Assessing Attendance by Peer Information Pan Deng1, Jianjun Zhou1*, Jing Lyu2, Zitong Zhao2 1Shenzhen Research Institute of Big Data, Shenzhen, China 2The Chinese University of Hong Kong, Shenzhen, China {pandeng, jinglyu, zitongzhao}@link.cuhk.edu.cn [email protected] ABSTRACT Traditionally, the attendance rate of a course or a student is isolated and might not be compared fairly, because in many Attendance rate is an important indicator of students’ study universities students are allowed to select some courses on their motivation, behavior and Psychological status; However, the own so that the course registration records of two students can be heterogeneous nature of student attendance rates due to the course different. In addition, each course can have its own attendance registration difference or the online/offline difference in a blended policy, making it harder to compare attendance rates. Courses that learning environment makes it challenging to compare attendance have mandatory attendance requirements usually have higher rates. In this paper, we propose a novel method called Relative attendance rates than those do not, so that it is not fair to compare Attendance Index (RAI) to measure attendance rates, which course attendance rates without considering attendance reflects students’ efforts on attending courses. While traditional requirements. Similarly, students who registered for courses with attendance focuses on the record of a single person or course, mandatory attendance requirements usually have higher relative attendance emphasizes peer attendance information of attendance rates than students who registered for courses with relevant individuals or courses, making the comparisons of voluntary attendance requirements, so that the attendance rate attendance more justified. Experimental results on real-life data does not always reflect the attainment of a student. Attendance of show that RAI can indeed better reflect student engagement. online and offline courses may not be compared directly as well, because the efforts to attend those courses can be significantly different. Attending online courses could be as simple as a mouse Keywords click away, while attending offline courses usually requires travelling from place to place physically. attendance, peer information, engagement, academic performance, comparison, clustering Traditionally it is not easy to fairly compare attendance in a university, due to not just the diversity of course registration and attendance requirements but also the difficulty of collecting campus-wide attendance data. Without attendance information of 1. INTRODUCTION peer students or courses, the attendance data of a student or a While studying offline is the norm for most schools, during course is isolated and difficult to adjust. However, in the era of epidemic periods all or a portion of students are forced to study Big Data, many new technologies [12, 18, 31] have been proposed online due to university closure. In such a blended learning to collect attendance data for many courses simultaneously, environment, tracking the study status and the wellbeing of making it possible to analyze the attendance structure of the students is an important issue for the university. Students’ student population, and develop new attendance calculation attendance in classes is a measure that reflects students’ methods. enthusiasm for the course and their status in the university [29]. Careful comparisons of attendance can also provide insight into Many studies suggest a correlation between attendance and students’ academic interest. If a student attends a course that has a attainment at university [5, 26, 14]. Several studies detect generally low attendance rate, it indicates that the student is more attendance rates using mobile devices and include attendance as a willing to attend the course than their classmates are; On the other feature to predict academic performance [27, 19, 28]. Attendance hand, if a student attends a course that has a generally high is also correlated with behavior and Psychological problems such attendance rate, it indicates that the student is just doing what as video game addiction [24] and depression [26]. Detecting others are doing. unusual attendance rate changes can help to identify abnormal behaviors and Psychological problems in an early stage and In this paper, we propose a novel method called Relative provide in time intervention to students in need. Attendance Index (RAI) to measure attendance, which reflects the efforts on attending courses and makes comparisons of attendance The successful applications of attendance data call for fair more justified. To our knowledge, this is the first study on fair comparisons among peer attendance, especially in universities. comparisons of attendance. While traditional attendance focuses on the record of a single person or course, we define a notion for attendance contribution to course attendance and add the attendance information of relevant individuals or courses to make the comparisons of attendance fairer. We perform a campus-wise Pan Deng, Jianjun Zhou, Jing Lyu and Zitong Zhao “Assess- ing attendance by peer information”. 