Neuroticism, Extrav ersion and Conscien tiousness, on the other. Duggan et al. [12] found that individuals high in Neuroticism may b e more vulnerable to exp eriencing dis- tress as they resp ond more negatively to daily stressors and report more daily stressful even ts and higher levels of daily stress. When p eople with high scores in Neuroticism en- coun ter stressful even ts, they tend to exp erience them as more a versiv e than those lo w in this trait [3], [21]. Finally , in a study with universit y studen ts, V olrath and T orgensen [52] sho w ed that students with more adaptive p ersonalities such as Extrav ersion and Conscien tiousness are more positive and sociable and hence less affected b y daily stress. In our study , Big Five p ersonality traits w ere measured b y asking sub jects to answer the online versio n of the 44 questions Big Five questionnaire develo p ed by John et al. [25], by means of 5-p oin t likert scales. The scores on the fiv e traits were the av erage ov er the raw scores (in v erted when needed) of the items p ertaining to each trait. 3.3 W eather data The question about the relationship b et w een moo d, health and wea ther has b een extensiv ely debated [22], [41]. Studies in environmen tal psychology inv estigated the role of weather as a stressor and show ed significant effects of temperature, hours of sunshine and humidit y on mo od [24], [41]. A large- scale study b y F aust and colleagues [15] on 16,000 students in Switzerland sho wed an asso ciation b et w een weather and sleep disorders, depressed moo d and irritabilit y . More re- cen tly , Denissen et al. [10] inv estigated the effects of six daily w eather v ariables (temperature, wind pow er, sunligh t, pre- cipitation, air pressure, photoperio d) on three moo d v ari- ables (p ositiv e affect, negativ e affect, and tiredness). Their results revealed main effects of temp erature, wind p o w er, and sunligh t on negativ e affect, while sunlight had also a main effect on tiredness and mediated the effects of precip- itation and air pressure on tiredness. In our exp erimen ts, we used the following weather v ari- ables: (i) mean temp erature, (ii) pressure, (iii) total pre- cipitation, (iv) h umidit y , (v) visibility and (vi) wind sp eed (measured in m/s) metrics. The source w eather data w ere collected from W olfram Alpha 2 . All w eather metrics are computed on a daily scale for the same day that is under in v estigation and the source data are extracted from the Boston area w eather stations (e.g. KBOS) located on the same relativ e elev ation as the campus where the data collec- tion was p erformed. 4. FEA TURE EXTRA CTION Based on previous works that characterize so cial interac- tions b y means of mobile phone data and use so cial intera c- tions data to predict p eople’s behaviors, states [2, 31], and traits [6, 9, 37], we derived the 25 call and sms basic fea- tures rep orted in T able 2 and the 9 prox imit y basic features reported in T able 3. F or each basic feature, we calculated second order fea- tures, such as mean, median, min, max, 99%, 95% quan- 2 h ttp://www.w olframalpha.com tiles, quantiles corresp onding to 0.5, 1, 1.5 and 2 standard deviations (applying Chebyshev’s inequality), v ariance and standard deviation functions. Moreov er, for each basic fea- ture we calculated the same functions as ab o v e for 2 and 3 da ys bac kward-mo ving window to accoun t for the possibility that past even ts influenced the curren t stress state. In the follo wing subsections we will describ e more in detail the 25 call and sms basic features and the 9 proximit y basic features. 4.1 Call and Sms Featur es The features reported in T able 2 fall under four broad categories: (i) general phone usage, (ii) active b ehaviors, (iii) regularity , and (iv) diversit y . T able 2: List of Basic F eatures General Phone Usage 1. T otal Number of Calls (Outgoing+Incoming ) 2. T otal Number of Incoming Calls 3. T otal Number of Outgoing Calls 4. T otal Number of Missed Calls 5. Number of SMS receiv ed 6. Number of SMS sen t Diversit y 7. Number of Unique Con tacts Called 8. Number of Unique Con tacts who Called 9. Number of Unique Con tacts Comm unicated with (Incoming+Outgoing) 10. Number of Unique Con tacts Associated with Missed Calls 11. Entrop y of Call Contacts 12. Call Contacts to Interactions Ratio 13. Number of Unique Con tacts SMS received from 14. Number of Unique Con tacts SMS sent to 15. Entrop y of SMS Contacts 16. Sms Contacts to Interactions Ratio Active Behaviors 17. Percen t Call During the Night 18. Percen t Call Initiated 19. Sms resp onse rate 20. Sms resp onse latency 21. Percen t SMS Initiated Regularity 22. Average In ter-ev ent Time for Calls (time elapsed b etw een tw o ev ents) 23. Average In ter-ev ent Time for SMS (time elapsed b etw een tw o ev ents) 24. V ariance Inter-ev en t Time for Calls (time elapsed betw een t wo events) 25. V ariance Inter-ev en t Time for SMS (time elapsed betw een t wo events) F eatures for gener al phone usage consist of: the total num- ber of outgoing, incoming and