A Non-negative Matrix Factorization Based Method for Quantifying Rhythms of Activity and Sleep and Chronotypes Using Mobile Phone Data
Talayeh Aledavood, Ilkka Kivimäki, Sune Lehmann, Jari Saramäki
AA Non-negative Matrix Factorization Based Methodfor Quantifying Rhythms of Activity and Sleep andChronotypes Using Mobile Phone Data
Talayeh Aledavood , Ilkka Kivim ¨aki , Sune Lehmann , and Jari Saram ¨aki School of Interactive Computing, Georgia Institute of Technology Department of Computer Science, Aalto University, Espoo, Finland Department of Applied Mathematics and Computer Science, Technical University of Denmark, Kongens Lyngby,Denmark The Niels Bohr Institute, University of Copenhagen, Copenhagen, Denmark * [email protected] ABSTRACT
Human activities follow daily, weekly, and seasonal rhythms. The emergence of these rhythms is related to physiology andnatural cycles as well as social constructs. The human body and biological functions undergo near 24-hour rhythms (circadianrhythms). The frequency of these rhythms is more or less similar across people, but its phase is different. In the chronobiologyliterature, based on the propensity to sleep at different hours of the day, people are categorized into morning-type, evening-type,and intermediate-type groups called chronotypes . This typology is typically based on carefully designed questionnaires ormanually crafted features drawing on data on timings of people’s activity. Here we develop a fully data-driven (unsupervised)method to decompose individual temporal activity patterns into components. This has the advantage of not including anypredetermined assumptions about sleep and activity hours, but the results are fully context-dependent and determined by themost prominent features of the activity data. Using a year-long dataset from mobile phone screen usage logs of 400 people,we find four emergent temporal components: morning activity, night activity, evening activity and activity at noon. Individualbehavior can be reduced to weights on these four components. We do not observe any clear emergent categories of peoplebased on the weights, but individuals are rather placed on a continuous spectrum according to the timings of their activities.High loads on morning and night components highly correlate with going to bed and waking up times. Our work points towardsa data-driven way of categorizing people based on their full daily and weekly rhythms of activity and behavior, rather thanfocusing mainly on the timing of their sleeping periods. a r X i v : . [ c s . C Y ] S e p Introduction
Human lives are defined by rhythms of different frequency: daily, weekly, seasonal, and annual, among others. The mostprominent rhythms in our lives are rooted in the day-night cycle . From body cell functions to social activities and interactions,many aspects of human lives follow diurnal rhythms . Sleep-wake cycles are regulated by circadian rhythms which are internalbody processes . Sleep is important to our health and well-being and enables us to restore physically and mentally to pursueour daily activities . Circadian rhythms have a similar near 24-hour length in all humans, but their phase varies from one personto another which results in different sleeping hours . There is a large body of literature that shows that people with later-phasecircadian rhythms are at risk of various physical and mental issues . Therefore understanding and measuring these rhythmsare crucial.In chronobiology, people are commonly divided into three categories referred to as “chronotypes” according to thedifference in the phase of circadian rhythms and the propensity to sleep at different hours. The three widely-acceptedchronotypes are morning-type, evening-type, and intermediate-type . Chronotypes are often measured with one of the variousavailable questionnaires which have been developed for this purpose since the 1970s . While chronotype questionnaires arewidely used, they come with some shortcomings. First, like any other questionnaire they rely on a person’s description of theirbehavior rather than measuring the behavior in situ . Also, most of them have cut-off boundaries, which might not be appropriatefor the population under study. For example, the most well-known chronotype questionnaire, the Morningness-EveningnessQuestionnaire (MEQ) , was originally designed based on a cohort of adults with ages between 18 to 32 years old