2021. In: Proceedings study on attendance and analyze its effects on course grades and of The 14th International Conference on Educational Data Mining GPA. Our experiment results show that RAI has a higher (EDM21). International Educational Data Mining Society, 400-406. correlation with academic performance than the traditional https://educationaldatamining.org/edm2021/ attendance rate. EDM’21June29-July022021,Paris,France * The corresponding author. 400 Proceedings of The 14th International Conference on Educational Data Mining (EDM 2021) The rest of this paper is organized as follows. Section 2 describes 3. METHOD the related studies. Section 3 introduces the RAI definition. The traditional attendance rate of a class or a student is defined in Section 4 presents the experiment results on real-life data from a a straightforward way. Only the information of the class or the university. Section 5 discusses an application of RAI on clustering student is involved. We give the definition of Attendance Rate student populations. Section 6 describes the limitations and future (AR) formally as in Definition 1. work. Section 7 lists the acknowledgments. Definition 1 (Attendance Rate rc and rs): Given class c and 2. RELATED STUDIES student s, let 𝑛𝑐𝑟𝑒𝑔and 𝑛𝑐𝑎𝑡𝑡 be the number of students registered c and the number of students attended c respectively; let 𝑛𝑟𝑒𝑔and With the advancement of technology, many new methods have 𝑠 𝑛𝑎𝑡𝑡 be the number of classes registered by s and the number of been proposed to collect campus-wide attendance data. Several 𝑠 classes attended by s. Then the Attendance Rate (AR) of class c studies [13, 11, 18] measured attendance via QR code systems in which QR codes are generated and then scanned by students to (rc) and of student s (rs) are defined as below respectively. authenticate themselves. Wang et al. [26] deployed an APP to 𝑟 = 𝑛𝑐𝑎𝑡𝑡, (1) students’ cell phones to detect attendance by GPS signals and 𝑐 𝑛𝑟𝑒𝑔 𝑐 WiFi tracing. A method independently developed in [19] and [28] 𝑟 = 𝑛𝑠𝑎𝑡𝑡 (2) used WiFi log to calculate attendance. Studies in [2, 25] proposed 𝑠 𝑛𝑟𝑒𝑔 𝑠 Bluetooth/Beacon based attendance prediction systems. Shoewu Students can have different sets of registered classes, and classes and Idowu [22] used fingerprints and Kar et al. [12] used face can have very different attendance requirements. When comparing recognition to detect individual attendance rates. the attendance rates of two students, it is necessary to analyze the Some studies measured offline and online attendance at the same set of classes attended by these two students and the attendance time. Brennan et al. [4] detected physical attendance by thermal rates of these classes. If a student attends a class attended by sensors and online behaviors by clickstream data. The change of almost everyone, the student makes little contribution to the online and physical attendance through time was observed. attendance rate of the class; on the other hand, if a student attends However, due to a technology limit, the method did not link a class that has a low attendance rate, the student makes a physical attendance to individuals and did not study the issue of significant contribution to the attendance rate. To capture the attendance comparison. Nordmann et al. [17] mixed the data of concept, we propose the notion of attendance contribution as in physical attendance and online recording clickstream together to Definition 2. form the total attendance rate instead of studying them separately. Definition 2 (Attendance Contribution 𝐷𝑠𝑐): Let rc be the However, attendance of live lectures is still a stronger predictor attendance rate of class c, and asc be a function indicting whether than recording use on students’ academic performance. student s attended class c or not, then the Attendance Contribution of student s on the attendance rate of class c is defined as Many studies confirmed the correlation between attendance and academic performance [1, 3, 7, 16]. See [15] for a survey. [15] 𝐷 =𝑎 −𝑟, with (3) 𝑠𝑐 𝑠𝑐 𝑐 also