missed calls and the total n um b er of sent/recei ved sms. Moreov er, we also computed the following ratios: outgoing to incoming calls, missed to (outgoing + incoming) calls, sen t to received sms. Then, we captured the active b ehaviors of an individual computing the following features: (i) percentage of calls done during the nigh t, (ii) p ercen tage of initiated calls dur- ing the night, (iii) the sms resp onse rate, (iv) the sms re- sponse latency , and (v) the percentage of initiated sms. In particular, we consider a text from a user (A) to be a re- sponse to a text received from another user (B) if it is sent within an hour after user A received the last text from user B. The resp onse rate is the p ercen tage of texts people re- spond to. The latency is the median time it takes p eople to answ er a text. Note that b y definition, latency will be less or equal to one hour. Diversity and r egular ity hav e been shown to b e impor- tan t for the characterization of differen t facets of human behavior. In particular, ent ropy , used as a measure of diver- sit y , has been successfully applied to predict mobility [47], spending patterns [28, 46], online b eha vior [44] and person- alit y traits [37]. Concerning r e gularity features, we mea- sured the time elapsed b et w een calls, the time elapsed b e- t ween sms exc hanges and the time elapsed b et ween call and sms. More precisely , w e consider b oth the av erage and the v ariance of the int er-even t time of one’s call, sms and sum thereof (call+sms). Noticeably , in fact, even when t w o users ha v e the same inter-ev en t time for b oth call and sms, that quan tit y can b e different for their sum. Diversity measures ho w evenly an individual’s time is dis- tributed among others. In our case, the div ersit y of user behavior is addressed by means of three kinds of features: (i) entrop y of con tacts, (ii) unique con tacts to interactions ratio, (iii) num ber of unique contacts, all computed b oth on calls and on sms. In particular, the entrop y of an individual is the ratio b et w een his/her total n um ber of contacts and the relative frequency at whic h he/she interacts with them. The more one interacts equally often with a large num ber of contacts, the higher the entrop y will b e. F or entrop y cal- culation, we applied Mil ler-Madow c orr e ction [35], which is explained in Equation 1. ˆ H M M ( θ ) ≡ − p X i =1 θ M L,i log θ M L,i + ˆ m − 1 2 N , (1) where ˆ m is a n um ber of bins with nonzero θ -probabilit y . The likel iho od function is given as the pro duct of probabil- it y density functions P ( θ ) = f ( x 1 ; θ ) f ( x 2 ; θ ) · · · f ( x n ; θ ) for a random sample X 1 , · · · , X n . θ M L is the maximum like- lihoo d estimate of θ , which maximizes P ( θ ). Miller-Madow correction was applied, dealing with the data quality prob- lems, to get bias-corrected empirical entrop y estimate. 4.2 Proximity F eatur es Starting from the Blueto oth hits collected, we filtered out all the cases with RS S I < 0. F rom the filtered Bluetooth pro ximit y data w e extracted the following basic Blueto oth pro ximit y features (T able 3). In this case, the extracted fea- T able 3: List of Basic Blueto oth Proximit y F eatures General Bluetooth Proximit y 1. Number of Bluetooth IDs 2. Times most common Bluetooth ID is seen 3. Blueto oth IDs accounting for n% of IDs seen 4. Blueto oth IDs seen for more than k time slots 5. Time interv al for whic h a Bluetooth ID is seen 6. Entrop y of Bluetooth con tacts Diversit y 7. Contacts to interactions ratio Regularity 8. Average Bluetooth intera ctions in ter-even t time (time elapsed betw een tw o ev ents) 9. V ariance of the Bluetooth interactions inter-ev ent time (time elapsed betw een tw o ev ents) tures fall under three broad categories: (i) general p rox imit y information, (ii) div ersit y , and (iii) regularity . As for call and sms, we applied Miller-Madow correction for entrop y calculation. 5. METHODOLOGY W e form ulated the automatic recognition of daily stress as a binary classification problem ( “not stressed” vs “stressed” ), with labels 0 for “not stressed” and lab el 1 for “stressed” . The t wo classes included all the cases with scores < = 4 and scores > 4, respectively . The sizes of the resulting tw o classes are 36.16% for ” stressed” and 63.84% for ” not stressed” . The inclusion of the cases with stress=4 in the 0 class meant to pro vide a more clearcut distinction b etw een the “stressed” and the “non-stressed” cases. The data set was then randomly split in to a training (80% of data) and a testing (20% of data) dataset, ca refully a void- ing that data for the same sub jects app eared in both the training- and in the test-set. In order to accelerate the con- v ergence of the models, w e normalize d eac h dimension of the feature v ector [4]. Additionaly , we also used a leav e-one- sub ject-out cross-v alidation strategy . Hence, 111 mo dels for eac h p ersonalit y trait were trained on 110-sub ject subsets, ev aluating them against the remaining ones and finally av er- aging the results. The results obtained are not significantly differen t from the ones obtained using the random split 80% vs 20%. In the rest of the paper, we will discuss only the results obtained with the random split 80% vs 20%. 