10 1 .In this work, instead of using questionnaires relying on people’s memories, we measure their sleep and activity patternsbased on digital traces that are recorded in an accurate and unobtrusive manner from mobile phones. Instead of matching oursubjects’ activity patterns to another cohort, we use Non-negative Matrix Factorization (NMF) to automatically learn thedominant patterns in the data. We use this unsupervised method to reduce the dimensionality of the data and to extract fourdominant rhythms within the population’s activity patterns. This approach minimizes the a priori assumptions on the sleep andactivity hours of the study participants and lets these patterns emerge from the data. Using NMF, we reduce each person’srhythm into weights on the extracted components. This way we can measure the typical timings of a person’s sleep and activitywithout having the common biases of questionnaires or enforcing arbitrary cut-off thresholds to the data. We also do not requireany prior knowledge about the population under study.Our work is part of a rapidly growing field in which digital traces that people leave behind, such as data produced by mobilephones, are used as a proxy for human activity and to measure temporal patterns of their behavior . These traces allow usto measure people’s behavioral patterns unobtrusively in the wild. Continued evolution of mobile phones and especially theubiquity of smartphones has opened up possibilities to collect high-resolution data from individuals and study their behaviorat a level of detail not previously possible . In the past years, multiple studies have gathered detailed data from individualswith the aim of studying and quantifying their behavior . More recently, high-resolution data from phones have been usedto study sleeping and resting patterns of people . Our work introduces a new method to study these patterns that can beapplied to data from any of these (and other similar) studies.In this work, we use data from the Copenhagen Networks Study, which has gathered data from around 1000 universitystudents for over two years . In this study, participants were given identical phones, and, with their consent, data on theirphone usage and behavior were collected and shared with researchers. We used timestamps of events where the phone screenturns on (screen-on events) to measure dominant activity and sleep rhythms of the study participants. In this work, we alsostudy in detail how the component weights of the decomposition of rhythms correlate with one another within the population.Furthermore, we investigate how these weights correlate with sleep and wake-time hours as well as the sleep duration fordifferent people. Out of over 800 students who have provided data to the Copenhagen Networks Study during the year 2014, we included 400in this analysis after filtering out individuals with low data quality (see section 3 for details on data pre-processing). We usescreen-on events as a proxy of the times that the person is active throughout the day. While this measure does not capture alltypes of activity, it is, for example, a good measure of when a person awake and using the phone. By studying the temporalactivity patterns of phone usage over a long period of time, we can form a picture of the behavioral patterns of the person. Weaggregated the weekly data for each person over the course of one year into 7 × =
168 one-hour bins (Monday morningto Sunday night). For each person we normalize the sum of the time series to unity and derive the activity rhythm of theperson. The average of the activity rhythms of all the 400 individuals is shown in Fig. 1. The average rhythm shows clear daily There are also questionnaires such as the Munich ChronoType Questionnaire (MCTQ) which are not based on pre-defined thresholds and assume thatchronotype is a continuous variable rather than a categorical one . eriodicity and a lower activity level during weekends, reflecting the common pattern in individual rhythms. The lowest levelof activity coincides with night times. Mo Tu We Th Fr Sa Su
Time F r a c t i o n o f E v e n t s Figure 1.
The average normalized activity rhythm of everybody included in the analysis. The x-axis shows a week (hourlybins). The red lines mark the beginning of each day and the grey dashed lines mark 12:00 noon on each day.We used NMF as an unsupervised method to extract the four main components of the activity rhythms of the 400 studentsincluded in the analysis (for details on the choice of method and number of components see section 3.3). The four NMFcomponents are displayed in Fig. 2. The components follow a daily periodicity, peaking at different hours of day. The timingsof the four peaks are associated with morning, noon, evening, and night.Out of the four components, the morning and noon components show lower levels of activity in the weekends compared toweekdays. The evening activity is lower on Friday and Saturday evenings and the night activity remains approximately similaron all days of the week. Taken together, as other components than nighttime activity are lower in the weekends, there is a shiftin activity patterns of people towards later hours in the weekend days, which is consistent with typical behavior of adults whooften go to bed earlier during the week and stay up later and catch up on sleep at the end of the week .Fig. 2, right column, shows the distribution of weights on different components. The distribution of weights on the noon andevening components resemble a normal distribution centered around a mean value. However, the distributions of the weights onthe morning and night components are more skewed so that the majority of people have low weights.The weights in decompositions on the four NMF components are not independent. Fig. 3 explores the correlationsbetween the weights of the decomposition of individuals’ activity rhythms on the components. While the correlation betweensome weights is mild (morning and evening, morning and night, noon and evening), some of the components show higheranti-correlations. For example, the Pearson correlation coefficients for weights on night and evening components is -0.69,which means people do not typically have high weights on both of these components simultaneously. Fig. 3, top right, showsthe weight for each person on the four components. In this plot we see that individuals cannot clearly be grouped into separategroups based on the activity rhythms and their weights on the four components form a continuous spectrum. This spectrumhowever has an elongated form (rather than a cloud) and shows that high weight of the evening component coincides with lowvalues on all other components.Sleep is a substantial part of people’s days and therefore one of the main determinants of how the daily rhythm of a personlooks. Sleeping times are represented in the daily rhythms with long periods of inactivity. While the NMF components and theweights for different people on them are associated with their hours of activity, they are also equally associated with their hoursof inactivity and sleep. Next, we calculate the most frequent going-to-sleep time, wake-up time, mid-sleep time, and sleepduration directly from the activity data (rather than using the NMF components or the daily rhythms). We then examine thecorrelation of these variables with weights on different components (see Fig. 4). In Section 3.4 the derivation of these variablesis explained.Sleep time shows a moderate anti-correlation with the weight on the evening activity and a high correlation with the weighton the night activity component (see table 1). This implies that going to sleep later is associated with higher activity on thephone on late hours of the day. This finding is non-trivial because staying up later does not necessarily mean that the person hasto be more active on the phone. For calculation of sleep times, even one data point at a late hour (e.g. while setting up the alarmbefore going to bed) would be sufficient for our algorithm to determine a late sleep time. However, for a high weight on thenight activity component the person has to have many data points (screen-on activity) at late hours. This also implies that thosewho have a high evening activity tend to be early sleepers, while those with high night activity levels are late sleepers. Similarly,wake-up time shows a moderate anti-correlation with the morning activity component, meaning that an earlier wake-up time is o Tu We Th Fr Sa Su Morning Activity N u m b e r o f i n d i v i d u a l s Weight on Morning Activity Component
Mo Tu We Th Fr Sa Su
Noon Activity N u m b e r o f i n d i v i d u a l s Weight on Noon Activity Component
Mo Tu We Th Fr Sa Su
Evening Activity N u m b e r o f i n d i v i d u a l s Weight on Evening Activity Component
Mo Tu We Th Fr Sa Su
Night Activity N u m b e r o f i n d i v i d u a l s Weight on Night Activity Component
Figure 2.
The four NMF components (left) and the distributions of weights of individuals’ activity rhythm decompositions onthe four components (right). The components are detected in the data without using any external information on the dailyrhythms, and we have named them after the detection as
Morning activity , Noon activity , Evening activity , and
Night activity for clarity. These names are chosen based on the times of the day when the activity peaks for each component.associated with higher activity level on the phone in the morning. The third and fourth columns in Fig. 4 show the mid-sleeppoint and sleep duration vs weight on different components. The mid-sleep time is the middle point between sleep time and igure 3.
Correlation between weight on one component vs. the other components. The color depicts the number of peoplewith values within each square. On top of each panel, the value of the Pearson correlation coefficient (left) and thecorresponding p-value (right) are shown. The highest (anti)-correlations are between the Evening and Night components (-0.69)and Morning and Noon components (-0.44), meaning that people tend to be more active in either of each the two componentsfor each pair. In the top right 3D plot, each dot represents one person so that the weights on three components (morning,evening and night) are the coordinates in the 3D space and the weight on the fourth component (noon) is represented by thecolor of the dot. Despite the existence of some outliers, people tend to form a spectrum within this space. High values ofweights for Evening activity coincide with low weights on all other dimensions. In this plot the sum of the four weights foreach person is normalized to unity.wake-up time, and therefore the mid-sleep time can be the same for two people with different sleep durations. The mid-sleeptime shows moderate to high (anti-)correlations with the morning activity, evening activity and night activity components. Sleepduration shows a moderate anti-correlation with the night activity component and a moderate correlation with the noon activitycomponent.