reviewed factors that affect attendance. To work around the 1，if 𝑠 attended 𝑐 issue of fair comparisons of attendance, many studies focused on 𝑎𝑠𝑐 ={ 0，if 𝑠 did not attend 𝑐 (4) samples from the same course or samples with similar registration Since 𝑟 ∊[0,1], Attendance Contribution is a number between records (e.g., first year students) [1, 3, 10]. [3] also controlled 𝑐 factors such as age, gender, nationality etc. in their regression -1 and 1. If s has registered c and s attended c, then the analysis. Studies in [7] and [16] divided the students into bands attendance rate of c cannot be zero and 𝐷 can approach 1 but 𝑠𝑐 according to grades and used the average attendance of each band never reach 1. for correlation studies. With Attendance Contribution, we can compare the attendance Student subtyping and clustering are widely used in analyzing rates of two students by computing the average Attendance learning process and predicting academic performance. Yang et Contribution on registered classes. We defined the notion as al. [30] applied EM-IRL to students learning behavior data and Relative Attendance Index (RAI) in Definition 3. observed significant differences between groups. Romero et al. Definition 3 (Relative Attendance Index 𝑅𝐴𝐼 ): Given student s, 𝑠 [20] used clustering on online forum data to predict students’ final Let 𝐾 be the set of classes registered by s, the Relative 𝑠 performance. Resulting model turned out to be suitable and highly Attendance Index (RAI) of s is defined as interpretable. Cerezo et al. [5] studied both learning process and clusters’ relation with performance using LMS logs data. 𝑅𝐴𝐼 = ∑𝑐∈𝐾𝑠𝐷𝑠𝑐 (5) Resulting clusters are well-interpreted and showed satisfying 𝑠 |𝐾𝑠| correlation with final marks. RAI considers both the student’s individual attendance status of a semester and the attendance status of the student’s classmates. Many studies explored the reasoning for student class attendance. The peer information is injected into the new measure through the Friedman et al. [9] and Moore et al. [14] reported positive course attendance rate in Attendance Contribution. relationship between class attendance and students’ motivation. LEMMA 1: −1<𝑅𝐴𝐼 < 1. Sloan et al. [23] further found that the level of interest has 𝑠 Proof: The 𝑅𝐴𝐼 definition only considers classes registered by s. significant impact on attendance. These studies indicated that 𝑠 When 𝑎 =0, 𝑟 ∊[0,1) ; When 𝑎 =1, 𝑠 attended 𝑐, attendance, along with other features, can better show students’ 𝑠𝑐 𝑐 𝑠𝑐 therefore 𝑟 ∊(0,1] . Thus 𝑎 −𝑟 ∊(−1,1), Therefore academic interest than traditional models. 𝑐 𝑠𝑐 𝑐 𝑅𝐴𝐼 = ∑𝑐∈𝐾𝑠(𝑎𝑠𝑐−𝑟𝑐) ∊(|𝐾𝑠|×−1,|𝐾𝑠|×1) = (-1, 1). □ None of the above studies has applied peer information to revise 𝑠 |𝐾𝑠| |𝐾𝑠| |𝐾𝑠| attendance measurements. RAI is a number between -1 and 1. When RAI approaches -1, 𝑎𝑠𝑐 is mostly 0 and 𝑟 approaches 1 for most classes, indicating that 𝑐 Proceedings of The 14th International Conference on Educational Data Mining (EDM 2021) 401 the student has skipped many well attended classes. On the other correlations. Since most students attending this university did not hand, when RAI approaches 1, 𝑎 is mostly 1 and 𝑟 approaches study Calculus in high school and Calculus accounts for a large 𝑠𝑐 𝑐 0 for most classes, indicating that the student has attended many faction in Mathematics courses, it is reasonable to see the MAT poorly attended classes. Therefore, RAI shows the difference of category having a much higher correlation. attitudes toward classes between students and their classmates. Table 1. Correlation in course categories CAT. Description AR RAI 4. Results FMA Financial Mathematics 0.65 0.65 4.1 Data and Setup MSE Material Science and Engineering 0.46 0.52 The anonymous attendance and grade data used in this paper were collected in 2018 and 2019 from a university1 in China. We BIM Bioinformatics 0.35 0.51 applied the method proposed in [19] and [28] to calculate the GED General Education D 0.48 0.46 attendance. Note that our Relative Attendance Index can be applied to attendance data collected using other methods such as GEB General Education B 0.42 0.45 QR code [18]. The student IDs were converted into hash codes, STA Statistics 0.32 0.39 then the attendance and grade data were connected through the MGT Management 0.26 0.39 hash codes. The university did not have a mandatory attendance policy; however, while most instructors followed the university GE Foundation: In Dialogue with GFN 0.28 0.37 policy, some instructors had their own attendance requirements. Nature Some instructors used in-class discussions and quizzes to FIN Finance 