5.1 F eature Selection As an initial step, we carried out a Pe arson c orr elation analysis to visualize and b etter understand the relations be- t ween v ariables in the feature space. W e found quite a large subset of features with strong mutual correlations and an- other subset of uncorrelated features. Hence, there was room for feature space reduction. W e excluded using princip al c omp onent analysis (PCA) b ecause the transformation it is based on pro duces new v ariables that are difficult to in ter- pret in terms of the original ones making the interpretation of the results more complex. Therefore, we turned to a pip elined variable sele ction ap- proac h, based on fe atur e r anking and fe atur e subset selec - tion , which was perfomed using only data from the training set. The metric used for feature ranking w as the mean de- crease in the Gini c o efficient of ine quality . This choice was motiv ated b ecause it outp erformed other metrics, such as m utual information, information gain and chi-square statis- tic with an av erage impro ve ment of appro ximately 28.5%, 19% and 9.2% resp ectively [45]. The Gini co efficient ranges betw een 0, expressing p erfect equality in predictive pow er and 1, expressing maximal inequality in predictive p o w er. The feature w ith maximum mean decrease in Gini coefficient is exp ected to hav e the maximum influence in minimizing the out-of-the-bag error. It is known in the literature that minimizing the out-of-the-bag error results in maximizing common p erformance metrics used to ev aluate mo dels ( e.g. accuracy , F1, AUC, etc.) [51]. The feature selection pro cedure produced a reduced subset of 32 features from an initial p ool of about 500 features. Hence, we obtained a low-dimensional feature space that mak es our approac h efficient to implement into mobile and m ultimedia applications. 5.2 Model Building W e trained a v ariety of classifiers: (i) an ensemble of tree classifiers based on a Random F orest algorithm [5], (ii) a Generalized Boosted Mo del (GBM) [16], (iii) Supp ort V ector Mac hines with linear and Gaussian radial basis kernel s, and (iv) Neural Netw orks. The b est solution of the classification problem was found using an ensem ble of t ree classifiers based on R andom F or est algorithm. In the rest of the pap er, we report the p erformance results only for Random F orest. Random forest algorithm pro duces a combination of sim- ple decision tree predictors, such that each tree i s dependent on the v alues of a random vector sampled independently with the same distribution for all the classification trees in the forest [5]. The decision b oundary is formed according to the margin function. Giv en an ensemble of tree classi- fiers h 1 ( x ) , h 2 ( x ) , ..., h K ( x ) and if the training set is dra wn at random from the empirical distribution of the random v ector Y , X the margin function is defined as: mg ( X , Y ) = av g k I ( h k ( X ) = Y ) − max j != Y av g k I ( h k ( X ) = j ) , (2) where I ( · ) is the c haracteristic function. The margin func- tion measures the distance b et w een the av erage votes at ( X , Y ) for the right class and the av erage vote for any other class. F or this mo del the generalization error function is: P E ∗ = P X ,Y ( mg ( X , Y ) < 0) , (3) where P X ,Y is the probabilit y o v er h X , Y i space. F or any ev en t A ⊂ Ω of the feature space the characteristic function I ( · ) of A is: I A ( x ) =  1 ⇐ ⇒ ( x ⊂ A ) 0 other wise   1 ⇐ ⇒ ∃ x 0 other w ise  (4) Random F orests classifiers were trained with a stepwise increase of the n um ber of trees equal to the upp er limit of 2 11 . Optimal num b er of trees for mo del generalization as measured by mean misclassification rate for 10-fold cross- v alidation strategy is estimated to b e 112 trees. In order to find the final mo del, we trained a n umber of models and selected the b est one based on κ metrics for the 10-fold v alidation strategy . The Cohen’s κ measures pair- wise agreement among a set of functions which are mak- ing classification decisions with correction for an exp ected c hance agreement [7]: κ = P ( A ) − P ( E ) 1 − P ( E ) (5) κ = 0 if there is no agreemen t more than expected b y chanc e follo wing the empirical distribution; while κ = 1 when there is a max agreement. κ is a state-of-the-art statistics about ho w significan tly the classification model is differen t from c hance. Importantly , κ is a more robust measure than the simple p ercen t agreemen t, given that it takes into accoun t c hance agreement occurring without being to o conserv ative. During the learning and mo del selection pro cess w e used a random sampling with replacemen t to generate a new set of data for eac h fold from the basic training set, and follo w ed lea v e-one-out 10-fold cross v alidation sc heme. W e adopted this strategy