In this work we used the same pre-processed data as in a previous study . The data used were from weeks 2–51 of the year2014. Data from weeks 1 and 52 were discarded, because the first week partially lies in 2013 and in week 52 the Christmasholidays lead to atypical temporal rhythms. There were in total N =
804 students that used their study phones during this year.We excluded study participants who did not use their phone actively or did not use it at all for part of the year. The inclusioncriteria were: (1) the person should have used the phone on 80% of the days, (2) during weeks 2–51, the participant should on M o r n i n g A c t i v i t y r: -0.221, p: 0.0 r: -0.47, p: 0.0 r: -0.424, p: 0.0 r: -0.063, p: 0.2105 N oo n A c t i v i t y r: -0.085, p: 0.0887 r: 0.21, p: 0.0 r: 0.105, p: 0.0361 r: 0.368, p: 0.0 E v e n i n g A c t i v i t y r: -0.552, p: 0.0 r: -0.184, p: 0.0002 r: -0.428, p: 0.0 r: 0.236, p: 0.0 sleep time N i g h t A c t i v i t y r: 0.721, p: 0.0 wake up time r: 0.339, p: 0.0 mid-sleep time r: 0.607, p: 0.0 sleep duration (hours) r: -0.428, p: 0.0 Figure 4.
Correlations of different NMF components with five different parameters, from right to left: sleep time, wake-uptime, mid-sleep time and sleep duration.average have 280 screen-on and screen-off events. The final number of study participants which were kept for further analysiswas N = In order to extract common temporal patterns from the data we use non-negative matrix factorization (NMF). NMF results in a(small) number of typical patterns from a dataset such that the original data can be approximated as weighted sums of thosetypical patterns. The constraint that the components must be non-negative makes interpretation of the decomposition moreintuitive than e.g. principal component analysis. Empirically, it has been found that NMF tends to produce components thatcorresponds to individual parts of a system – when faces are decomposed, the components become eyes, noses, mouths, andmoustaches .Formally the decomposition is achieved by approximating a data matrix X of non-negative elements with a product of twoother non-negative matrices as X ≈ WH T , where the dimension (i.e. the number of columns) of the factorization matrices W and H is smaller than the dimension of the data matrix. The dimension of the factorization means the number of typical able 1. Pearson correlation coefficient for different NMF components and sleep and activity parameters: sleep time, wake-uptime, mid-sleep time and sleep duration.
Sleep time Wake-up time Mid-sleep time Sleep duration r -0.221 -0.47 -0.424 -0.063
Morning activity p-value 0.0 0.0 0.0 0.2105r -0.085 0.21 0.105 0.368
Noon activity p-value 0.0887 0.0 0.0361 0.0r -0.522 -0.184 -0.428 0.236
Evening activity p-value 0.0 0.0002 0.0 0.0r 0.721 0.339 0.607 -0.428
Night activity p-value 0.0 0.0 0.0 0.0patterns, called components , that are sought from the data.The approximation above is achieved by minimization of an error between the actual data matrix and the factorization.Any meaningful error function can be used, but we used the most standard one, i.e. the squared Frobenius distance, whichgeneralizes the Euclidean distance from vectors to matrices: E = (cid:107) X − WH T (cid:107) = ∑ i , j (cid:16) x i j − ∑ k w ik h jk (cid:17) . (1)For computing NMF, we use the scikit-learn Python package , which contains an off-the-shelf implementation ofthe Hierarchical Alternative Least Squares (HALS) algorithm for minimizing the error (1) . The implementation is stochastic,because of which we always ran the algorithm with 1000 different random seeds and picked the run that resulted in the smallesterror.In more detail, we store our activity data in an N × M matrix X , where N is the number of data vectors (in our case thenumber of individuals studied) and M their dimensionality (the number of hours in a week). Non-negative matrix factorization(NMF) means an approximation of X as X ≈ WH T , where H is an M × K and W an N × K matrix, and K is the number of components sought from the data. The components arethe K column vectors of matrix H (i.e. rows of H T ). In other words, the components are simply vectors in the original dataspace, meaning, in our case, histograms of weekly mobile phone usage patterns. The contribution of each component to theapproximation of each data point is given by the elements of the other factor matrix, W . For example, the