0.34 0.36 encourage attendance. The data came from 4838 students from Cohort 14 to 19 in 44 majors, spanning over 3 semesters, with MAT Mathematics 0.34 0.36 1489 courses grouped into 37 categories by the university. For EIE Electronic Information Engineering 0.34 0.36 most courses, students received letter grades from A to F. Courses with other grades such as P/F were excluded. The traditional GFH GE Foundation: In Dialogue with 0.34 0.36 attendance rate (AR) for a student was calculated using Formula Humanity (2) in Definition 1, and the corresponding RAI was calculated ECO Economics 0.27 0.36 using Formula (5) in Definition 3. GNB Genomics and Bioinformatics 0.35 0.35 4.2 Correlation with Academic Performance GEA General Education A 0.32 0.35 Many previous studies show that attendance is correlated with ACT Accounting 0.32 0.34 academic performance. Given that the purpose of student PHY Physics 0.22 0.29 attendance comparison is usually to assess the attainment of the students, we calculated the correlation between attendance rates HSS Humanities and Social Science 0.27 0.28 and GPA to assess the fairness of attendance comparisons. A ERG General Engineering courses 0.27 0.27 more correctly calculated attendance assessment will have a higher correlation with the GPA. The Pearson correlation between GEC General Education C 0.24 0.27 RAI and GPA is 0.48, which is significantly higher than that of CHM Chemistry 0.25 0.25 AR (0.37). The p-values of the two correlation values are FRN French 0.28 0.23 3.7x10-225 and 2.6x10-129 respectively. Since they are well below the 0.05 threshold, the correlation values are generally considered CSC Computer Science 0.23 0.23 significant. MKT Marketing 0.12 0.22 We also calculated the correlation between attendance and CHI Chinese 0.13 0.19 academic performance within each course category. The result is shown in Table 1 (sorted by the RAI correlation). Some course ENG English 0.06 0.12 categories were filtered out because they had small enrollments and did not generate correlation values with low enough p-values (<0.05) to be statistically significant. For 19 out of the 26 course categories, RAI has a higher correlation than AR. Only for two 4.3 RAI Distribution categories, GED and FRN, AR has a higher correlation than RAI (The descriptions of the categories are listed in Table 1). AR and To illustrate the different distributions on RAI for high and low RAI are tie for the five categories of FMA, GNB, ERG, CHM and course grade students, we collected two sets of samples, with one CSC. We remark that language related courses such as ENG set having a course grade no less than B+ and the other set no (English) have low correlations because those courses usually greater than C. Each sample is a triplet with a hashed student ID, a have in-class discussions resulting in an unofficial mandatory course ID, and the corresponding grade received by the student in attendance requirement. Categories that rely on prior knowledge the course. We then calculated the RAI of the student in the in high school, such as Chemistry and Physics, also have low corresponding course. Figure 1 (a) shows the distribution of the first set. It shows that more than 50% of samples have RAI > 0 (better than normal). Figure 1 (b) shows the distribution of the low 1 T he use of the data by our project has been approved by the course grade set. It shows that the majority of samples have RAI < 0 (worse than normal), with some down to -0.8. For easier u niversity management and the committee in charge of personal comparisons of both sets, the values in both subfigures have been information in this university. 402 Proceedings of The 14th International Conference on Educational Data Mining (EDM 2021) Table 2. Parameters to tune for the clustering. Parameters Range Number of PCA components [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15] Eps of DBSCAN [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0] MinPoints of DBSCAN [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20] We applied the same clustering procedure to AR and RAI attendance values respectively. For AR, the procedure failed to generate meaningful clustering (the result contained one big cluster only). For RAI the procedure produced 8 clusters with 616, (a) Samples with grades ≥ B+ 472, 665, 520, 130, 1799, 61, and 549 students respectively. 