SmartGPA: How Smartphones Can Assess and Predict …campbell/smartGPA.pdfSmartGPA: How Smartphones Can Assess and Predict Academic Performance of College Students ... traits may differentially

SmartGPA: How Smartphones Can Assess and PredictAcademic Performance of College Students

Rui Wang, Gabriella Harari†Peilin Hao, Xia Zhou, and Andrew T. Campbell

Dartmouth College, The University of Texas at Austin†,{ruiwang, hao, xia, campbell}@cs.dartmouth.edu,

[email protected]

ABSTRACTMany cognitive, behavioral, and environmental factors im-pact student learning during college. The SmartGPA studyuses passive sensing data and self-reports from students’smartphones to understand individual behavioral differencesbetween high and low performers during a single 10-weekterm. We propose new methods for better understandingstudy (e.g., study duration) and social (e.g., partying) behav-ior of a group of undergraduates. We show that there area number of important behavioral factors automatically in-ferred from smartphones that significantly correlate with termand cumulative GPA, including time series analysis of ac-tivity, conversational interaction, mobility, class attendance,studying, and partying. We propose a simple model basedon linear regression with lasso regularization that can accu-rately predict cumulative GPA. The predicted GPA stronglycorrelates with the ground truth from students’ transcripts(r = 0.81 and p < 0.001) and predicts GPA within ±0.179of the reported grades. Our results open the way for novelinterventions to improve academic performance.

Author KeywordsSmartphone sensing; data analysis; mental health; academicperformance; behavioral trends

ACM Classification KeywordsH.1.2 User/Machine Systems; I.5 Pattern Recognition; J.3Life and Medical Sciences

General TermsAlgorithms, Experimentation.

INTRODUCTIONCollege life is complex. Students have to balance going toclasses and performing well academically with competing de-mands for their time and energy, such as, extracurricular ac-tivities, busy social lives, working because of financial con-cerns, being members of under-represented minorities, fitting

Permission to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor profit or commercial advantage and that copies bear this notice and the full citationon the first page. Copyrights for components of this work owned by others thanACM must be honored. Abstracting with credit is permitted. To copy otherwise,or republish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee. Request permissions from [email protected].

To be presented at ACM Conference on Ubiquitous Computing (UbiComp2015), Osaka, Japan from Sep. 7-11, 2015

in on campus (e.g., first-generation college student), dealingwith friends and families, and trying to stay active and phys-ically and mentally healthy during the ebb and flow of theterm’s workload and commitments. As a result, succeedingin a demanding educational environment is challenging.

When we think of “academic performance” we usually as-sociate it with educational outcomes best represented by astudent’s cumulative GPA. This measure typically captures acontinuous assessment of a student’s academic achievementin terms of results from assignments, quizzes, tests, midterms,and final examinations as they move through their collegeyears. Academic performance is linked to a student’s intel-lectual curiosity and ability (e.g., as measured by IQ), theirdrive and motivation, the educational environment, health,prior test results (e.g., SATs), and personality traits (e.g., con-scientiousness). There is no general agreement, however, onwhy students with similar academic capability at the same in-stitution do better or worse than one another. Furthermore, itis not clear which behavioral patterns (e.g., study habits, classattendance, time management, sleep patterns, partying behav-ior) significantly contribute to the individual differences inthe academic achievement among students. Many questionsarise. Are there distinct differences in the behavioral patternsbetween high (e.g., GPA ≥ 3.5) and low performers (e.g.,GPA ≤ 3) at the same college? If such behavioral differencesexist could we use these correlations as a basis to predict aca-demic performance? How do differing psychological states,such as personality, mental health, affect, and stress collec-tively contribute to GPA?

This paper makes the following contributions. First, we pro-pose new methods to automatically infer study (e.g., studyduration and focus) and social (e.g., partying) behaviors us-ing passive sensing from smartphones. Next, we use time se-ries analysis of these and other behavioral states derived fromthe StudentLife dataset [43], a longitudinal study of collegestudents, to find what behaviors significantly impact term andcumulative GPA. We use this behavior analysis as input tomodel the individual differences between high and low per-formers in a population of undergraduate students at Dart-mouth College. Third, in order to understand changes in be-havior students experience across the term we propose twonew behavioral metrics: 1) behavioral slope, which capturesthe magnitude of behavioral change (e.g., studying, stress)and its direction (i.e., class attendance increase or decrease)over the complete term, as well as the first and second half

1

of the term; and 2) behavioral breakpoints, which capture thespecific points in the term where a student experiences an in-dividual behavior change (e.g., increase or decrease). Thetime series analysis of student behavioral streams and thesechange metrics are used as input to correlation analysis andprediction of GPA. Finally, we propose for the first time amodel that can predict a student’s cumulative GPA using au-tomatic behavioral sensing data from smartphones. We usethe Lasso (Least Absolute Shrinkage and Selection Opera-tor) [40] regularized linear regression model as our predic-tive model. Our prediction model indicates that students withbetter grades are more conscientious, study more, experiencepositive moods across the term but register a drop in posi-tive affect after the midterm point, experience lower levels ofstress as the term progresses, are less social in terms of con-versations during the evening period, and experience changein their conversation duration patterns later in the term. Thepredicted GPA strongly correlates with the ground truth withr = 0.81 and p < 0.001, mean absolute error (MAE) of0.179, and R2 of 0.559, which measures the goodness of fitof a model and indicates that our model explained 56% of thevariance in students’ GPAs. We do this without the use ofany prior data that has been traditionally used for academicassessment [15, 22, 25], such as IQ and standardized test re-sults (e.g., SAT scores). As a result, our work opens the doorto using passive sensing data from smartphones as a predictorof academic performance.

The StudentLife study [43] looked at correlations betweenacademic performance and the averages of the low level sen-sor data (i.e., activity, conversation, and mobility) for all stu-dents across a term. StudentLife did not, however, study thetime series of each individual behavior, nor analyze the in-dividual differences between high and low performers. TheSmartGPA study advances the state-of-the-art by inferringnew behaviors, proposes new behavioral change metrics, dis-covering new correlations, and showing for the first time thatpassive smartphone sensing data can be used to accuratelypredict GPA.

RELATED WORKIn the computer science community, many efforts have beenmade to predict grades from students’ self-report data ande-learning behaviors using various machine learning mod-els. However, only a few studies have examined the rela-tionships between students’ performance and sensed behav-iors. The StudentLife study [9, 43] found correlations be-tween students’ GPAs and automatic sensing data obtainedfrom smartphones. In addition, Watanabe and colleagues[44, 45] investigated the correlations between scholastic per-formance and face-to-face interaction among students duringbreak times using a wearable sensor device. Our research ex-tends this work by building a predictive model of academicperformance based on students’ self-reports and sensed be-havior features obtained from their smartphones.

In the fields of education and psychology, much research hasfocused on identifying the predictors of college student’s aca-demic performance. Overall, the existing studies tend to fo-cus on whether students’ personality traits (e.g., extraversion,

conscientiousness), lifestyle behaviors (e.g., physical activ-ity, sociability, sleep), and mental states (e.g., stress, positiveaffect) are related to their course grades or GPAs, However,the existing research findings are primarily based on students’self-reports (i.e., one-time surveys asking about general phys-ical activity or sleep tendencies), which may be susceptible toa range of limitations (for a review see [26]). Thus, one aimof our study is to use unobtrusive and longitudinal measuresof students’ lifestyle behaviors to predict performance. Next,we review the existing research that links academic perfor-mance with students’ personality, behaviors, and emotions.

