11th IEEE International Conference on Cognitive Infocommunications – CogInfoCom 2020 • September 23-25, 2020 • Online on MaxWhere 3D Web

Different learning predictors and their effects for

Moodle Machine Learning models
László Bognár Tibor Fauszt
University of Dunaújváros Budapest Business School
Department of Information Technology Faculty of Finance and Accountancy
Dunaújváros, Hungary Budapest, Hungary
[email protected] [email protected]

Abstract—In this paper 16 different Moodle Machine In online education the teaching-learning process is
Learning models for predicting the success of 57 full-time primarily implemented in Learning Management Systems
students enrolled in the Applied Statistics course at the (LMS) such as Moodle, Edmodo, Canvas, Schoology,
University of Dunaújváros in Hungary have been developed and Blackboard, Learn, etc. The first definition of cognitive
tested in terms of “goodness”. The success can be affected by infocommunication was provided by Baranyi and Csapo [26].
several factors, but here only students’ cognitive activities are Based on this definition these systems can be interpreted as an
examined. The predictors used in the models are based on: infocommunication system [22], [25], [27]. LMS collects and
number of view of PDF lecture notes, number of views of video stores information and allow the students to interact with it.
lectures, number of views of books of solved exercises, number of
Storing, retrieving and interacting with recorded data is an
views of Minitab videos (videos for problem solving with a
statistical software), number of attempts of quizzes and best
obligatory part in online services [28]. This finding is also true
grades achieved by students on quizzes. The models differed in for LMS. The systems record the various student interactions in
the number and in the types of predictors. Binary Logistic relational databases or log files thus student activities can be
Regression was used for model training and evaluation. The obtained from them. Interactions are manifested in the form of
target of the models indicates whether a student is at risk of not different data that depend on the given learning resource or
achieving the minimum grade to pass the course. The impact of student activity. Nowadays artificially cognitive functionalities
cognitive predictors that are part of the Moodle core Analytics are increasingly appearing in these systems. Student’s
API on predictive power was also examined. Having evaluated cognitive activity is inevitable in these systems, so the
the goodness of the different models, it was shown that students’ cognitive infocommunication can improve the effectiveness of
success can be predicted purely from cognitive activities, but learning [29].
their predictive powers are very diverse. The predictors of
quizzes have the largest impact on the success, however, Numerous research studies analyze the relationships
supplementing the model with other even less effective predictors between these recorded data and student success identifying
much better model can be made. Models built from purely which student activities have significant impact on student
Moodle core cognitive predictors give much less reliable results. success. Several LA tools have been developed using different
plugins, R or Phyton programs for analyzing students’ behavior
Keywords—machine learning; online learning; Moodle; in the given system [16], [17], [18], [4], [19].
student success; Learning Management Systems; Learning
Analytics The use of Machine Learning (ML) models in education is
also on the increase. With these models it is possible to make
I. INTRODUCTION predictions for students’ success individually. However, it is a
question when in the learning period we need to intervene and
Nowadays there is a growing demand for online warn a student that he/she will probably fail to complete the
educational materials, especially in higher education. These course. Reliable warning requires accurate ML models based
materials provide flexibility in learning but also give students on student’s activity.
greater responsibility. In face-to-face learning in small groups,
teachers have direct feedback on students' activities and their It is important to send the warnings for the students in time
academic progress. Teacher can get impressions about students' and the students confront with the possible consequences of
academic attainments verbally interacting with them, and their own performances. An effective way of doing this is to
he/she can continuously monitor students’ activities and integrate ML models into an LMS like Moodle. A good model
identify those students who are lagging. This is different in gives as few false predictions and alarms as possible. Although
online learning especially in mass education. Much less time is since Moodle 3.4 release it is possible to create ML models
spent on face-to-face communication between teacher and within the LMS system very few studies have been published
students, so Learning Analytics (LA) tools and educational so far about the conditions of its proper application. Using
data mining technics are needed to identify students who are at these models as black boxes hold serious risks to get unreliable
risk of not completing the course. predictions and false alarms.