in order to preven t data o v erfitting and to deal with p otential data loss in cases where calls, sms and Blue- tooth proximit ies existed in the real world but were not reg- istered b y the smartphone logger softw are. Our structur al risk minimization , as opp osed to empirical risk minimiza- tion solution to prev en t data ov erfitting, w as incorp orated b y working with a regularization p enalt y into the learning process, balancing the mo del’s complexit y against training data fitting and by sampling the model training sets in such a wa y that they mimic the empirical distributions without most probable erroneous outliers. Model parameter estimation selection w as done iteratively on the basis of our exploratory analysis, inferred knowledge of the relationships b etw een v ariables and model p erfor- mance metrics ( κ and Accuracy). Confounding v ariables are iden tified but not remo ved from the dataset during training and test phases. 6. EXPERIMENT AL RESUL TS The performance metrics used to ev aluate our approac h are: accuracy , κ , sensitivity , and specificity . The recognition model based on random forest algorithm shows 90.68% ac- curacy on the training set and 72.28% accuracy on the test set. In T able 4 we pro vide the final stress recognition mo del performances on the test set along with their statistical sig- nificance [17]. Metric V alue Accuracy 0.7228 95% CI (0.7051, 0.7399) No Information Rate 0.6384 P-V alue [Acc > NIR] < 2.2e-16 Kappa 0.3752 Sensitivit y 0.5272 Specificity 0.8335 ’P ositiv e’ Class “stressed” T able 4: Recognition Mo del P erformance Metrics Information ab out accuracy and κ metrics distribution us- ing 10-fold cross v alidation strategy is provi ded in T able 5. As w e can see, the distribution of the estimated performance metrics do es not v ary substantially among folds, signaling a goo d generalization despite the p ossible existence of hetero- geneous data in eac h fold and the “noise” coming from the resampling pro cedure. Accuracy Kappa Min. 0.6959 0.2995 1st Qu. 0.7156 0.3535 Median 0.7282 0.3817 Mean 0.7232 0.3684 3rd Qu. 0.7312 0.3869 Max. 0.7404 0.4010 T able 5: 10-fold Cross-V alidation Metrics W e also compared our approach based on combining m ul- tiple indicators with simpler approaches using as predictors (i) only p ersonality traits, (ii) only weather conditions, (iii) only activities inferred from mobile phone data, (iv) a com- bination of personality traits and w eather conditions, (v) a com bination of personality traits and activities inferred from mobile phone data, (vi) a com bination of weather conditions and activities inferred from mobile phone data. T able 6 re- ports accuracy , κ , sensitivit y , sp ecificit y and F1 for each approac h. In this table we also rep ort the p erformance of (vii) a simple ma jorit y classifier, which alw a ys returns the ma jority class as prediction (accuracy = 63.84%). Finally , w e also ran exp erimen ts with three classes (” not stressed” , ” neutral” , ” stressed” ), with labels − 1 for ” not stressed” , la- bel 0 for ” neutral” , and label 1 for ” stressed” . The class ” not stressed” included all the cases with scores < 4, the class ” neutral” included all the cases with scores = 4, and the class ” stressed” included all the scores > 4. The sizes of the resulting three classes are 42.83% for ” not stressed” , 20.98% for ” neutral” , and 36.15% for ” stressed” . The global accuracy obtained b y our multifactorial model, 59.57%, significantly outperformed the p erformance of simple ma jority classifier, whic h alwa ys returns the class ” not stressed” as prediction. 7. DISCUSSION The comparison among the p erformance of the v arious models in T able 6 provides con vincing evidence that none of the features s ets (personality , weather, smartp hone activit y) considered alone is endo wed with a goo d enough predictive pow er. This conclusion applies also to pairwise combina- tions of the same features sets to the extent that neither personality+smartphone activity , nor p ersonality+w eather, nor weather+smartphone activit y do any b etter than the ma jority classifier (accuracy=63.84%). Interestingly , signif- ican t improv emen ts ov er the latter can only b e obtained by the simultaneous usage of the three features sets: our fi- nal mo del based on a Random F orest classifier using 32- dimensional feature vectors obtained a 72.28% accuracy for our 2-class classification problem. As p oin ted out in Section 2, some recent works hav e used mobile phones data for stress recognition [1, 42]. Bauer and Luk o wicz [1] reported a 53% of accuracy in detecting the transition from stressful p erio ds (a tw o w eek exam session) to non-stressful p eriods (tw o weeks after the exam session). Our multifactorial approac h outp erforms the approach pro- posed by [1] altough a direct comparison may be not ad- equate given the different task. More recently , Sano and Picard rep orted an accuracy p erformance of 75% using a com bination of features from mobile phones and more ob- trusiv e wearable sensors. Ho wev