Personality. Research that examines the links between aca-demic performance and personality tends to adopt the BigFive personality framework [20], which consists of five broadtraits: extraversion, agreeableness, conscientiousness, neu-roticism, and openness [20]. Taken together, a meta-analyticreview of the literature on academic performance and per-sonality suggests that student performance is associated withagreeableness, conscientiousness, and openness to experi-ence [28]. However, some studies also find extraversionand neuroticism to be negatively associated with perfor-mance [16]. Some researchers [8] suggest that personalitytraits may differentially impact academic performance; forexample, by impairing performance in the case of neuroti-cism, or increasing academic achievement in the case of con-scientiousness.

Physical Activity. The majority of research focused onacademic performance and physical activity tends to sug-gest that grade averages are higher among students meet-ing health guidelines for moderate-vigorous physical activity[42]. However, a large cross-sectional study of health behav-iors of students in forty U.S. colleges and universities foundthat more than half of the students (58%) did not meet pub-lic health recommendations for moderate-vigorous physicalactivity [42]. Another self-report based study [19] has foundthat physical activity self-reports are not associated with stu-dent GPAs. This study, however, focused on health sciencegraduate students whose physical activity reports met or ex-ceeded the recommended levels for adults, suggesting that theresults may not generalize to other student populations. Onlya few studies that we know of have found relationships be-tween academic performance and sensor-based physical ac-tivity measures. One study measured student physical activityusing a sensor armband in addition to self-reports and foundthat changes in physical activity were associated with GPA[30]. Specifically, total step count was associated with in-creases in GPA, whereas reported moderate physical activitywas associated with decreases in GPA [30]. The StudentLifestudy measured student physical activity using accelerometerdata from students’ smartphones and found that lower overallGPAs was associated with greater average levels and variabil-ity of activity durations aggregated over a term [43].

Sociability. Behaviors related to sociability (e.g., duration orfrequency of engaging in conversation, partying, and spend-ing time alone or with others) have been difficult to study,which has led to differences in the way sociability is opera-

2

tionalized. For instance, a meta-analysis of predictors of col-lege performance found that social involvement (e.g., socialintegration, involvement in campus activities) was associatedwith higher GPAs among college students [31]. However,night outings (i.e., social events such as partying, movies,etc.) have been associated with poorer performance [18]. Inaddition, social support has been linked to higher academicperformance among college freshmen [13]. The StudentLifestudy measures sociability using conversation data inferredfrom the audio collected by smartphone’s microphone [43].It shows that greater conversation durations aggregated overa term was associated with higher spring term GPAs.

Sleep. The majority of research focused on academic per-formance and sleep patterns tends to suggest that grade aver-ages are higher among students meeting guidelines for goodsleep habits [42]. However, a large cross-sectional study ofhealth behaviors of students in forty U.S. colleges and uni-versities found that only a quarter of students (24%) actu-ally met public sleep recommendations [42]. In addition,some studies that examine the relationship between perfor-mance and sleep duration find a negative relationship betweenself-reported number of hours slept and students’ GPA [17].However, other studies have found a quadratic relationshipbetween performance (i.e., cumulative GPA) and total sleepduration, such that too little or too much sleep is associatedwith poorer performance [39]. Other research studies havefound that wake-up and bed times are important for perfor-mance, such that later bed and wake-up times are associatedwith poorer performance [39, 41]. In addition, variability insleep behaviors (e.g., bed times, wake times, total sleep dura-tion) has been linked to performance, such that greater vari-ability it associated with poorer performance [39].

Class Attendance and Studying. In general, a meta-analyticreview of college performance found that academic-relatedskills (e.g., study skills and habits) were associated withhigher GPAs among college students [31]. Research that ex-amines academic behaviors has also found that absenteeismand class attendance predict academic performance, suchthat students who attend class more often performed bet-ter than those who missed class [8, 11, 18]. For example,a meta-analysis of studies that examine the relationship be-tween class attendance and performance found attendance tobe strongly related to class grades and GPA among collegestudents [12]. However, the StudentLife study [43] foundno correlation between class attendance and academic per-formance.

Positive Affect. Relatively few studies have examined therelationship between academic performance and positive af-fect. Those studies that have focused on affect found thatpositive affect was associated with higher grades and GPAs,while negative affect during the second half of the semesterwas associated with lower grades and GPAs [32] .

Stress. The relationship between academic performance andstudents’ stress has received much attention in the researchliterature. In general, college students tend to report more

stressful daily hassles than major life events, with the topfive sources of stress being change in sleeping habits, vaca-tion/breaks, change in eating habits, increased workload, andnew responsibilities [34]. The existing research tends to showthat moderate stress is associated with decreases in students’GPAs [30], and that perceived stress during the end of thesemester is associated with lower GPAs [29]. Previous re-search tends to find a curvilinear relationship between stressand performance, such that too little or too much stress is as-sociated with poorer performance [39].

Performance Prediction. Previous research [15] aimed atpredicting performance has used a neural network model topredict student’s grades from their placement test scores. Var-ious data collected from entering students are used in [25] topredict student academic success using discriminant functionanalysis. [22] proposes a regression model to predict the stu-dent’s performance from their demographic information andtutor’s records. [33] applies web usage mining in e-learningsystems to predict students’ grades in the final exam of acourse. In [48], the authors propose an approach based onmultiple instance learning to predict student’s performancein an e-learning environment. Recent work [38] showed thatthey can predict a student is at risk of getting poor assessmentperformance using longitudinal data such as previous test per-formance and course history. To the best of our knowledgethere is no work on using passive sensor data from smart-phones as a predictor on academic success.

ACADEMIC PERFORMANCE DATASETIn this paper, we use a subset of the StudentLife dataset toanalyze and predict academic peformance. The StudentLifedataset is a large, longitudinal dataset that is publicly avail-able [43]. The dataset is collected from 30 undergrads and 18graduate students over a 10-week term in spring 2013. Eachstudent takes three classes during a term at Dartmouth Col-lege. The dataset includes over 53 GB of continuous sens-ing data from smartphones, including: 1) objective sensingdata: sleep (bedtime, duration, wake up), face-to-face con-servation duration, face-to-face conversation frequency andphysical activity (stationary, walk, run); 2) location-baseddata: location, co-location, indoor/outdoor mobility and dis-tance covered; 3) other phone data: light, Bluetooth, audio,Wi-Fi, screen lock/unlock, phone charge, and app usage. Thedataset also comprises 32,000 daily self-reports covering af-fect (PAM [27]), stress, exercise, mood, loneliness, social andstudy spaces; and pre-post surveys including PHQ9 depres-sion scale [23, 24, 37] , UCLA loneliness scale [35], posi-tive and negative affect schedule (PANAS) [46], perceivedstress scale (PSS) [10], big five personality [20], flourishingscale [14] and the Pittsburgh sleep quality index [6]. Finally,the dataset includes academic assessment data, such as, classinformation, deadlines, academic performance (i.e., grades,term GPA, cumulative GPA from transcripts), class atten-dance rates (from phone location data), Piazza data [4], andstudent dinning data including meals data, location and time.

In this paper, we only use undergraduate student (N=30) databecause only undergraduates have GPAs. In contrast, gradu-ate students [5] do not have GPA and only receive High Pass,

3

Table 1: Behaviors classified from automatic sensing andEMA data.

Sensing data

activity durationaudio inferences (voice/noise/silence)conversation frequencyconversation durationdistance coveredindoor mobilitysleep durationlocation

EMAs stresspositive affect

Table 2: Psychological surveys.

Personality

OpennessConscientiousnessExtraversionAgreeablenessNeuroticism

Mental health

PHQ-9 (pre)PHQ-9 (pre)Perceived Stress Scale (pre)Perceived Stress Scale (post)UCLA Loneliness Scale (pre)UCLA Loneliness Scale (post)Flourishing (pre)Flourishing (post)

Pass, Low Pass and No Credit for their classes. Undergrad-uates receive grades A-E, Incomplete, No Credit for classes,which is reflected in their term and cumulative GPA.