26 student samples were labeled as noise by DBSCAN and excluded in the follow-up study. For each cluster, we identified the top five most popular majors among the samples in the cluster to analyze students’ academic interest and performance. Figure 2 shows the profiles of the 8 clusters from the RAI clustering, labeled as cluster 1 to 8. We combined the dimensions of student numbers, distribution of majors, RAI attendance rates, top 10% academic performance ratio, and last 10% academic performance ratio to show how RAI attendance is related with academic performance and how the analysis can provide guidance on major selection. While some of the findings are interesting, we admit that not all phenomena can be fully explained due to the complexity behind attendance and attainment [15]. For all subfigures in Figure 2, the X is the major of the students. Figure 2(a) shows the number of students in each major for the 8 clusters in a row. Some of the clusters are very specific. Cluster 5 contains two majors only, TRAN (Translation) and PSY (Psychology); Cluster 7 contains the major of FE (Financial Engineering) only. (b) Samples with grades ≤ C Figure 2(b) shows the distribution of majors among the clusters (whether a cluster accounts for a significant portion of the Figure 1. RAI of high and low course grade samples. students in a major), with each bar representing a fraction of the corresponding major in the university. For example, as shown in normalized as the proportion values. We can see that the first set Figure 2(b), close to 70% of the students majoring in PSY are in has a more concentrated distribution than the second set. This cluster 5; close to 50% of the students majoring in CSE i ndicates that students receiving grade C or lower have a much (Computer Science and Engineering) are in Cluster 6, with other higher probability of having extreme attendance behaviors large portions of CSE students in Cluster 2, 3 and 4. Figure 2(c) (skipping many courses). shows the RAI attendance of the clusters. Students in Cluster 6 have significantly lower RAI values than the other clusters. Figure 2(d) shows the ratio of students with a GPA in the top 10% of the 5. DISCUSSION major. If a bar of major m is higher than the 0.1 line, it means that the students from the cluster in major m outperform the average In this section, we showcase an application of RAI on clustering level of students in major m. Similarly, Figure 2(e) shows the the student population. portion of students with a GPA in the last 10% of the major. The We formatted the attendance values in the 37 course categories as higher the value, the worse the performance of the students, which a vector for each student, then applied a clustering algorithm on is the opposite of Figure 2(d). the vectors. Since the dimensionality of 37 is too high for most Figure 2 illustrates how RAI correlates with academic clustering algorithms, we applied PCA to reduce the performance. Figure 2(c) shows that Cluster 6 has the overall dimensionality. The clustering algorithm we applied was the lowest RAI values, with all the five majors having negative RAI DBSCAN clustering algorithm [8] using the Euclidean distance. values. Cluster 6 also has the worst top 10% ratio in Figure 2(d) DBSCAN performs density-based clustering and does not require (only one major is barely over the average cutline), and the worst the input of the cluster number. The parameters we tuned in this last 10% ratio in Figure 2(e) (all five majors worse than the experiment are specified in Table 2. We applied silhouette score average). The TRAN major has about the same number of [21] to select the best set of parameters with the highest silhouette students in Cluster 5 and Cluster 8. The TRAN in Cluster 8 has a score. higher RAI value as well as a higher top 10% ratio and a much lower last 10% ratio than TRAN in Cluster 5. There are exceptions though. CSE in Cluster 3 has a negative RAI, but its Proceedings of The 14th International Conference on Educational Data Mining (EDM 2021) 403 top 10% ratio is the highest in Cluster 3. However, this is 6. LIMITATIONS AND FUTURE WORK consistent with our result in Table 1, which shows that CSE In this study, we defined the Relative Attendance Index to adjust courses have a relatively low RAI correlation with academic the attendance measurement, with the objective of better performance (CSE major students usually take many CSE reflecting students’ attainment and interest. While attendance is courses). affected by many factors [15], the new information we introduced Another interesting finding is that in all the 7 clusters with more is only the attendance of the peer. Further improvements should than one major, the major that has the highest RAI value also has address more factors of attendance. the lowest last 10% ratio except for Cluster 2. In Cluster 2 it is