er, the limited n um ber of sub jects (only 18) and the limited num ber of days (only 5) mak e the results preliminary . Other approac hes used video and audio features for stress recognition [18, 32]. F or in- stance, StressSense, an application for stress detection from h uman voice, achiev ed a 76% of accuracy in outdo or envi- ronmen ts. How ever, this metho d dep ends on sound qualit y and it ma y p ose priv acy concerns for p eople perceiving v oice recording and analysis as a threat to their priv acy . Hence, our performance of 72.28 shows that the prop osed multifac- torial approach is a reliable and less obtrusive alternative. An in v estigation of th e most important pred ictors of daily stress rev eales interesting asso ciations. T able 1 rep orts the 32 features selected and used in our model rank ed by their mean reduction in accuracy . All the personality traits con- tributes significantly in predicting the daily stress v ariable. These results are interesting b ecause the previous studies in social psychology focused their analyses mainly on the asso- ciations b et w een stress and Neuroticism, Extrav ersion and Conscien tiousness. Instead, our work sho ws the imp ortan t con tribution play ed also by Agreeableness and Op enness to Experience to the automatic classification of daily stress. Moreo v er, these results op en us the p ossibilit y of creating a m ulti-step stochastic model in whic h we first estimate the personality and then w e use those estimates as indep enden t v ariables for the daily stress recognition problem. Our cur- ren t approach uses self-rep orted information on personality and this strategy could b e a limitation for scaling to larger sample of users. Ho wev er, recent studies sho w ed that p er- sonalit y traits may be reliable reco gnized from mobile phone data [6, 9, 37, 48]. With regard to w eather, we found confirmation for the association b etw een temp erature and stress. Moreov er, sig- T able 6: Mo del Metrics Comparison for F eature Subsets Model Accuracy Kappa Sensitivity Specificity F1 Our Multifactorial Mo del 72.28 37.52 52.72 83.35 57.89 Baseline Majority Classifier 63.84 0.00 100.00 0.00 0.00 W eather Only 36.16 0.00 100.00 0.00 0.00 Personalit y Only 36.16 0.00 100.00 0.00 0.00 Bluetooth+Call+Sms 48.59 6.80 73.80 34.32 50.94 Personalit y+W eather 43.55 2.96 81.90 21.83 51.20 Personalit y+Blueto oth+Call+Sms 46.40 7.01 83.17 25.57 52.88 W eather+Bluetooth+Call+Sms 49.60 -5.45 38.45 55.91 35.55 nifican t effects of other meteorological v ariables – humidit y , visibilit y , and wind sp eed – for predicting daily stress were also found. Regarding the mobile phone data, it is interest ing to note the con tribution of proximit y features. Out of the selected 32 features, 11 features are pro ximit y ones, 6 comes from call data and 6 from sms data. In particular, an interesting predictiv e role is play ed by the num b er of time interv als for whic h an id is seen. The results obtained using proximit y features seem to confirm previous findings in so cial psychol- ogy: in particular, the relev an t role pla y ed by face-to-face in teractions and by in teractions with strong ties in deter- mining the stress level of a sub ject [27]. F or sure, this result requires further in v estigation. In addition, t w o features cap- turing the en trop y in pro ximity interactions are among the selected ones. This finding seems t o confirm results av ailable in the social psyc hological literature ab out the asso ciations betw een stress and the ric hness/div ersit y of so cial interac- tions [20]. F urther confirmation to this conclusion comes from the similalry imp ortan t role play ed by entrop y-based call and sms features. The remaining selected features related to sms interac- tions are (i) the latency in replying to a text message, de- fined as the median time to answer a text message and (ii) the amount of sms communications (outgoing+incoming). 8. IMPLICA TIONS AND LIMIT A TIONS Stress has b ecome a ma jor problem in our so ciety . Ubiq- uitous connectivity , information ov erload, increased men tal w orkload and time pressure are all elements contributing to increase general stress lev els. While in some cases people ma y realize that they are undergoing stressful situations, sev ere and c hronic stress may b e more difficult to detect. Moreo v er, stress may b e considered the norm in a mo d- ern and demanding so ciety . Nonetheless, while slightly in- creased stress levels ma y b e functional for pro ductivit y , pro- longed and severe stress can b e at the source of sev eral p hy si- cal dysfunctions lik e headache, sleep or i mmu nological disor- ders, unhealthy b ehaviours such as smoking and bad eating habits, as well as of psychological and relational problems. Beside manifest so cial costs, stress also entails considerable financial c osts for our economies, whic h are estimated b y the W orld Health Organization in 300 billion dollars a year for American en terprises, and 20 billion euro for Europe ones, in terms of absenteeism and low productivity . Our