Table 1 summarized the automatic sensing data and EMAdata we use from the broader StudentLife dataset. Automaticsensing data captures daily behaviors. The EMA data cap-tures positive affect and stress level. Table 2 summarized thepre and post psychological surveys data we use.

For details on the StudentLife study and how behavioral states(e.g., sleep, face-to-face conservation) are inferred see [43].

ASSESSING STUDY AND SOCIAL BEHAVIORThe StudentLife dataset provides a number of low-levelbehaviors (e.g., physical activity, sleep duration, sociabil-ity based on face-to-face conversational data) but offers nohigher level data related to study and social behavior, whichare likely to impact academic performance. In what follows,we discuss how we attribute meanings or semantics to loca-tions – called behavioral spaces – as a basis to better under-stand study and social behavior – that is, we extract high levelbehaviors, such as, studying, study duration, study focus, andsocial behaviors, such as, partying and partying duration byfusing multiple sensor streams with behavioral spaces.

Behavioral SpacesThe StudentLife dataset has two types of location data: GPSand Wi-Fi location. Wi-Fi indicates the location of the Wi-Fiaccess points (APs) and are mapped to specific buildings orarea of buildings (e.g., libraries). We use clustering analy-sis of GPS and WiFi APs to label each campus building and

specific areas in each building with semantically meaning-ful labels, such as, study areas (e.g., libraries, specific cafeswhere students study), fraternity, sororities, classrooms, labo-ratories, department buildings, gyms, movie theaters, concerthalls, shops, food halls, TV/games rooms, student dorms, etc.These labels provide clues about a student’s behavior at thesebehavioral spaces. Importantly, we associate a number of at-tributes with behavioral spaces to give them more contextualmeanings, specifically: (i) dwell time, we are interested theamount of time a student spends at these locations – class-room, study area, dorm, laboratory, party or social spaces; (ii)activity, we compute the percentage of stationary labels fromthe activity classifier [43] among all activity inferences whena student is at specific areas – for example, in study areas orthe classrooms, a high level of stationary labels (e.g., not in-teracting with their phone) might coarsely be associated withbeing “focused” on studying or paying attention in class; andfinally (iii) audio, we calculate the percentage of silent labelsamong all audio inferences from the classifier [43] – this mayfor example be used to indicate the context of a space – highlevels of silence in a study area might indicate focused workand vice versa. We use behavioral spaces and their attributesto better assess study and social behavior, as discussed next.

Study BehaviorEach student takes three classes, which are scheduled at spe-cific periods during the week [2]. Classes fall into three cat-egories: 65-minute periods three times weekly, 50-minuteperiods four times weekly, and 110-minute periods twiceweekly. In addition, each class has an additional X-periodof 50 minutes that a lecturer may or may not use. The earliestclasses start at 8.45 AM and the latest finishes at 5.50 AM.Without knowing in advance what classes students take wecan simply use cluster analysis based on location during theacademic week and automatically determine which classesthey take using the class directory (which includes the classname, its location and schedule). We use location, date (i.e.,weekday M-F) and time to automatically determine if a stu-dent attends a class or not, checking the dwell time at the lo-cation at least equals 90% of the scheduled period (e.g., 110minutes). Using this approach the phone can automaticallydetermine the classes a student is taking and their attendancerates. We also cross check our inferred class schedule with theclass schedule provided by each student. Figure 1(b) showsthat class attendance rates across the term.

We use behavioral space information to determine study be-havior. We heuristically determine if a student’s dwell time ata study areas (e.g., library, labs, study rooms, cafes where stu-dent primarily work) is at least 20 minutes. We consider peri-ods shorter that 20 minutes are less likely to be real study pe-riods. In addition to dwell time, we use activity and audio at-tributes to determine a student’s level of focus at a study area.The value of activity indicates how often the phone moves –the person is either moving around in the study area or sta-tionary but using the phone. We consider a number of scenar-ios. If a student is in a study (e.g., a library) and moves aroundwe consider this contributes to a lack of focus. If the phone ismostly stationary in a study area, we consider this contributesto focus. We also use the audio attribute to determine the level

4

0

8

16

24

32

40

mon tue wed thu fri sat sun 0

0.2

0.4

0.6

0.8

1

stu

dy d

ura

tion (

hours

)

focus

weekday

study durationstudy focus (activity)

study focus (audio)

(a)

0

5

10

15

20

25

1 2 3 4 5 6 7 8 9 0

0.2

0.4

0.6

0.8

1

midterm

dura

tion (

hours

)

attendance

week

deadlinesattendance

study

(b)

0

2

4

6

8

2 2.5 3 3.5 4

num

ber

of stu

dents

cumulative GPA

(c)Figure 1: (a) Weekday average study duration and study focus trend. (b) Weekly study duration and lecture duration trend withcontrast to the number of deadlines. Note, the number of deadlines is scaled. (c) Cumulative GPA distribution.

of ambient noise in study areas. We consider quiet environ-ments contribute to study focus and noisy environments donot. Figure 1(a) shows the changing study duration and focusfor all the students across the term. In term of focus, a higheractivity value indicates that the student moves around less andthus is more focused and a higher audio value indicates thatthe student is in a quieter environment which is more con-ducive to being focused. We do not combine these valuesbut use them as independent variables in the analysis section.We acknowledge that both activity and audio attributes canonly represent coarse estimations of study focus. For exam-ple, noisy environments in cafes where students study maysuit certain personalities and be more conducive to studyingthan quiet libraries. We also cannot determine if a student isactually studying or on social networks on their computers.Furthermore, if students study in a group their conversationaldata would be considered as a noisy environment. We argue,however, that behavioral spaces combined with dwell times,activity and audio attributes provide an new unobtrusive, ifcoarse estimation of study duration and focus.

Social behaviorsWhile the original StudentLife dataset uses the term “socialbehavior” to mean the number of face-to-face conversationsbetween students, we extend that to include other higher levelsocial behaviors, such as: How often do students party duringthe week or across the term? How long to they party for?

Dartmouth is located in a small college town in Hanover, NewHampshire with few other partying alternatives for drink-ing other than fraternities and sororities. There are threebig drinking nights that many undergraduates attend at Dart-mouth College [1]: Wednesday, Friday, and Saturday nights.Wednesday night is when the Greek houses – particularly fra-ternities – hold their weekly “meetings” (a colloquial termfor parties). Both Friday and Saturday are big party nights oncampus. Sunday is a day of rest where students buckle downto academics to make progress on assignments before the startof the academic week. In addition, Monday nights senior so-cieties hold their “meetings”. Importantly, much of the legaland illegal drinking occurs at these parties, which are locatedin the basements of the fraternities. Frat parties are open toall students across campus and consist of playing a drinkinggame called pong with paddles and dance music. There havebeen a lot of discussions in the press about the safety issuesaround such a social scene and the new president of the col-

lege is wisely trying to create alternative venues for studentsto party [3].

We consider behavioral spaces (e.g., fraternities, sororities,dorms) and their attributes to infer if a student is partying. ifa student is at a party location we assume that they will bemoving and around acoustic sound of conversation or music.We also consider the day of the week as being significant forthe fraternity and sorority parties (i.e., Monday, Wednesday,Friday and Saturday). We discard dwell times under 30 min-utes at any partying locations.