first data vector, x ,i.e. the first row of matrix X , is approximated in NMF as x = w h T + w h T + . . . + w K h T K , where h T i are the rows of matrix H T . he complete factorization can be represented graphically as a decomposition of rank-one matrices as: M (features) N (vectors) X ≈ K (weights) N W MK (components) H T = w (cid:104) h T (cid:105) + w (cid:104) h T (cid:105) + · · · + w K (cid:104) h T K (cid:105) , (2)where the i -th value of each weight vector w j indicates the contribution of component h j to the representation of data vector i in the reduced space. The extraction of dominant rhythms of the system can be approached with several different tools, including principal andindependent component analyses, factor analysis, topic modeling, as well as some functional data analysis methods. We did alsoexperiment with some of these methods, but found the results obtained with NMF most interesting and informative. In essence,we chose NMF for its interpretability (e.g. compared to component analysis methods which do not have the non-negativityconstraint) and its conceptual simplicity (over more involved methods such as topic modeling and functional data analysismethods). Also, the phone screen activity data that we analyze is non-negative, which further makes NMF a natural choice.For the optional number of components we started with 3 components which is the number of commonly used chronotypes.Using the cophenetic correlation coefficient, which is a measure of finding optimal number of components in NMF based onstability of components in different runs of the algorithm, we could see that 4 is where this number is maximized (see Fig. 5).For calculating the cophenetic correlation coefficient we used Nimfa, a Python library for NMF . We assume that the longest period of inactivity on the phone during each 24-hour period for each person is their sleep timeand calculate their sleep and wake-up times based on this assumption. This method does not give exact sleep and wake-uptimes for each and every night, but by looking at these values over a longer period, typical sleeping times of a person can beinferred . In Fig. 6, 7 days of data for one person is shown, where hours of the day when the person has been active onthe phone and the hours when they have not touched the screen are depicted. For calculating the common sleep and wake-uptimes for each person, we look at data from all days, find the longest period of inactivity and pick the first hour of the longestperiod of inactivity as the start of the sleep time. Similarly we take the first hour of activity after the long period of inactivity asthe wake-up time. For doing this, we do not go by calendar days (starting at midnight), but rather go with 24 hours starting at3pm (assuming that sleeping is most likely happening at night hours), to reduce the chance of splitting the longest period ofinactivity for the day when separating the data for consecutive days. We calculate the sleep, wake-up, and mid-sleep hours inthis way for every day and find the most common values (mode) for each distribution. Using the mode we reduce the effect ofoutlier nights on the calculations. For calculating the common sleep duration for each person we simply use the number ofhours of inactivity and find the mean of the distribution for each person.
In this work we study the sleep and activity of a group of 400 university students, using data on smartphone screen-ontimestamps. We see clear daily and weekly rhythms where there is higher levels of activity during the daytime and lowactivity levels at night. Also, we observe a delayed phase for the peak of activity and the lowest activity levels (rest times)during weekends. This holds both for individuals and group level activity patterns. Even though all patterns have a 24-hourperiodicity, the times of activity peaks and lowest activity do not coincide for everybody. This is consistent with findings fromchronobiology studies: different people have propensity to sleep at different hours of the day and the timing when they are mostactive or alert also varies.We use NMF for extracting main temporal components of the data and show that four meaningful components emerge. Theactivity peak for each of these components is at a different hour of the day (morning, noon, evening, and night). When we C o p h e n e t i c C o rr e l a t i o n C o e ff i c i e n t Figure 5.