tec hnology provides a cost-effective, unobtrusiv e, widely a v ailable and reliable tool for stress recognition. It detects daily stress levels with a 72.28% accuracy combining real life data from differen t sources, such as p ersonalit y traits, social relationships (in terms of calls, sms and pro ximit y in teractions), and weather data. The dev elopment of a re- liable stress recognition system is a first but essential step to w ard wellbeing and sustainable living, and its scop e can be extended to differen t areas of applicability . Providing people with a to ol capable of gathering rich data ab out real life, and transforming them into meaningful insights ab out stress levels, pav es the wa y to a new generation of con text- a wa re tec hnologies that ca n target therapists, en terpises and common citizens. This tec hnology can inform the design of automatic sys- tems for the assessment and treatment of psychological stress. With such a tool, therapists could monitor and record pa- tien ts’ daily stress levels, access longitudinal data, identify recurren t or significan t stressors and mo dulate treatment ac- cordingly . In wor k environmen ts, where stress has b ecome a seri- ous problem affecting pro ductivit y , leading to occupational issues and causing health diseases, our system could b e ex- tended and emplo y ed for early detection of stress-related conflicts and stress con tagion, and for supp orting balanced w orkloads. Awareness is a first but crucial step to moti- v ate p eople to c hange their b eha viour and take informed and concrete steps tow ard a healthy lifestyle and appropri- ate stress coping strategies. Mobile applications developed on the basis of our technology could provide feedback to in- crease people’s aw areness of their stress lev els, alerts when they reach a warning threshold, and suggest stress manage- men t and relaxation techniques when appropriate. Ho w ever, our study has also some limitations. W e can list the follo wing ones: (i) our sample comes from a p opulation living in the same environmen t. Our sub jects were all mar- ried graduate students living in a campus facility of a ma jor US univ ersity; and (ii) the non-av ailabilit y of proximit y data concerning the interaction with people not participating in the data collection, a fact that is common to many other relev ant studies and that has b een also p oin ted out by [40]. The first problem is at least partially attenuated b y the large v ariability of the sample in terms of pro v enance and cultural bac kground (in our sample w e hav e sub jects from 16 coun- tries and from all the con tinen ts), whic h can be expected to correspond to a wide palette of interaction b ehaviors that efficaciously coun terbalance the effects of living-place homo- geneit y . 9. CONCLUSION The goal of this pap er was to inv estigate the automatic recognition of p eople’s daily stress from three differen t sets of data: a) p eople activit y , as detected through their smart- phones (data p ertaining to transitory prop erties of indi- viduals); b) w eather conditions (data p ertaining to tran- sitory prop erties of the environmen t); and c) p ersonalit y traits (data concerning p ermanen t dispositions of individ- uals). The problem was mo deled as a 2-w ay classification one. The results convincing ly suggest that all the three t yp es of data are necessary for attaining a reasonable pre- dictiv e pow er. As long as one of those information sources is dropp ed, p erformances drop below those of the baselines. Moreo v er, the distributional data for accuracy and κ show the robustness and generalization p o wer of our m ultifacto- rial approach. T aken together, and despite the limitations discussed ab o v e, our results not only pro vide evidence that individual daily stress can be reliably predicted, but they also p oin t to the necessit y of considering at the same time people’s transi- tory prop erties (smartphone activity), transitory prop erties of the environmen t and information ab out stable individual c haracteristics. F or the sak e of transitory individual proper- ties, mobile phone usage patterns hav e important adv antage o ver alternative metho ds: they are l ess unobtrusiv e and raise limited priv acy problems as compared to, e.g., voice analysis or the exploitation of data from physiological sensors. More- o ver , and imp ortantly , automatic stress detection based on mobile phone data can ta ke adv antage of the extensiv e usage and diffusion of these devices, it can b e applied in several real world situations and it can b e exploited for a v ariety of applications that are delivered by means of the same de- vice. F or example, applications used to inform the design of clinical decision supp ort systems or self-monitoring applica- tions of stress lev els in work settings and in other daily life situations, which allows people to identify personal stressors and enforces their proactive role in stress preven tion and managemen t. Acknowledgment This w ork was partially supported b y T elecom Itali a Seman- tics