We partition each fraternity/sorority dwell periods (i.e., visitor stay) into 10-minute windows and calculate the audio pro-file and the activity attributes. We hypothesize that the au-dio and the activity attributes should be significantly differentwhen the student is partying or not partying. We use k-meansclustering [47] to find the partying thresholds for both the au-dio (e.g., music or being surrounded by a large group of peo-ple) and activity (e.g., dancing) attributes. Figure 2(a), showsthat a student is more likely to be in a party when the audioattribute (i.e., the percentage of silent labels) is below 40%.Surprisingly, we did not find significant differences in the ac-tivity attribute. By fusing audio, dwell time, and location wecan distinguish if a student is partying even if they live in thefraternity/sorority or are just visiting. To validate our partyinference method, we compare the daily inferred party dura-tion each week with the known party days across the term.Figure 2(b) shows our inferred party data for all the studentsfor each weekday averaged over the 10-week term in terms ofthe number of hours partied. Clearly Wednesday and Fridayare the big nights on campus. In addition, Thursday and Sat-urday are also popular party nights. Sunday and Monday arenot party nights and students are clearly catching up with aca-demics. Our data from smartphones strongly aligns with theparty weekly pattern groundtruth as discussed earlier. Fig-ure 2(c) shows the partying trends across the full term. Theparty season peaks during the second week of term, steadilydrops until after the mid terms when it picks up at week 7,when there is a campus-wide spring festival called Green KeyWeekend. Finally, dormitories are another places where stu-dents socialize. Clearly, these events are not large parties.We use the same approach as discussed above to determine ifa student is socializing at a dorm.

CAPTURING BEHAVIORAL CHANGEIn what follows, we discuss behavioral change features ex-tracted from the low-level automatic sensing (e.g., sleep du-

5

0

0.2

0.4

0.6

0.8

1

0 0.2 0.4 0.6 0.8 1

partyingother

% o

f non-s

ilent la

bels

% of non-stationary labels

(a)

0

50

100

150

200

mon tue wed thu fri sat sun

part

y d

ura

tion (

hours

)

weekdays

(b)

0

50

100

150

1 2 3 4 5 6 7 8 9

midterm

part

y d

ura

tion (

hours

)

weeks

(c)Figure 2: (a) shows clustering of audio and activity feature. 32567 audio and activity pairs are clustered into two clusters. Wedefine the cluster denoted by “+” as the party cluster, which contains 9921 pairs. We can clearly see that we can apply simplethresholding on audio profile (> 0.4) to find the party cluster. (b) shows that Wednesdays, Fridays, and Saturdays are primaryparty nights, which is in line with the reality [1]. (c) shows students party less during midterms and finals.

ration) and EMA data (e.g., stress) and high-level study andsocial behaviors discussed in the previous section. We createtime series of each behavior for each student. We use datapreprocessing [43] to convert the behavioral data in variousforms to a uniform time series format. The behavior timeseries samples each behavior each day. After the data pre-processing, each time series summarizes a different behavior(e.g., physical activity, conversation frequency and duration,sleep, social behavior, and study behaviors). In order to un-derstand behavior changes across the term we propose twofeatures: behavioral slope, which captures the magnitude ofchange (e.g., increase or decrease in sleep) over the completeterm as well as the first and second half of the term for allstudents – from the start of term to the midterm point, andthen from the midterm point to the end of term; and behav-ioral breakpoints, which capture the specific points in theterm where individual behavior change occurs – the num-ber of breakpoints a student experiences indicates the rate ofchange that occurs. In addition to looking at a student’s be-havior during a day, we partition a day into three epochs asdescribe in [43]. Specifically, we label the period between12am and 9am as the night epoch, 9am to 6pm as the dayepoch, and 6pm to 12 am as the evening epoch.

Behavioral SlopeWe are interested in quantifying behavioral change of stu-dents during the term. For example, is a student more or lessactive, social, studious, etc., as the term progresses. We cap-ture the behavioral change by computing a slope for each be-havioral time series (e.g., indoor mobility, stress, affect) foreach student using linear regression. The value of the slopeindicates the direction and strength of behavioral changes. Apositive slope with a greater absolute value indicates a fasterincrease in behavioral change (e.g., partying). In contrast, anegative slope with a greater absolute value indicates a fasterdecreasing behavior level (e.g., class attendance). For exam-ple, consider the the number of independent conservationsa student has each day as a time series over the term. Aslope = 0 means the student has the same number of con-versations each day across a complete term – this is highlyunlikely. A slope < 0 means the student has fewer conversa-tions as the term proceeds. And finally, a slope > 0 meansthat the student has an increasing number of conversations asthe term proceeds. The slope of a behavior allows us to takeinto account the dynamics of behavior and understand indi-vidual differences among students. As discussed in the Stu-

dentLife study [43] the midterm period (shown in Figure 1(b)as week 4 and 5) is a significant milestone in the term. Weselect the “midterm point” as a point to measure behavioralslope up to and then after. This point is the center day of themidterm period and mid point of the complete 10 week termperiod. The workload students experience increases from thebeginning of term, as shown in in Figure 1(b). After midtermsstudents have projects and larger assignments culminating infinal exams. We partition the behavioral time series at themidterm point and use two linear regression to fit time seriesfrom the beginning of term to the midterm point (i.e., firsthalf of term) and from the midterm point to the end of term(i.e., second half of term). We use the terms pre-slope andpost-slope to capture students’ behavioral change during thefirst and second half of the term, respectively. In addition, wecompute a term-slope for each behavior taken over the com-plete term for all students.

Behavioral BreakpointsThe pre-slope and post-slope points are used collectively forall students to understand behavioral change. However, manystudents may change behaviors at different timescales thanthe midterm point. Students may enact or experience change(e.g., attend class more or less, study more or less) for manydifferent reasons. For example, some students may changetheir study behaviors early in the term to adapt to increasingworkload, whereas others may react later. We compute “be-havioral breakpoints” for each student using the time series ofeach of their behaviors. We can find a day in the term, beforeand after which the student’s behavior change patterns differ.We call this day a behavioral breakpoint. For example, con-sider a student that spends a similar number of hours study-ing. However, after a certain day the student spends moreand more time studying. We consider the point of changeas a breakpoint. Many factors influence breakpoints. Dif-ferent behaviors may have different breakpoints (e.g., an in-crease or decrease in stress, affect, and studying). The num-ber of breakpoints may indicate how quickly a student en-acts changes because of an event. We use two linear regres-sions to fit the data and use the Bayesian information criterion(BIC) [36] to select the best model. BIC is a model selec-tion criterion that selects a model with good predictive per-formance using as few model parameters as possible. LowerBIC value indicates a better model. In our analysis, we con-sider a good piecewise fitting model as the lowest BIC among

6

all piecewise models and also lower than the single regres-sion model. If the single regression model is selected, thebreakpoint is set to the last day. Using per-student behav-ioral breakpoints we can analyze the rate of changes occur-ring across the term and understand individual differences.

RESULTSIn this section, we first conduct correlation analysis to findwhich time series features have significant connection withacademic performance, specifically, the spring term and cu-mulative GPA. Figure 3(c) and Figure 1(c) shows the distri-bution of the term and cumulative GPA for the students inthe study. Cumulative GPA indicates a student’s overall longterm academic performance. The spring term GPA captureshow a student performs in a single 10-week term. In thispaper, we introduce new behaviors (e.g., study and social)and methods (i.e., behavioral slope and breakpoints) to quan-tify individual differences between students using time seriesanalysis. After correlation analysis we discuss our model forpredicting cumulative GPA.

Correlation AnalysisTo best understand the relationship between student behav-iors, emotions, mental health, and personality, and academicoutcomes we conduct Pearson correlation analysis. We iden-tify a number of strong and significant correlations.

Spring Term GPA. The mean of spring term GPA is 3.3306and the standard deviation is 0.7983. The skewness, however,is -1.7725, meaning that most students receive high GPAsfor the term and only a small portion of students get lowGPAs (Figure 3(c)). In the StudentLife paper, we found thespring term GPA negatively correlates with the means of in-door mobility and positively correlates with the conversationfrequency and duration. Here we present results from newlydesigned features as shown in Table 3.