Cophenetic correlation coefficient versus the number of componentsdecompose individuals’ activity rhythms and look into the weights on the four components, we see that these weights form acontinuous spectrum and do not exhibit clear-cut clusters.Despite clear individual differences in the phase of the circadian rhythms (which are internal rhythms), exogenous factorscan play a big role in how an individual behaves and when they go to sleep and wake up. Also, differences in cultures, climates,environment, and so on, can make a difference in who is perceived as a morning person or a night owl within a population. Forexample, waking up at 8 am in one community might result in the label ‘morning person’ but 8 am could be considered a ‘laterriser’ within another group or population. While some chronotype questionnaires try to capture these external factors, they stilldepend on data from a pool of people to set up the cut-off thresholds for different chronotypes. Our method for extractingthe most dominant activity rhythms in the system does not depend on arbitrary thresholds on activity hours derived from datafrom other populations, and it can be used for groups of almost any size. The analyses developed here can further be used tocategorize people based on their weights on different components and give a measure of how one person compares to otherpeople in the group under study. For example, we can rank people based on their weights on the morning component andlabel a person as a morning person within that cohort. However, if the available data is from too few individuals, our methodwould pick up small nuances of the rhythms and would not necessarily give robust results. Additionally, our study suggests thateven when categorizing people into groups based on their circadian patterns, it would be useful to take into account their sleepduration as well as the timing of their activity peaks during the day, in addition to their sleep timing.A future line of study would be to look into the robustness of the components by studying different population sizesand different life styles and ages. It will also be informative to use a dataset where both phone activity data and chronotypequestionnaires are available. We can compare questionnaire categories with those derived by using NMF and a clusteringmethod on the weights on the components. In addition to that, we can apply the same method on other types of data gatheredfrom phones and other wearables which have (at least) an hourly resolution. This allows us to study the behavioral patterns forother activity types and compare them across different data streams. For example, in previous work, we showed that individualstend to have persistent daily rhythms for communication across different channels (calls and text messages) but the shapes ofthese rhythms and peak hours differ . Studying rhythms of different data streams we can for example investigate how rhythmsof social activity vary from those of physical activity.Some of the previous questionnaire-based works on chronotypes have pointed out that using thresholds in these question-naires for separating different groups is rather meaningless , and it has been suggested that there is a wide and continuous a y d a y d a y d a y d a y d a y d a y Figure 6.
This figure shows presences of screen-on events in different hours and different days (each row is one day of data)from one person in the study. The higher dots show presence of activity in that hour and lower dots show the lack of it. We findthe longest period of of inactivity within the 24 hours (area filled with blue). These hours are marked as sleeping times. Thefirst hour of inactivity is recognized as (going to) “sleep time” and the first hour right after of the period of inactivity isdetermined to be the “wake-up time”. This processes repeated for all days for each person.spectrum of individual types. We show this based on a data-driven method and unobtrusive method rather than by usingquestionnaires. Our results do not support the use of strictly categorical variables such as chronotypes for describing people’sactivity and sleep habits. Considering the chronotype as a continuous variable instead might improve the results of studiesrelated to adverse health impacts of late phase of the circadian rhythms. Given the work presented here, we believe that there isa need to rethink how to quantify temporal patterns in human behavior and to extend the concept so that it takes the nuances ofthe activity level during non-sleep periods better into account.
References Foster, R. G. & Kreitzman, L. The rhythms of life: what your body clock means to you!
Exp. physiology , 599–606(2014). Panda, S., Hogenesch, J. B. & Kay, S. A. Circadian rhythms from flies to human.
Nature , 329–335 (2002). Edery, I. Circadian rhythms in a nutshell.
Physiol. genomics , 59–74 (2000). Irwin, M. R. Why sleep is important for health: a psychoneuroimmunology perspective.
Annu. review psychology (2015). Kerkhof, G. A. Inter-individual differences in the human circadian system: a review.
Biol. psychology , 83–112 (1985). Fabbian, F. et al.
Chronotype, gender and general health.
Chronobiol. international , 863–882 (2016). Antypa, N., Vogelzangs, N., Meesters, Y., Schoevers, R. & Penninx, B. W. Chronotype associations with depression andanxiety disorders in a large cohort study.
Depress. anxiety , 75–83 (2016). Romo-Nava, F. et al.
Evening chronotype as a discrete clinical subphenotype in bipolar disorder.
J. Affect. Disord. ,556–562 (2020). Adan, A. et al.
Circadian typology: a comprehensive review.
Chronobiol. international , 1153–1175 (2012). Levandovski, R., Sasso, E. & Hidalgo, M. P. Chronotype: a review of the advances, limits and applicability of the maininstruments used in the literature to assess human phenotype.
Trends psychiatry psychotherapy , 3–11 (2013). Horne, J. A. & Östberg, O. A self-assessment questionnaire to determine morningness-eveningness in human circadianrhythms.
Int. journal chronobiology (1976).
Roenneberg, T., Wirz-Justice, A. & Merrow, M. Life between clocks: daily temporal patterns of human chronotypes.