978-1-7281-8213-1/20/$31.00 ©2020 IEEE 000405

Authorized licensed use limited to: Cornell University Library. Downloaded on May 22,2021 at 22:35:38 UTC from IEEE Xplore. Restrictions apply.
 L. Bognár and T. Fauszt • Different learning predictors and their effects for Moodle Machine Learning models

Several educational ML models have been set up primarily values of the indicators are the independent variables in the
with the goal of predicting student performance and then model.
examining their predictive goodness [5], [6], [7], [8], [9]. In
these models, different ML algorithms were used. Some of One specific feature of the MAA is that it automatically
them are: Logistic Regression (LR) [1], Decision Trees (DT), appends the averages of the individual indicator values as new
Bayesian Networks (BN), Artificial Neural Networks (ANN), indicators to the model (Average Predictors AP), hence it
Support Vector Machines (SVM), K-nearest neighbour (kNN). doubles the number of the indicators in the models. It must be
The metrics commonly used in these studies to express ML mentioned because this feature may have influence of the
model goodness are: Accuracy, F1 score, Matthews’s results. Too many indicators can make the model overfitted
correlation coefficient (MCC) [10]. resulting in unreliable predictions.

In our experiment 16 different ML models were created in a) Moodle core Indicators

the Moodle LMS and examined how accurately the different Based on Garrison, Anderson, and Archer theory [3], [21], in
models can predict student failure. The models can be Moodle core there are two types of pre-prepared indicators to
categorized into two groups. 5 models belonged to the group in express student's social and cognitive activities. These are
which only Moodle core predictors that are part of Moodle Social breath and Cognitive depth indicators and these
Analytics API (MAA) were used. 11 models were created with indicators are prepared for each item (activity/resource).
self-defined predictors. The models were nested, so the effect
of the type and the number of the sequentially added predictors The exact method of calculating the values of these
on the models’ goodness could have been analyzed. indicators is not described in the Moodle documentation.
Moodle is written in PHP, so the exact calculation method
II. MOOLDE MACHINE LEARNING ANALYTICS API could be decoded from the code. In general, however, Moodle
Moodle is one of the world’s most popular open source core indicators quantify student’s activities related to a
learning platform with millions of users and widely used in particular Moodle item.
online education [14], [15], [30]. Starting with version 3.4, it is b) Moodle self-defined indicators
possible to create a ML model in the Moodle Analytic API
(MAA) that is part of the core [11], [2]. The creators of Moodle Moodle is an open source system that allows us to integrate
provided a great tool for course builders with this new feature. our own indicators into the system specific to the target we
Using the tool, scheduled analysis can be executed based on want to predict. Indicators are classes written in PHP and the
ML models. During the course, those students can be identified values of the indicators can be calculated in methods in such
automatically who are lagging and at a risk of dropping out. classes.
2) Moodle targets and backends
A. Moodle Activities and Resources MAA is responsible for generating CSV files containing
In Moodle there are two types of items through which the values of the indicators and target. This CSV file can be
student activities are realized. These items are the activities and downloaded from the system and the data can be statistically
the resources. An activity is a general term for a group of analyzed by other professional statistical software packages.
features in a Moodle course. Usually an activity is something Moodle ML backend is responsible for processing the CSV
that a student will do that interacts with other students and or files and providing results about the model goodness.
the teacher [12]. An activity usually refers to a resource. A The metrics commonly used to express ML model
resource is an item that a teacher can use to support learning, goodness are: Accuracy, F1 score, Matthews Correlation
such as a file or link [13]. In the standard Moodle there are 14 Coefficient (MCC), normalized Matthews Correlation
different types of activities (Assignments, Chat, Choice, Coefficient (nMCC) [24]. The built-in backend gives the
Database, Feedback, Forum, Glossary, Lesson, LTI External Accuracy of the model and the variance of nMCC. These
tool, Quiz, SCORM, Survey, Wiki, Workshop) and 7 different values are intended to indicate the predictive power of the
types of resources (Book, File, Folder, IMS Content package, model. The system does not provide F1 score which sometimes
Label, Page, URL). There are several other plugins which would be better to use.
represent other user activities like Games, Feedback, Wiki and
so many other. From these the following are relevant in our The backends only support binary classifications so the
research: Quiz from the group of activities, the Book, Page value of the target must be either 0 or 1. The backend can be
(can contain File or Video) from the group of resources. written in any programming language. The core backend is
written in PHP and uses Logistic Regression algorithm.
B. Moodle indicators, targets, backends, time splitting 3) Time splitting methods
methods There is an option for splitting up a model into time
1) Moodle indicators segments: in weeks, quarters, 8 parts, 10 parts, etc. This is
In MAA the predictors for a given student are determined important when we want to make predictions during the course.
by their course activities and expressed by indicators. The Based on a time-split model created from a previous realization
indicators can practically express the weight of a student of the course, the system can automatically make predictions
activity and always have the value in the range [-1,1]. The for each time-split of the current course. It is useful to warn the
student that he/she will probably not be able to complete the