We find a number of significant correlations between studyand social behavior and GPA. In terms of social behav-ior, we find that students who spend more time partying atfraternities or sororities are less likely to have high GPAs(r = −0.398, p = 0.029). In addition, students that social-ize more at their dorms rather than fraternities prior to themidterm point are more likely to have higher GPAs (r =0.363, p = 0.049). In terms of study behavior, we findstudents who spend more time studying have higher GPAs(r = 0.381, p = 0.038). In addition, students who show anincrease in the amount of time they devote to studying priorto the midterm point (r = 0.397, p = 0.030) are more likelyto have better grades.

In terms of other behaviors inferred from automatic sensing,we find a decrease in physical activity throughout the termparticularly after the midterm point negatively correlates withspring term GPA (r = −0.576, p = 0.001), meaning thatstudents who experience a decrease in their physical activitylevels are more likely to have higher GPAs. We find simi-lar activity trends during the day, night, and evening epochs.We find similar trends for indoor mobility were students whohave higher GPAs tend to have decreasing indoor mobility

Table 3: Spring Term GPA Correlations.

features r p-value

auto

mat

icse

nsin

g

activity term-slope -0.551 0.002activity post-slope -0.576 0.001activity night term-slope -0.431 0.017activity night post-slope -0.654 < 0.001activity day term-slope -0.411 0.024activity day post-slope -0.442 0.016activity evening term-slope -0.485 0.007conversation freq night breakpoint 0.379 0.039indoor mobility term-slope -0.606 < 0.001indoor mobility pre-slope 0.423 0.020indoor mobility post-slope -0.515 0.004indoor mobility night term-slope -0.529 0.003indoor mobility night pre-slope 0.365 0.047indoor mobility night post-slope -0.543 0.002indoor mobility day term-slope -0.568 0.001indoor mobility day post-slope -0.371 0.048indoor mobility evening term-slope -0.552 0.002dorm duration term-slope 0.437 0.016social duration dorm pre-slope 0.363 0.049party duration mean -0.398 0.029study duration mean 0.381 0.038study duration pre-slope 0.397 0.030

survey Perceived Stress Scale (post) -0.405 0.050

throughout the term (r = −0.606, p < 0.001). Interestingly,students who have higher GPAs tend to increase their indoormobility during the first half of the term (r = 0.423, p =0.020), and decrease their indoor mobility during the sec-ond half of the term (r = −0.515, p = 0.004). In addition,these high achievers tend to spend more time in their dormsthroughout the term (r = 0.437, p = 0.016). Finally, we findthe perceived stress scale negatively correlates with the springterm GPA, meaning students who are less stressed are morelikely to have higher GPAs (r = −0.405, p = 0.050).

Cumulative GPA. The mean of cumulative GPA is 3.4215and the standard deviation is 0.3978. The GPA distribution isshown in Figure 1(c). In the StudentLife paper, we found thecumulative GPA negatively correlates with the means of ac-tivity duration and indoor mobility; and positively correlateswith the number of Bluetooth co-locations. Here we presentresults from newly designed features as shown in Table 4.

We find a number of significant correlations between studyand social behavior and GPA. Students who spend moretime studying are more likely to have higher GPAs (r =0.518, p = 0.003). In addition, students who are more fo-cused in terms of their activity (i.e., their phone is more sta-tionary, r = 0.430, p = 0.018) and audio (i.e., study in qui-eter environments, r = 0.380, p = 0.038) attributes are morelikely to have higher GPAs. The study focus trends, how-ever, show that students who have higher GPAs tend to havea decreasing focus (i.e., activity attribute) before the midtermpoint and prefer to study at locations that are not quiet (e.g.,cafe area). The attendance rate does not correlate with theGPA as discussed in the StudentLife paper [43]. However, wefind that the change of attendance before midterm positivelycorrelates with the GPA, meaning that students who increasetheir attendance before midterm point are more likely to havehigher GPAs (r = 0.470, p = 0.009).

7

2

2.5

3

3.5

4

0.2 0.4 0.6 0.8

high GPA

cum

ula

tive G

PA

attendance

(a)

1

2

3

4

0.2 0.4 0.6 0.8

high GPA

spring term

GP

A

attendance

(b)

0

2

4

6

8

10

1 1.5 2 2.5 3 3.5 4

num

ber

of stu

dents

spring term GPA

(c)Figure 3: (a) Attendance and the cumulative GPA. (b) Attendance and the spring term GPA. (c) Spring term GPA distribution.

In terms of other behaviors inferred from automatic sensing,we find significant correlations between GPA and behaviorchange trends. Students whose physical activity level aremore likely to increases before midterm (r = 0.418, p =0.022) or decreases after midterm (r = −0.449, p = 0.015)are more likely to have higher GPAs. Similarly, studentswho move around indoors more toward the midterm (r =0.425, p = 0.019) or move around more after the midterm(r = −0.426, p = 0.021) are more likely to have higherGPAs. Looking at the overall behavioral changes, studentswhose indoor mobility increases more slowly or decreasesthroughout the term are more likely to have higher GPAs(r = −0.387, p = 0.035). In term of daily conversation du-ration, students who have increasing conversation durationsafter the midterm point are more likely to have higher GPAs(r = 0.443, p = 0.016). Regarding the conversation fre-quency, students with later breakpoints (i.e., changed theirdaily conversation frequency pattern later in the term) aremore likely to have higher GPAs (r = 0.641, p < 0.001 forthe night epoch and r = 0.498, p = 0.005 for the eveningepoch).

In terms of the psychological features from pre-post surveydata, we find PHQ-9 [23, 24, 37] score negatively correlateswith the GPA (r = −0.470, p = 0.027), meaning that stu-dents who are more depressed tend to have lower grades. Stu-dents who are more conscientious from the Big 5 [20] surveytend to have higher GPAs (r = 0.551, p = 0.004) and stu-dents who tend to have more neurotic are more likely to havelower GPAs (r = −0.423, p = 0.035).

Prediction AnalysisIn this section, we present a simple model that can predictGPA. We use linear regression with lasso regularization toidentify non-redundant predictors among a large number ofinput features. Theses predictors include a combination ofautomatic sensing time series behavioral data (i.e., conversa-tional and study features), EMA time series data (e.g., pos-itive affect and stress), mental health data (i.e., depression),and personality data (i.e., conscientiousness).

Predictive Model. Predicting GPA is a regression problem;that is, predicting an outcome variable (i.e., GPA) from a setof input predictors (i.e., features). We evaluate various regres-sion models such as regularized linear regression, regressiontrees, and support vector regression using cross-validation.We select the Lasso (Least Absolute Shrinkage and Selec-tion Operator) [40] regularized linear regression model as

Table 4: Cumulative GPA Correlations.

features r p-value

auto

mat

icse

nsin

g

activity pre-slope 0.418 0.022activity post-slope -0.449 0.015activity day pre-slope 0.477 0.008activity day post-slope -0.391 0.036activity night pre-slope 0.427 0.019activity night post-slope -0.411 0.027conversation duration post-slope 0.443 0.016conversation duration night post-slope 0.407 0.028conversation duration evening post-slope 0.368 0.050conversation freq night breakpoint 0.641 < 0.001conversation freq evening breakpoint 0.498 0.005indoor mobility term-slope -0.387 0.035indoor mobility pre-slope 0.425 0.019indoor mobility post-slope -0.426 0.021indoor mobility night term-slope -0.396 0.031indoor mobility night pre-slope 0.433 0.017indoor mobility night post-slope -0.448 0.015indoor mobility day post-slope -0.386 0.039class attendance pre-slope 0.470 0.009study duration mean 0.518 0.003study focus activity mean 0.430 0.018study focus activity pre-slope -0.372 0.043study focus audio mean 0.380 0.038study focus audio post-slope -0.548 0.002

surv

eys PHQ-9 depression scale (post) -0.470 0.027

conscientiousness 0.551 0.004neuroticism -0.423 0.035

our predictive model. Lasso is a method used in linear re-gression; that is, Lasso minimizes the sum of squared errors,with a bound on the sum of the absolute values of the coeffi-cients. We hypothesize that there is a linear relationship be-tween the features and the GPA outcome; that is, GPA can berepresented using a linear combination of the feature values.Therefore, we use linear regression as the predictive model.