the When clustering on student data, the clustering algorithm second highest RAI major EIE that has the lowest last 10% ratio. DBSCAN worked better on RAI than AR data, and the clustering The highest RAI major in Cluster 2 is BIFC (Bioinformatics), a analysis confirmed the correlation between RAI and academic new major with a relatively small enrollment. Students facing the performance. We admit that we have not been able to confirm the risk of poor academic performance may consider selecting or correlation between RAI and students’ academic interest, which is switching to the major with the highest RAI in the same cluster. an interesting topic to be further explored. While we admit that this is by-no-mean a correlation between RAI and students’ academic interest, we remark that the interest in a The raw attendance data was collected using a WiFi based method subject is generally believed to be a weapon to fight against poor [19, 28]. It is possible that some students closed the WiFi performance. Together with the fact that DBSCAN worked better connection on their cell phones or even closed their cell phones all on RAI than AR, we believe this phenomenon may suggest that together before class, leading to a false label of absence. If a RAI has a better potential than AR for exploring students’ student had less than 50 WiFi connection records in a week, their academic interest. data in that week were excluded from the statistics. While we admit that this could generate some noise in the attendance, we (a) Number of students in each major (b) Fraction of the corresponding major in the university (c) RAI (d) Top 10% ratio (e) Last 10% ratio Figure 2. Student Clustering. 404 Proceedings of The 14th International Conference on Educational Data Mining (EDM 2021) observed that this situation only occurred rarely. For example, in be interesting to see how RAI performs on other attendance data Fall term 2019 we found that only 2 out of 4838 students turned such as QR code scanning and online attendance data. off WiFi completely, and only 3.34% of the students had some weeks of data filtered out. We conjecture two reasons for this 7. ACKNOWLEDGMENTS phenomenon. First, usage of laptops and tablets is popular among This work was supported by Shenzhen Research Institute of Big the students in this university. Many students carry them to the Data. We also want to thank anonymous reviewers for helpful classroom to view course materials. Laptops and tablets usually suggestions. can connect WiFi only. Secondly, students use time confetti before and after class to chat with friends or read the news. It will [12] Kar, N., Debbarma, M. K., Saha, A., & Pal, D. R. (2012). 8. REFERENCES Study of implementing automated attendance system using [1] Alexander, V., & Hicks, R. E. (2016). Does class attendance face recognition technique. International Journal of computer predict academic performance in first year psychology and communication engineering, 1(2), 100. tutorials. International Journal of Psychological Studies, 8(1), [13] Masalha, F., & Hirzallah, N. (2014). A students attendance 28-32. system using QR code. International Journal of Advanced [2] Avireddy, S., Veerapandian, P., Ganapati, S., Venkat, M., Computer Science and Applications, 5(3), 75-79. Ranganathan, P., & Perumal, V. (2013, March). MITSAT— [14] Moore, S., Armstrong, C., & Pearson, J. (2008). Lecture An automated student attendance tracking system using absenteeism among students in higher education: A valuable Bluetooth and EyeOS. In 2013 International Mutli- route to understanding student motivation. Journal of Higher Conference on Automation, Computing, Communication, Education Policy and Management, 30(1), 15-24. Control and Compressed Sensing (iMac4s) (pp. 547-552). IEEE. [15] Moores, E., Birdi, G. K., & Higson, H. E. (2019). Determinants of university students’ attendance. Educational [3] Bijsmans, P., & Schakel, A. H. (2018). The impact of Research, 61(4), 371-387. attendance on first-year study success in problem-based learning. Higher Education, 76(5), 865-881. [16] Newman‐Ford, L., Fitzgibbon, K., Lloyd, S., & Thomas, S. (2008). A large‐scale investigation into the relationship [4] Brennan, A., Sharma, A., & Munguia, P. (2019). Diversity of Online Behaviours Associated with Physical Attendance in between attendance and attainment: a study using an Lectures. Journal of Learning Analytics, 6(1), 34-53. innovative, electronic attendance monitoring system. Studies in Higher Education, 33(6), 699-717. [5] Cerezo, R., Sánchez-Santillán, M., Paule-Ruiz, M. P., & Núñez, J. C. (2016). Students' LMS interaction