000406

Authorized licensed use limited to: Cornell University Library. Downloaded on May 22,2021 at 22:35:38 UTC from IEEE Xplore. Restrictions apply.
 11th IEEE International Conference on Cognitive Infocommunications – CogInfoCom 2020 • September 23-25, 2020 • Online on MaxWhere 3D Web

course in time. In this type of model additional predictors are The self-defined indicators:
given to the model to indicate each time segment. For a model
Quiz Self-check Total Attempt (QSTATT) indicator
with 10 time-splits, 10 additional time-split indicators appear
indicates the Total of Attempts (TATT) the student has
automatically.
completed all self-tests within a chapter.
III. THE COURSE AND THE ML MODELS

A. The course
The course analyzed in this study was attended by 57 full- ()
time students at the University of Dunaújváros in Hungary and
delivered through Moodle LMS. The subject of the course was where n the number of quizzes in a chapter, A i represents
Applied Statistics. The course was a blended learning course. how many times the student attempted the i-th quiz. If TATT
The students learned online from the course materials in >= 10 then QSTATT = 1, otherwise QSTATT= TATT /10.
Moodle, but they got additional teacher support. The course Quiz Self-Test Max Grade (QSTMGR) indicator is the ratio
has 7 chapters using the resources: Lecture videos, Minitab of the Total of Best Grades (TBG) achieved by the student on
videos (videos for problem solving with a statistical software), self-tests within a chapter and the Total of Achievable
PDF lecture notes, Books of solved exercises, Quizzes for Self- Maximum Grades (TAMG) of self-tests within a chapter.
testing. Throughout the course the students wrote four
midterm-tests, 25 points of each, and the sum of the points they
earned had to reach the minimum points of 70 to get a grade to
pass the course.
(2)
B. Moodle ML models of the study where n the number of quizzes in a chapter, m the number
The 16 different ML models used in the study could be of attempts of a given quiz, GRi the grade achieved by the
categorized into two groups. The 5 models in the first group student of the i-th attempt.
had some of the 3 Moodle core indicators, while the 11 models
in the second group had some of the 57 self-defined indicators.
The target in each model was an educational target, namely for
a given student the value of the target is 0 if the course is
passed and 1 if the course is failed by the student. In all model
time splitting method was used with 10 time-splits. ()
The goodness of the models was evaluated by the metrics: where n the number of quizzes in a chapter, AG i the
Accuracy, F1 score (F1), normalized Matthiew correlation achievable maximum grade of the i-th quiz. If TBG = 0 then
coefficient (nMCC). QSTMGR = 0, otherwise QSTMGR = 1 – ((TAMG - TBG) /
The Average Predictors were removed from the model. TAMG).
Binary Logistic Regression (LR) algorithm was used to train QMTATT indicator indicates the Total of Attempts (TATT)
and evaluate the models in MATLAB (MATLAB 2008, release the student has completed the chapter module test. If TATT >=
2018b). 5-fold cross-validation technics were applied. 10 then QMTATT = 1 otherwise QMTATT = TATT / 10.
1) Group of Moodle core indicator models QMTMGR indicator is the ratio of the Best Grade
In Moodle Analytic API there are site-level or course-level Maximum (BGM) achieved by the student on module test in a
analysis. The Moodle core indicators can express a student’s chapter and the Achievable Maximum Grade of the Module
overall site or overall course activity. This means, for example, test in the chapter (AMGM).
that a site level Quiz Cognitive indicator expresses a student’s
total activity for all quizzes that the student has solved in any
of the courses in which he/she enrolled while a course level ()
indicator expresses a student’s overall activity for all tests that
he/she has solved within a given course. These indicators where GRi the grade achieved by the student of the i-th
basically express how intensive the student’s interactions with attempt. If BGM = 0 then QMTMGR = 0, otherwise
course activities. QMTMGR = 1 – ((AMGM - BGM) / AMGM).
Only Moodle core Cognitive depth indicators were used PDF indicator indicates the Number of Downloads
here: Quiz Cognitive (QC), Page Cognitive (PC) and Book (NDPDF) the student has downloaded PDF documents in a
Cognitive (BC). chapter. If NDPDF >= 15 then PDF = 1, otherwise PDF =
QMTMGR = NDPDF / 15.
2) Group of Self-defined indicator models
Self-defined linear type indicators were developed for VIDEO indicator related to Lecture videos indicates the
indicating overall course cognitive activity. The values of the Number of Views of Lecture Videos in a chapter (NVLV). If
indicators were floating point numbers between [0, 1].