Lasso solves the following optimization problem:

minβ0,β

(1

2N

N∑i=1

(yi − β0 − xTi β)2 + λ

p∑j=1

|βj |)

where N is the number of observations; yi is the ground truthof observation i; xi is the p degree feature vector at obser-vation i; λ is a nonnegative regularization parameter, whichcontrols the number of nonzero components of β (i.e., num-ber of the selected features); β0 is the intercept; and β is theweight vector. The regularization parameter λ is selected us-ing cross-validation. The optimization problem is essentiallyto minimize the mean square error 1

2N

∑Ni=1(yi−β0−xTi β)2

8

Table 5: Lasso Selected GPA Predictors and Weights.

features weight

survey conscientiousness 0.0449

EMApositive affect 0.0930positive affect post-slope -0.1215stress term-slope -2.6832

sensingconversation duration night breakpoint 0.3467conversation duration evening term-slope -0.6100study duration 0.0728

of fitting while keeping the model as simple as possible (i.e.,select minimal number of features to avoid overfitting). Thus,Lasso automatically selects more relevant features and dis-cards redundant features. We have D = 193 input featureandN = 30 training cases. This is a typical smallN , largeDproblem (i.e., we have more features than training instances).Lasso helps us select relevant features (i.e., predictors) toavoid overfitting.

Evaluation Metric. We use the mean absolute errors (MAE),the coefficient of determination (R2) [7], and Pierson cor-relation to measure the performance of outcome prediction.MAE measures how close predictions are to the outcomes.The mean absolute error is given by

MAE =1

n

N∑i=1

|yi − β0 − xTi β|

Smaller MAE is preferred because it indicates that the predic-tions are closer to the ground truth. R2 is a another statisticthat measures the goodness of fit of a model and indicates thathow much of the variance can our model explain. R2 rangesfrom 0 to 1, where 1 indicates that the model perfectly fits thedata. R2 can be seen to be related to the unexplained vari-ance where R2 = 0 if the feature vector X tells us nothingabout the outcome. We use Pearson correlation to measurethe linear relations between the ground truth and the predic-tive outcome.

GPA Prediction Results. We apply leave-one-subject-outcross validation [21] to determine the parameters for Lassoand the weights for each feature. In order to make the weightregularization work properly, each feature is scaled within therange [0, 1]. Selected features have non-zero weights. TheMAE of our predicted cumulative GPA is 0.179, indicatingthat the predictions are within ±0.179 of the groundtruth.TheR2 is 0.559, which indicates that the features can explain55.9% of the GPA variance. The predicted GPA strongly cor-relates with the ground truth with r = 0.81 and p < 0.001,which further indicates that our predictions can capture out-come differences using the given features.

Table 5 shows the selected features to predict the cumulativeGPAs and their weights. Interestingly, lasso selects a singlelong term measure (i.e., conscientious personality trait, timeseries self-reports for affect and stress), and three automaticsensing data behaviors (i.e., conversational and study behav-ior data). Students who are conscientious are concerned withdoing something they have to do correctly. The weights in-

dicate the strength of the predictors. Students who are moreconscientious, study more, experience positive moods (e.g.,joy, interest, alertness) across the term but register a drop inpositive affect after the midterm point, experience lower lev-els of stress as the term progresses, are less social in termsof conversations during the evening period between 6-12 pm,and experience latter change (i.e., a behavioral breakpoint) intheir conversation duration pattern. In the case of spring termGPA, Lasso does not select features for prediction. Instead, itchooses to use the intercept alone to predict the spring termGPA outcomes; that is, it chooses using 3.40 to predict thespring GPA for all the students. We believe this is due to theskewness of the spring term GPAs, as shown in Figure 3(c).Note that average spring term GPA is 3.3306. The intersectlasso selected is close to the average GPA but adjusted to theskewness.

DISCUSSIONIn this section, we contextualize our findings with regard tothe existing literature on the connections between academicperformance and students’ automatically sensed behaviors,academic-related behaviors, personality, affect, stress, andlifestyle.

We found a number of behavioral change patterns that stu-dents experience (i.e., slopes and breakpoints) significantlycorrelate with academic performance. Previous work [43]studied the level of behaviors or averages across a 10-weekterm, such as physical activity levels and sociability lev-els. However, this fails to capture the individual differencesamong students. It fails to recognize that different people mayhave different behavioral baselines. For example, extrovertssocialize more whereas introverts socialize less because so-cialization provides more joy to extroverts, thus extrovertsand introverts who have the same academic outcome mayhave different behavior level. In addition, behavior level failsto capture how behavior changes occurs overtime. By mod-eling behavioral change using behavioral slopes and break-points, we get insights into how different students react tothe events in the lives albeit social commitments, academicworkload, or other hidden triggers that impact students (e.g.,stressors or pressures). In this paper, we found that time se-ries analysis allows us to quantify individual behavioral dif-ferences over a large timescale of 10 weeks. We also foundthat slopes and breakpoints correlate with performance and insome cases are strong predictors, as discussed in the Predic-tion Analysis Section.

In terms of sociability, our results showed that changes in stu-dents’ conversation durations were significant predictors ofperformance. Specifically, students who showed change intheir conversation durations later in the term for the nightepoch had higher GPAs. Students who showed decreas-ing in their conversation durations during the evening epochthroughout the term also had higher GPAs. These findingsare novel and extend previous work that demonstrates a rela-tionship between academic performance and social involve-ment [31]. Our findings contribute to this body of work bysuggesting that changes in students sociability patterns areimportant predictors of academic performance. For example,

9

our results suggest that students who change their night timesocializing durations later in the term performed better, com-pared to those who change their night time socializing earlierin the term. Additionally, students who decrease their eveningsocializing durations during the term perform better, com-pared to students who increase their evening socializing dura-tions during the term. We suspect that these students may bepreparing for their examinations and focusing on other tasksduring the evening (e.g., studying), which could contributeto the observed decreases in ambient conversation duration.In addition, our results are consistent with previous researchthat found greater student outings at night to be associatedwith lower performance [18].

Turning to academic-related behaviors, our results showedthat study duration was a significant predictor of perfor-mance. More specifically, students with longer average studydurations had higher GPAs at the end of the term, comparedto students with shorter study durations. This finding is con-sistent with research that found academic-related skills (e.g.,study skills and habits) to be associated with higher GPAs[31]. Our results extend this work by going beyond self-reported study habits to show that unobtrusively measuredstudying habits (e.g., via WiFi and GPS) can also predictstudent performance. In contrast to previous research, wedid not find class attendance to be a significant predictor ofperformance, and we did not observe simple correlations be-tween class attendance and GPAs as other studies have sug-gested [12]. After inspecting the distribution of the studentsgrades and attendance as shown in Figure 3, we find studentswho have higher GPAs have either a high or low attendancerate, whereas students who have medium to lower GPAs havea medium attendance rate. The data shows for some high aca-demic performers, attending lectures or not does not affecttheir grades. We believe students’ attendance is determinedby the classes they take. Since all of them take at least oneprogramming classes, high achievers may not need to attendlectures to perform well.