and Knowledge Innov ation Laboratory (SKIL) with re- searc h grant T. 10. REFERENCES [1] G. Bauer and P . Luko wicz. Can smartphones detect stress-related changes in the b eha viour of individuals? In Pervasive Computing and Communic ations Workshops (PERCOM Workshops), 2012 IEEE International Confer enc e on , pages 423–426. IEEE, 2012. [2] A. Bogomolo v, B. Lepri, and F. Pianesi. Happiness recognition from mobile phone data. Pro ce edings of the 2013 International Confer ence on So cial Computing (So cialCom 2013) , pages 790–795, 2013. [3] N. Bolger and E. A. Schilling. Personalit y and the problems of everyda y life: The role of neuroticism in exposure and reactivity to daily stressors. Journal of Personality , 59(3):355–386, 1991. [4] G. E. P . Box and D. R. Cox. An Analysis of T ransformations. Journal of the R oyal Statistic al So ciety. Series B (Metho dolo gic al) , 26(2):211–252, 1964. [5] L. Breiman. Random forests. Mach. L e arn. , 45(1):5–32, Oct. 2001. [6] G. Chittaranjan, J. Blom, and D. Gatica-Perez. Mining large-scale smartphone data for p ersonality studies. Personal Ubiquitous Comput. , 17(3):433–450, Mar. 2013. [7] J. Cohen. A Co efficient of Agreemen t for Nominal Scales. Educ ational and Psycholo gic al Me asur ement , 20(1):37–46, Apr. 1960. [8] S. Cohen, K. R. C., and L. U. Gordon. Me asuring str ess: A guide for he alth and so cial scientists . Oxford Univ ersit y Press, USA, 1997. [9] R. de Oliveira , A. Karatzoglou, P . Concejero Cerezo, A. Armenta Lop ez de Vicu ˜ na, and N. Oliver. T o w ards a psychographic user mo del from mobile phone usage. In CHI’11 Extende d A bstr acts on Human F actors in Computing Systems , pages 2191–2196. ACM, 2011. [10] J. Denissen, L. Butalid, L. Penk e, and M. V an Aken. The effects of weather on daily mo od: A multilev el approac h. Emotion R ese ar cher , 8(5):662–667, 2008. [11] W. Dong, B. Lepri, and A. Pen tland. Mo deling the co-ev olution of b eha viors and so cial relationships using mobile phone data. In MUM , pages 134–143, 2011. [12] C. Duggan, P . Sham, A. Lee, C. Minne, and R. Murray . Neuroticism: a vulnerabilit y marker for depression evidence from a family study . Journal of Affe ctive Disor ders , 35(3):139 – 143, 1995. [13] N. Eagle and A. Pen tland. Reality mining: sensing complex so cial systems. Personal and Ubiquitous Computing , 10(4):255–268, 2006. [14] N. Eagle, A. S. Pen tland, and D. Lazer. Inferring friendship netw ork structure by using mobile phone data. Pr o c e e dings of the National Ac ademy of Scienc es , 106(36):15274–15278, 2009. [15] V. F aust, M. W eidmann, and W. W ehner. The influence of meteorological factors on children and y ouths: A 10% random selection of 16,000 pupils and appren tices of basle city (switzerland). Acta Pae dopsychiatric a: International Journal of Child & A dolesc ent Psychiatry , 1974. [16] Y. F reund and R. E. Schapire. A decision-theoretic generalization of on-line learning and an application to bo osting. Journal of Computer and System Scienc es , 55(1):119 – 139, 1997. [17] S. Garcia and F. Herrera. An Extension on ” Statistical Comparisons of Classifiers ov er Multiple Data Sets” for all Pairwise Comparisons. Journal of Machine L e arning R ese ar ch , 9:2677–2694, Dec. 2008. [18] D. Giakoumis, A. Drosou, P . Cipresso, D. Tzov aras, G. Hassapis, A. Gaggioli, and G. Riv a. Real-time monitoring of b eha vioural parameters related to psyc hological stress. Studies in he alth te chnolo gy and informatics , 181:287, 2012. [19] M. Gonzalez, C. Hidalgo, and L. Barabasi. Understanding individual mobility patterns. Natur e , 453(7196):779–782, 2008. [20] W. J. Go ode. A theory of role strain. Americ an So ciolo gic al R eview , 25(4):pp. 483–496, 1960. [21] L. H. A. S. Gunthert, Kathleen Cimbolic; Cohen. The role of neuroticism in daily stress and coping. Journal of Personality and So cial Psycholo gy , 77:1087–1100, 1999. [22] J. Hardt and H. Gerb ershagen. No changes in mo od with the seasons: observ ations in 3000 chronic pain patien ts. A cta Psychiatric a Sc andinavic a , 100(4):288–294, 1999. [23] J. Healey and R. Picard. Detecting stress during real-w orld driving tasks using physiolo gical sensors. Intel ligent T r ansp ortation Systems, IEEE T ransa ctions on , 6(2):156–166, 2005. [24] E. How arth and M. S. Hoffman. A multidimensional approac h to the relationship b et ween mo o d and w eather. British Journal of Psycholo gy , 75(1):15–23, 1984. [25] O. P . John and S. Sriv astav a. The Big Five trait taxonom y: History , measurement, and theoretical persp ectives. In L. A. Pervin and O. P . John, editors, Handb o ok of Personality: The ory and R esear ch , pages 102–138. Guilford Press, New Y ork, second edition, 1999. [26] E. Jov anov, A. O’Donnell Lords, D. Rask ov ic, P . Cox, R. Adhami, and F. Andrasik. Stress monitoring using a distributed wireless intelligen t sensor system. Engine ering in Me dicine and Biolo gy Magazine, IEEE , 22(3):49–55, 2003. [27] D. Krackhardt. The strength of strong ties: The importance of philos in organizations. pages 216–239, 