000407

Authorized licensed use limited to: Cornell University Library. Downloaded on May 22,2021 at 22:35:38 UTC from IEEE Xplore. Restrictions apply.
 L. Bognár and T. Fauszt • Different learning predictors and their effects for Moodle Machine Learning models

NVLV >= 15 then VIDEO = 1, otherwise VIDEO = NVLV / An interesting comparison can be made between Model 3
15. and Model 15. Both models include quiz-related indicators
only. The two models differ in the number of indicators (NoP:
MINITAB indicator indicates the Number of Views of 3, NoP: 28) and in the way how the values of the indicators are
Videos with Minitab software in a chapter (NVV). If NVV >= calculated (See calculation of QMTATT and QMTMGR). The
15 then MINITAB = 1, otherwise MINITAB = NVV / 15. Model 3 which contains only 1 Moodle core indicator gives not
BOOK indicator indicates the Number of Views of a Book interpretable result, while Model 15 is one of the best models
Chapters in a chapter (NVBC). If NVBC >= 15 then BOOK = among all. (Accuracy: 0.89, F1 score: 0.6, nMCC: 0.77). We
1, otherwise BOOK = NVBC / 15. can state that much better model can be built with properly
defined indicators based on same student activity.
IV. RESULTS The goodness of models built from less effective indicators
are getting surprisingly better as the number of indicators are
A. Moodle core indicator models’ results increasing. Model 14 includes all non-quiz-type indicators,
The results are summarized in Table I. which are defined separately in Model 6, Model 7, Model 8,
and Model 9. Yet the goodness of the Model 14 is much better
The models (Model1, Model 2, Model 3 and Model 4) than these separated models. While Model 6 and Model 7 F1
containing only one or two Moodle core indicators give Not score is not interpretable, Model 8 F1 score is 0.10, Model 9 F1
Interpretable (NI) results of F1 score and nMCC. Actually 9 score is 0.12, then Model 14 F1 score is 0.48 which
students of the 57 failed but these models falsely classified all significantly exceeds these individual F1 score values or even
students being successful. the sum of them.
The Accuracy of the full model containing all 3 cognitive Model 16 with all the 57 self-defined indicators is the best.
indicators (Model 5) is 0.84 (the same as for all other models) (Accuracy 0.91, F1 score: 0.71, nMCC: 0.83).
and nMCC 0.54. This relatively large Accuracy value is
misleading. It suggests a good model but this nMCC value and However, it is important to note that this finding namely
the F1 score are very low indicating that even this full model is that among the models investigated the model with all the
inappropriate for reliable predictions. indicators proved to be the best is not always the case. Creating
too many indicators may result in overfitting. This question is
investigated in detail in [20].
TABLE I. GROUP OF MOODLE CORE INDICATOR MODELS
Model 1 Model 2
PC BC TABLE II. GROUP OF SELF-DEFINED INDICATOR MODELS
NoP: 1 NoP: 1
Acc: 0.84 Acc: 0.84 Model 6 Model 7 Model 8 Model 9
F1: NI F1: NI PDF BOOK VIDEO MINITAB
nMCC: NI nMCC: NI NoP: 7 NoP: 7 NoP 7: NoP: 8
Model 4 Model 3 Acc: 0.84 Acc: 0.84 Acc: 0.84 Acc: 0.84
PC + BC QC F1: NI F1: NI F1: 0.10 F1: 0.12
NoP: 2 NoP: 1 nMCC: 0.49 nMCC: NI nMCC: 0.55 nMCC: 0.56
Acc: 0.84 Acc: 0.84 Model 12 Model 13 Model 10 Model 11
F1: NI F1: NI PDF+BOOK VIDEO + MINITAB QMTATT QMTMGR
nMCC: NI nMCC: NI NoP: 14 NoP: 15 NoP: 14 NoP: 14
Model 5 Acc: 0.83 Acc: 0.84 Acc: 0.85 Acc: 0.85
QC + PC + BC F1: 0.02 F1: 0.27 F1: 0.45 F1: 0.41
NoP: 3 nMCC: 0.49 nMCC: 0.61 nMCC: 0.69 nMCC: 0.67
Acc: 0.84 Model 14 Model 15
F1: 0.08 PDF+BOOK+VIDEO + MINITAB QATT+ QMTMGR
nMCC: 0.54 NoP: 29 NoP: 28
Acc: 0.86 Acc: 0.89
F1: 0.48 F1: 0.6
nMCC: 0.70 nMCC: 0.77
Model 16
B. Self-defined indicator models’ results FULL - PDF+BOOK+VIDEO + MINITAB + QMTATT+ QMTMGR
NoP: 57
The results are summarized in Table II. Acc: 0.91
F1: 0.71
Models containing only PDF and only BOOK indicators nMCC: 0.83