patterns and [17] Nordmann, E., Calder, C., Bishop, P., Irwin, A., & Comber, their relationship with achievement: A case study in higher D. (2019). Turn up, tune in, don’t drop out: The relationship education. Computers & Education, 96, 42-54. between lecture attendance, use of lecture recordings, and achievement at different levels of study. Higher [6] Clark, G., Gill, N., Walker, M., & Whittle, R. (2011). Education, 77(6), 1065-1084. Attendance and performance: Correlations and motives in lecture-based modules. Journal of Geography in Higher [18] Patel, A., Joseph, A., Survase, S., & Nair, R. (2019, April). Education, 35(2), 199-215. Smart Student Attendance System Using QR Code. In 2nd International Conference on Advances in Science & [7] Colby, J. (2005). Attendance and Attainment-a comparative Technology (ICAST). study. Innovation in Teaching and Learning in Information and Computer Sciences, 4(2), 1-13. [19] Pytlarz, I., Pu, S., Patel, M., & Prabhu, R. (2018). What Can We Learn from College Students' Network Transactions? [8] Ester, M., Kriegel, H. P., Sander, J., & Xu, X. (1996, Constructing Useful Features for Student Success Prediction. August). A density-based algorithm for discovering clusters International Educational Data Mining Society. in large spatial databases with noise. In Kdd (Vol. 96, No. 34, pp. 226-231). [20] Romero, C., López, M. I., Luna, J. M., & Ventura, S. (2013). Predicting students' final performance from participation in [9] Friedman, P., Rodriguez, F., & McComb, J. (2001). Why on-line discussion forums. Computers & Education, 68, 458- students do and do not attend classes: Myths and realities. 472. College teaching, 49(4), 124-133. [21] Rousseeuw, P. J. (1987). Silhouettes: a graphical aid to the [10] Fryer, L. K., Ginns, P., Howarth, M., Anderson, C., & interpretation and validation of cluster analysis. Journal of Ozono, S. (2018). Individual differences and course computational and applied mathematics, 20, 53-65. attendance: why do students skip class?. Educational Psychology, 38(4), 470-486. [22] Shoewu, O., & Idowu, O. A. (2012). Development of attendance management system using biometrics. The Pacific [11] Hosen, P., Himel, N., Adil, M. A., Moon, M. N. N., & Nur, Journal of Science and Technology, 13(1), 300-307. M. F. N. (2019). Music Playing and Wallpaper Changing System Based on Emotion from Facial Expression. In [23] Sloan, D., Manns, H., Mellor, A., & Jeffries, M. (2020). Emerging Technologies in Data Mining and Information Factors influencing student non-attendance at formal Security (pp. 79-87). Springer, Singapore. teaching sessions. Studies in Higher Education, 45(11), 2203-2216. Proceedings of The 14th International Conference on Educational Data Mining (EDM 2021) 405 [24] Terry, M., & Malik, A. (2018). Video Gaming as a Factor [28] Wang, Z., Zhu, X., Huang, J., Li, X., & Ji, Y. (2018). That Affects Academic Performance in Grade Nine. Online Prediction of Academic Achievement Based on Digital Submission. Campus. International Educational Data Mining Society. [25] Varshini, A., & Indhurekha, S. (2017). Attendance system [29] Westerman, J. W., Perez‐Batres, L. A., Coffey, B. S., & using beacon technology. International Journal of Scientific Pouder, R. W. (2011). The relationship between & Engineering Research, 8(5), 38. undergraduate attendance and performance revisited: [26] Wang, R., Chen, F., Chen, Z., Li, T., Harari, G., Tignor, S., Alignment of student and instructor goals. Decision Sciences ... & Campbell, A. T. (2014, September). StudentLife: Journal of Innovative Education, 9(1), 49-67. assessing mental health, academic performance and [30] Yang, X., Zhou, G., Taub, M., Azevedo, R., & Chi, M. behavioral trends of college students using smartphones. In (2020). Student Subtyping via EM-Inverse Reinforcement Proceedings of the 2014 ACM international joint conference Learning. International Educational Data Mining Society. on pervasive and ubiquitous computing (pp. 3-14). [31] Zhou, M., Ma, M., Zhang, Y., SuiA, K., Pei, D., & [27] Wang, R., Harari, G., Hao, P., Zhou, X., & Campbell, A. T. Moscibroda, T. (2016, September). EDUM: classroom (2015, September). SmartGPA: how smartphones can assess education measurements via large-scale WiFi networks. In and predict academic performance of college students. In Proceedings of the 2016 acm international joint conference Proceedings of the 2015 ACM international joint conference on pervasive and ubiquitous computing (pp. 316-327). on pervasive and ubiquitous computing (pp. 295-306). 406 Proceedings of The 14th International Conference on Educational Data Mining (EDM 2021)