In terms of personality, our results showed that conscientious-ness was a significant predictor of performance, such that stu-dents higher in conscientiousness had higher GPAs comparedto students lower in conscientiousness. This finding is consis-tent with psychological research that examines the relation-ship between self-reported personality traits and academicperformance of college students [28]. More specifically, wefound support for the relationship between conscientiousnessand performance as shown by the correlations between con-scientiousness and spring term GPA, as well as cumulativeGPA. However, we did not find relationships between GPAsand students’ extraversion, agreeableness, or openness, whichsome studies have found to be significant predictors of studentperformance [8,16,28]. Although neuroticism was not a sig-nificant predictor of students’ GPAs, an examination of thesimple correlations between the traits and GPAs shows thatneuroticism was negatively associated with cumulative GPAperformance (r = −0.42), suggesting that students who arehigher in neuroticism have lower GPAs.

In terms of affect, our results showed that positive affect lev-els and change were significant predictors of performance.Specifically, students with higher average levels of positiveaffect had higher GPAs at the end of the term, compared tostudents with lower average levels. Students with decreasingpositive affect after the midterm point also had higher GPAs,compared to students with increasing positive affect. Thesefindings are consistent with previous studies that demon-strate a relationship between greater positive affect and highergrades and GPAs [32]. Interestingly, previous studies havefound that increase in negative affect during the second halfof a semester is associated with lower grades and GPAs [32].However, we found that decreases in positive affect during thesecond half of the term were associated with higher GPAs.We suspect that students who focused on their academic-related tasks and performance during the second half of theterm are less likely to report feeling great positive affect (e.g.,excited, enthusiastic), compared to students who did not .

In terms of stress, our results showed that the change in stressduring the term was a significant predictor of performance,such that students with decreases in stress levels throughoutthe term had higher GPAs at the end of the term. This find-ing is consistent with previous research that finds a negativerelationship between student stress and GPAs (e.g., [30]). Wesuspect that students who performed well in their classes be-came less stressed as the semester progressed, compared tostudents who performed poorly in their classes. This possi-bility is consistent with previous research that found greaterstress during the end of the semester to be associated withlower GPAs [29].

We are the first to the best of our knowledge that has used au-tomatic sensing data from smartphones and time series EMAsto predict GPA. The predicted GPA strongly correlates withthe groundtruth with r = 0.81 and p < 0.001, MAE is 0.179,R2 is 0.559. We predict GPA without using any priors such asSAT, IQ test results, or knowing students’ grades during theterm. While a student in day-to-day life would likely mea-sure their success via assignment grades and midterm per-formance, our results show that there are a number of otherpredictors of academic success to consider. Our predictionmodel indicates that students getting better grades are moreconscientious, study more, experiences positive moods acrossthe but register a drop in positive affect after the midtermpoint, experience lower levels of stress as the term progresses,are less social in terms of conversations during the evening,and experience change in their conversation duration patternlater in the term. The correlations and prediction model dis-cussed in this paper naturally lead to a consideration of po-tential interventions to improve academic performance. Wehypothesize that our work could serve as a catalyst for newforms of real-time interventions to help under-performingstudents improve their academic performance.

We also recognize the limitations of our work. While thedataset [43] is large, rich and deep the number of studentsin the study is small (N=30). Such a small dataset is limitingbecause we cannot use more sophisticated predictive modelsor features because it may lead to overfitting. We see from the

10

predictive results that Lasso selects only 7 features from 193features despite that we found many more correlations. Next,Dartmouth is an Ivy league liberal arts college. Its undergrad-uates are among the top high school performers. Therefore,our sample is skewed to high performers with good GPAs. Fi-nally, while the students in the sample were not all computerscience majors they all took one class in common [43]: An-droid programming. The samples therefore could be biasedto science students and do not represent the larger cross sec-tion of students found in liberal arts, for example. We believethat a larger scale study with more diverse college studentsacross different universities would present better samples forour study. Such a large scale, cross institutional study wouldoffer more diverse and representative samples allowing us torefine and revalidate our predictive model accordingly.

CONCLUSIONThe SmartGPA study has shown that there a number of signif-icant correlations between the GPA and a number of behav-iors automatically inferred from smartphone sensing data. Wealso presented a number of novel automatic sensing methodsfor assessing the study and social behavior of students, in-cluding, partying instances and duration, and study durationand focus. In our previous StudentLife study [43], we usedsimple averages of all student behaviors over the term andpresented a number of correlations with performance basedon this approach. The SmartGPA study goes much deeper inour analysis of academic performance and proposes time se-ries analysis of each student’s data streams to best understandthe individuals’ differences between high and low performers.As part of that analysis we proposed novel methods to assessbehavioral changes experienced by students over the 10-weekterm – that is, we proposed behavioral slopes and breakpointsto capture changing behaviors. Furthermore, we proposed asimple predictive model that use linear regression with lassoregularization on a number of performance predictors. Thepredicted GPAs are within ±0.179 of the groundtruth. Ourresults open the way for novel interventions to improve aca-demic performance.

REFERENCES1. Dartmouth bans hard liquor in effort to end high-risk

and harmful behavior. http://college.usatoday.com/2015/01/31/dartmouth-bans-hard-liquor-in-effort-to-end-high-risk-and-harmful-behavior/.

2. Dartmouth College Weekly Schedule Diagram.http://oracle-www.dartmouth.edu/dart/groucho/timetabl.diagram.

3. Moving dartmouth forward. http://www.dartmouth.edu/~president/forward/.

4. Piazza. https://piazza.com/.

5. Regulations for Graduate Study.http://www.dartmouth.edu/~regarchive/catalog/regulations04/grad.html/.

6. BUYSSE, D. J., REYNOLDS III, C. F., MONK, T. H.,BERMAN, S. R., AND KUPFER, D. J. The pittsburghsleep quality index: a new instrument for psychiatricpractice and research. Psychiatry research 28, 2 (1989),193–213.

7. CAMERON, A. C., AND WINDMEIJER, F. A. R-squaredmeasures for count data regression models withapplications to health-care utilization. Journal ofBusiness & Economic Statistics 14, 2 (1996), 209–220.

8. CHAMORRO-PREMUZIC, T., AND FURNHAM, A.Personality predicts academic performance: Evidencefrom two longitudinal university samples. Journal ofResearch in Personality 37, 4 (2003), 319 – 338.

9. CHEN, F., WANG, R., ZHOU, X., AND CAMPBELL,A. T. My smartphone knows i am hungry. InProceedings of the 2014 Workshop on Physical Analytics(New York, NY, USA, 2014), WPA ’14, ACM, pp. 9–14.

10. COHEN, S., KAMARCK, T., AND MERMELSTEIN, R.A global measure of perceived stress. Journal of healthand social behavior (1983), 385–396.

11. CONARD, M. A. Aptitude is not enough: Howpersonality and behavior predict academic performance.Journal of Research in Personality 40, 3 (2006), 339 –346.

12. CREDÉ, M., ROCH, S. G., AND KIESZCZYNKA, U. M.Class attendance in college a meta-analytic review of therelationship of class attendance with grades and studentcharacteristics. Review of Educational Research 80, 2(2010), 272–295.

13. DEBERARD, M. S., SPIELMANS, G., AND JULKA, D.Predictors of academic achievement and retentionamong college freshmen: A longitudinal study. Collegestudent journal 38, 1 (2004), 66–80.

14. DIENER, E., WIRTZ, D., TOV, W., KIM-PRIETO, C.,CHOI, D.-W., OISHI, S., AND BISWAS-DIENER, R.New well-being measures: Short scales to assessflourishing and positive and negative feelings. SocialIndicators Research 97, 2 (2010), 143–156.