(Model 6, Model 7) gave not interpretable results. It should be

V. CONCLUSIONS
noted, however, that these resources could be downloaded and
printed by the students, thus the number of clicks may not In a comprehensive experiment several educational
necessarily reflect properly the use of these resources. machine learning models have been analyzed and compared in
terms of their predictive power. The indicators composed from
Model 10 and Model 11 with quiz-related indicators students’ quiz activities proved to be the most effective
(QMTATT, QMTMGR) have the highest F1 scores (0.45, 0.41). individual predictors but including additional, even less
These are qualitatively better indicators than all the others. effective, predictors related to other activities can significantly
Interestingly, the indicators for viewing video materials improve the efficiency of the model. To assess the goodness of
(VIDEO, MINITAB) and used in Model 8 and Model 9 did not ML models their accuracy can be misleading. F1 score or
have large impact on the goodness of the model (F1 scores: nMCC are better metrics to reflect the goodness of the models.
0.10, 0.12).

000408

Authorized licensed use limited to: Cornell University Library. Downloaded on May 22,2021 at 22:35:38 UTC from IEEE Xplore. Restrictions apply.
 11th IEEE International Conference on Cognitive Infocommunications – CogInfoCom 2020 • September 23-25, 2020 • Online on MaxWhere 3D Web