J.biological rhythms , 80–90 (2003). Roenneberg, T. et al.
Human activity and rest in situ. In
Methods in enzymology , vol. 552, 257–283 (Elsevier, 2015).
Cichocki, A., Zdunek, R., Phan, A. H. & Amari, S.-i.
Nonnegative matrix and tensor factorizations: applications toexploratory multi-way data analysis and blind source separation (John Wiley & Sons, 2009).
Cuttone, A., Lehmann, S. & Larsen, J. E. Inferring human mobility from sparse low accuracy mobile sensing data.In
Proceedings of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous Computing: AdjunctPublication , 995–1004 (2014).
Aledavood, T. et al.
Daily rhythms in mobile telephone communication.
PloS one , e0138098 (2015). Aledavood, T. Temporal patterns of human behavior (2017).
Ureña-Carrion, J., Saramäki, J. & Kivelä, M. Going beyond communication intensity for estimating tie strengths in socialnetworks. arXiv:2007.14238 (2020).
Stopczynski, A. et al.
Measuring large-scale social networks with high resolution.
PloS one , e95978 (2014). Eagle, N. & Pentland, A. S. Reality mining: sensing complex social systems.
Pers. ubiquitous computing , 255–268(2006). Wang, R. et al.
Studentlife: assessing mental health, academic performance and behavioral trends of college students usingsmartphones. In
Proceedings of the 2014 ACM international joint conference on pervasive and ubiquitous computing ,3–14 (2014).
Mattingly, S. M. et al.
The tesserae project: Large-scale, longitudinal, in situ, multimodal sensing of information workers.In
Extended Abstracts of the 2019 CHI Conference on Human Factors in Computing Systems , 1–8 (2019).
Monsivais, D., Bhattacharya, K., Ghosh, A., Dunbar, R. I. & Kaski, K. Seasonal and geographical impact on human restingperiods.
Sci. reports , 1–10 (2017). Cuttone, A. et al.
Sensiblesleep: A bayesian model for learning sleep patterns from smartphone events.
PloS one ,e0169901 (2017). Aledavood, T., Lehmann, S. & Saramäki, J. Social network differences of chronotypes identified from mobile phone data.
EPJ Data Sci. , 1–13 (2018). Aledavood, T. et al.
Smartphone-based tracking of sleep in depression, anxiety, and psychotic disorders.
Curr. psychiatryreports , 49 (2019). Martinez, G. J. et al.
Improved sleep detection through the fusion of phone agent and wearable data streams. In , 1–6 (IEEE,2020).
Lee, D. D. & Seung, H. S. Learning the parts of objects by non-negative matrix factorization.
Nature , 788–791 (1999).
Pedregosa, F. et al.
Scikit-learn: Machine learning in Python.
J. Mach. Learn. Res. , 2825–2830 (2011). Cichocki, A. & Phan, A.-H. Fast local algorithms for large scale nonnegative matrix and tensor factorizations.
IEICEtransactions on fundamentals electronics, communications computer sciences , 708–721 (2009). Zitnik, M. & Zupan, B. Nimfa: A python library for nonnegative matrix factorization.
J. Mach. Learn. Res. , 849–853(2012). Ciman, M. & Wac, K. Smartphones as sleep duration sensors: Validation of the isensesleep algorithm.
JMIR mHealthuHealth , e11930 (2019). Borger, J. N., Huber, R. & Ghosh, A. Capturing sleep–wake cycles by using day-to-day smartphone touchscreen interactions.
NPJ digital medicine , 1–8 (2019). Aledavood, T. et al.
Channel-specific daily patterns in mobile phone communication. In
Proceedings of ECCS 2014 ,209–218 (Springer, 2016).
Acknowledgements
TA and JS acknowledge support from the Academy of Finland, project number 297195. TA also acknowledges support fromJames S. McDonnell Foundation and thanks Mikko Kivelä for providing constructive feedback on the manuscript. uthor contributions statement
SL collected the data. TA, SJ and JS designed the study. TA, SL, IK, JS conducted the study. TA analyzed the data. All authorscontributed to the writing of the manuscript.
Additional information
Competing interests