15. FAUSETT, L., AND ELWASIF, W. Predictingperformance from test scores using backpropagation andcounterpropagation. In Neural Networks, 1994. IEEEWorld Congress on Computational Intelligence., 1994IEEE International Conference on (Jun 1994), vol. 5,pp. 3398–3402 vol.5.

16. FURNHAM, A., CHAMORRO-PREMUZIC, T., ANDMCDOUGALL, F. Personality, cognitive ability, andbeliefs about intelligence as predictors of academicperformance. Learning and Individual Differences 14, 1(2002), 47 – 64.

17. GILBERT, S. P., AND WEAVER, C. C. Sleep quality andacademic performance in university students: Awake-up call for college psychologists. Journal ofCollege Student Psychotherapy 24, 4 (2010), 295–306.

11

18. GOMES, A. A., TAVARES, J., AND DE AZEVEDO, M.H. P. Sleep and academic performance inundergraduates: a multi-measure, multi-predictorapproach. Chronobiology International 28, 9 (2011),786–801.

19. GONZALEZ, E. C., HERNANDEZ, E. C., COLTRANE,A. K., AND MANCERA, J. M. The correlation betweenphysical activity and grade point average for healthscience graduate students. OTJR: occupation,participation and health 34, 3 (2004), 160–167.

20. JOHN, O. P., AND SRIVASTAVA, S. The big five traittaxonomy: History, measurement, and theoreticalperspectives. Handbook of personality: Theory andresearch 2 (1999), 102–138.

21. KOHAVI, R., ET AL. A study of cross-validation andbootstrap for accuracy estimation and model selection.In Ijcai (1995), vol. 14, pp. 1137–1145.

22. KOTSIANTIS, S., AND PINTELAS, P. Predictingstudents marks in hellenic open university. In AdvancedLearning Technologies, 2005. ICALT 2005. Fifth IEEEInternational Conference on (July 2005), pp. 664–668.

23. KROENKE, K., AND SPITZER, R. L. The phq-9: a newdepression diagnostic and severity measure. PsychiatricAnnals 32, 9 (2002), 509–515.

24. KROENKE, K., SPITZER, R. L., AND WILLIAMS, J. B.The phq-9. Journal of general internal medicine 16, 9(2001), 606–613.

25. MARTINEZ, D. Predicting student outcomes usingdiscriminant function analysis.

26. PAULHUS, D. L., AND VAZIRE, S. The self-reportmethod. Handbook of research methods in personalitypsychology (2007), 224–239.

27. POLLAK, J. P., ADAMS, P., AND GAY, G. PAM: aphotographic affect meter for frequent, in situmeasurement of affect. In Proc. of SIGCHI (2011).

28. POROPAT, A. E. A meta-analysis of the five-factormodel of personality and academic performance.Psychological bulletin 135, 2 (2009), 322.

29. RAFIDAH, K., AZIZAH, A., NORZAIDI, M. D.,CHONG, S. C., SALWANI, M. I., AND NORAINI, I.Stress and academic performance: Empirical evidencefrom university students. Academy of EducationalLeadership Journal 13, 1 (2009), 37.

30. RETTINGER, V. W. The relationship between physicalactivity, stress, and academic performance. Doctoraldissertation.

31. ROBBINS, S. B., LAUVER, K., LE, H., DAVIS, D.,LANGLEY, R., AND CARLSTROM, A. Do psychosocialand study skill factors predict college outcomes? ameta-analysis. Psychological bulletin 130, 2 (2004),261.

32. ROGATEN, J., MONETA, G., AND SPADA, M.Academic performance as a function of approaches tostudying and affect in studying. Journal of HappinessStudies 14, 6 (2013), 1751–1763.

33. ROMERO, C., ESPEJO, P. G., ZAFRA, A., ROMERO,J. R., AND VENTURA, S. Web usage mining forpredicting final marks of students that use moodlecourses. Computer Applications in EngineeringEducation 21, 1 (2013), 135–146.

34. ROSS, S. E., NIEBLING, B. C., AND HECKERT, T. M.Sources of stress among college students. Socialpsychology 61, 5 (1999), 841–846.

35. RUSSELL, D. W. UCLA loneliness scale (version 3):Reliability, validity, and factor structure. Journal ofpersonality assessment 66, 1 (1996), 20–40.

36. SCHWARZ, G., ET AL. Estimating the dimension of amodel. The annals of statistics 6, 2 (1978), 461–464.

37. SPITZER, R. L., KROENKE, K., WILLIAMS, J. B.,GROUP, P. H. Q. P. C. S., ET AL. Validation and utilityof a self-report version of prime-md: the phq primarycare study. Jama 282, 18 (1999), 1737–1744.

38. TAMHANE, A., IKBAL, S., SENGUPTA, B.,DUGGIRALA, M., AND APPLETON, J. Predictingstudent risks through longitudinal analysis. InProceedings of the 20th ACM SIGKDD InternationalConference on Knowledge Discovery and Data Mining(New York, NY, USA, 2014), KDD ’14, ACM,pp. 1544–1552.

39. TAYLOR, D. J., VATTHAUER, K. E., BRAMOWETH,A. D., RUGGERO, C., AND ROANE, B. The role ofsleep in predicting college academic performance: Is it aunique predictor? Behavioral sleep medicine 11, 3(2013), 159–172.

40. TIBSHIRANI, R. Regression shrinkage and selection viathe lasso. Journal of the Royal Statistical Society. SeriesB (Methodological) 58, 1 (1996), pp. 267–288.

41. TROCKEL, M. T., BARNES, M. D., AND EGGET, D. L.Health-related variables and academic performanceamong first-year college students: implications for sleepand other behaviors. Journal of American college health49, 3 (2000), 125–131.

42. WALD, A., MUENNIG, P. A., O’CONNELL, K. A.,AND GARBER, C. E. Associations between healthylifestyle behaviors and academic performance in usundergraduates: a secondary analysis of the americancollege health association’s national college healthassessment ii. American Journal of Health Promotion28, 5 (2014), 298–305.

43. WANG, R., CHEN, F., CHEN, Z., LI, T., HARARI, G.,TIGNOR, S., ZHOU, X., BEN-ZEEV, D., ANDCAMPBELL, A. T. Studentlife: Assessing mental health,academic performance and behavioral trends of collegestudents using smartphones. In Proceedings of the 2014ACM International Joint Conference on Pervasive andUbiquitous Computing (New York, NY, USA, 2014),UbiComp ’14, ACM, pp. 3–14.

12

44. WATANABE, J.-I., MATSUDA, S., AND YANO, K.Using wearable sensor badges to improve scholasticperformance. In Proceedings of the 2013 ACMconference on Pervasive and ubiquitous computingadjunct publication (2013), ACM, pp. 139–142.

45. WATANABE, J.-I., YANO, K., AND MATSUDA, S.Relationship between physical behaviors of students andtheir scholastic performance. In Ubiquitous Intelligenceand Computing, 2013 IEEE 10th InternationalConference on and 10th International Conference onAutonomic and Trusted Computing (UIC/ATC) (2013),IEEE, pp. 170–177.

46. WATSON, D., CLARK, L. A., AND TELLEGEN, A.Development and validation of brief measures ofpositive and negative affect: the panas scales. Journal ofpersonality and social psychology 54, 6 (1988), 1063.

47. XU, R., WUNSCH, D., ET AL. Survey of clusteringalgorithms. Neural Networks, IEEE Transactions on 16,3 (2005), 645–678.

48. ZAFRA, A., ROMERO, C., AND VENTURA, S. Multipleinstance learning for classifying students in learningmanagement systems. Expert Systems with Applications38, 12 (2011), 15020 – 15031.

13