1992. [28] C. Krumme, A. Llorente, M. Cebrian, A. Pen tland, and E. Moro. The predictability of consumer visitation patterns. Scientific R ep orts , (1645), 2013. [29] N. D. Lane, E. Miluzzo, H. Lu, D. Peebles, T. Choudhury , and A. T. Campb ell. A survey of mobile phone sensing. IEEE Communic ations Magazine , 48(9):140–150, 2010. [30] N. Lathia, K. K. Rach uri, C. Mascolo, and P . J. Ren tfro w. Contextual dissonance: Design bias in sensor-based exp erience sampling metho ds. Pr o c e e dings of Ubic omp 2013 , pages 183–192, 2013. [31] R. LiKamW a, Y. Liu, N. D. Lane, and L. Zhong. Moo dscop e: building a mo od sensor from smartphone usage patterns. In Pr o c e e ding of the 11th annual international c onfer enc e on Mobile systems, applic ations, and servic es , MobiSys ’13, pages 389–402, New Y ork, NY, USA, 2013. ACM. [32] H. Lu, D. F rauendorfer, M. Rabbi, M. S. Mast, G. T. Chittaranjan, A. T. Campb ell, D. Gatica-Perez, and T. Choudhury . Stresssense: detecting stress in unconstrained acoustic environmen ts using smartphones. In Pr o c ee dings of the 2012 ACM Confer enc e on Ubiquitous Computing , UbiComp ’12, pages 351–360, New Y ork, NY, USA, 2012. ACM. [33] A. Madan, M. Cebri´ an, S. T. Moturu, K. F arrahi, and A. Pen tland. Sensing the ” health state” of a comm unit y . IEEE Pervasive Computing , 11(4):36–45, 2012. [34] D. Ma jo e, P . Bonhof, T. Kaegi-T rachsel, J. Gutknech t, and L. Widmer. Stress and sleep quality estimation from a smart wearable sensor. In Pervasive Computing and Applic ations (ICPCA), 2010 5th International Confer enc e on , pages 14–19, 2010. [35] G. Miller. Note on the bias of information estimates. Information the ory in psycholo gy: Pr oblems and metho ds , 2(95):100, 1955. [36] G. Miller. The smartphone psychology manifesto. Persp e ctives on Psycholo gic al Scienc e , 7(3):221–237, 2012. [37] Y.-A. Mont joy e, J. Quoidbach, F. Robic, and A. Pen tland. Predicting p ersonality using nov el mobile phone-based metrics. In A. Greenberg, W. Kennedy , and N. Bos, editors, So cial Computing, Behavior al-Cultur al Mo deling and Pr e diction , volu me 7812 of L e ctur e Notes in Computer Science , pages 48–55. Springer Berlin Heidelb erg, 2013. [38] A. Muaremi, B. Arnrich, and G. T r ¨ oster. A survey on measuring happiness with smart phones. In 6th International Workshop on Ubiquitous Health and Wel lness (UbiHe alth 2012) , 2012. [39] K. Plarre, A. Raij, S. Hossain, A. Ali, M. Nak a jima, M. Al’absi, E. Ertin, T. Kamarck, S. Kumar, M. Scott, D. Siewiorek, A. Smailagic, and L. Wittmers. Contin uous inference of psyc hological stress from sensory measurements collected in the natural environmen t. In Information Pro cessi ng in Sensor Networks (IPSN), 2011 10th International Confer enc e on , pages 97–108, 2011. [40] D. Quercia, J. Ellis, L. Capra, and J. Crow croft. T racking ” gross communit y happiness” from tw eets. In CSCW , pages 965–968, 2012. [41] J. L. Sanders and M. S. Brizzolara. Relationships betw een weather and mo od. The Journal of Gener al Psycholo gy , 107(1):155–156, 1982. [42] A. Sano and R. W. Picard. Stress recognition using w earable sensors and mobile phones. Pr o c e e dings of the 2013 Humaine Asso ciation Confer enc e on Affe ctive Computing and Intel ligent Inter action , pages 671–676, 2013. [43] K. R. Sche rer, D. Grandjean, T. Johnstone, G. Klasmeyer, and T. B ¨ anziger. Acoustic correlates of task load and stress. In INTERSPEECH , 2002. [44] R. Sinatra and M. Szell. Entrop y and the predictabilit y of online life. Entr opy , (16):543–556, 2014. [45] S. R. Singh, H. A. Murthy , and T. A. Gonsalves. F eature selection for text classification based on gini coefficient of inequality . Journal of Machine L e arning R ese ar ch-Pr oc ee dings T r ack , 10:76–85, 2010. [46] V. K. Singh, L. F reeman, B. Lepri, and A. S. P en tland. Predicting sp ending b eha vior using socio-mobile features. In So cial Computing (So cialCom), 2013 International Confer enc e on , pages 174–179. IEEE, 2013. [47] C. Song, Z. Qu, N. Blumm, and A. Barabasi. Limits of predictabilit y in human mobility . Scienc e , (327):1018–1021, 2010. [48] J. Staiano, B. Lepri, N. Aharony , F. Pianesi, N. Seb e, and A. Pen tland. F riends don’t lie: inferring personality traits from so cial netw ork structure. In Pr o c e e dings of the 2012 A CM Confer enc e on Ubiquitous Computing , pages 321–330. ACM, 2012. [49] A. Stop czynski, V. Sek ara, P . Sapiezynski, A. Cuttone, M. M. Madsen, J. E. Larsen, and S. Lehmann. Measuring large-scale so cial netw orks with high resolution. PL oS ONE , (9), 2014. [50] J. Suls, P . Green, and S. Hillis. Emotional reactivity to everyda y problems, affective inertia, and neuroticism. Personality and So cial Psycholo gy Bul letin , 24(2):127–136, 1998. [51] E. T uv, A. Borisov, G. Runger, and K. T orkkola. F eature selection with ensemb les, artificial v ariables, and redundancy elimination. The Journal of Machine L e arning R ese ar ch , 10:1341–1366, 2009. [52] M. V ollrath and S. T orgersen. Personalit y types and coping. Personal. Individ. Differ. , 29:367–378, 2000.