VI. FUTURE WORK [12] Moolde docs: https://docs.moodle.org/39/en/Activities

[13] Moodle docs: https://docs.moodle.org/39/en/Resources
The PHP source of indicators developed for this research
[14] Nagy B., Váraljai M. Mihalovicsné K. A., “E-learning Spaces to
are not currently merged into Moodle core. The Moodle Empower Students Collaborative Work Serving Individual Goals”,
development community uses the GIT distributed version ACTA POLYTECHNICA HUNGARICA 17 : 2 pp. 97-114. , 18 p. 2020
control system [23]. Development in an open source system [15] Ujbányi T., Stankov G. Nagy B., “A transparent working environment in
has its own rules. New source code that planned to be built into MaxWhere virtual space”, In: Proceedings of the 10th IEEE
the system must goes through a strict source code review and International Conference on Cognitive Infocommunications:
only after that can the code become the part of the core. The CogInfoCom 2019 Piscataway (NJ), USA : IEEE, pp. 475-478. , 3 p. ,
2019
contact of the Moodle development community has been made
[16] S. Graf, C. Ives, N. Rashman, A. Ferri, “AAT – A Tool for Accessing
and after the testing and verification phase, the PHP code of the and Analysing Students’ Behaviour Data in Learning Systems” LAK
indicators may appear as part of the Moodle LMS. '11: Proceedings of the 1st International Conference on Learning
Analytics and Knowledge, pp: 174–179, 2011
In Moodle there is an option to switches a model in
[17] N.Kadoić, D. Oreški, “Analysis of Student Behavior and Success Based
production mode. In this mode depending on the prediction on Logs in Moodle”, In: 41st International Convention on Information
models target notifications can be received about that student and Communication Technology, Electronics and Microelectronics
who are at risk of dropping out. Additional help can be given to (MIPRO), 2018
students identified by the system who are at risk of failure. [18] V. Rachel, G. Sudhamathy, M. Parthasarathy, “Analytics on Moodle
Teacher assistance, extra attention, mentoring support, Data Using R Package for Enhanced Learning Management”,
additional tasks can all help student success reducing dropout. International Journal of Applied Engineering Research ISSN 0973-4562
Volume 13, Number 22 pp. 15580-15610, (2018)
In the next semesters the model will be tested on ongoing
[19] E. Młynarska, D. Greene, P. Cunningham, “Indicators of Good Student
courses to identify students lagging. Performance in Moodle Activity Data” In arXiv preprint arXiv:1601.
02975, 2016.
REFERENCES [20] L. Bognár, T. Fauszt, “How to Avoid Unreliable Predictions in
Educational Machine Learning Analytics” unpublished.
[21] E. Dalton “Learning Analytics and the Community of Inquiry Model”,
[1] D R Cox. 1958. the Regression Analysis of Binary Sequences. Journal
MoodleMoot US 2016
of the Royal Statistical Society. Series B (Methodological) 20, 2 (1958),
215–242. https://doi.org/10.2307/2983890 [22] P. Baranyi, Á. Csapó “Definition and Synergies of Cognitive
Infocommunications”, Acta Polytechnica Hungarica, Vol. 9 No. 1, pp:
[2] D. Monllaó, D.Q. Huynh, M. Reynolds, M. Dougiamas and D. Wiese,
67-83, 2012.
“A Supervised Learning framework for Learning Management Systems”
In: Proceedings of the First International Conference on Data Science, [23] Git: https://git-scm.com/
E-Learning and Information Systems, p. 18. ACM, October 2018. [24] Chicco, D., Jurman, G. “The advantages of the Matthews correlation
[3] D.R. Garrison, T. Anderson, W. Archer, “Critical Inquiry in a Text- coefficient (MCC) over F1 score and accuracy in binary classification
Based Environment: Computer Conferencing in Higher Education”, The evaluation.”, BMC Genomics 21, 6. https://doi.org/10.1186/s12864-019-
Internet and Higher Education. pp: 87-105, ISSN: 1096-7516. 1999. 6413-7 2020.
[4] B. Minaei-Bidgoli, D. A. Kashy, G. Kortemeyer, W. F. Punch, [25] P. Baranyi and A. Csapo, “Cognitive infocommunications”,
“Predicting student performance: An applicationof data mining methods Coginfocom. In Computational Intelligence and Informatics (CINTI),
with an educational we-based system.”, Session F1E 0-7803-7444- 2010 11th International Symposium on, pages 141 –146, 2010.
4/02/$17.00 2002 IEEE November 6 - 9, Boston, MA. 2002. [26] P. Baranyi, A. Csapo, G. Sallai, “Cognitive Infocommunications
[5] G. Đambić, M. Krajcar and D. Bele, “Machine learning model for early (CogInfoCom)” Springer International Publishing, 208 pages, ISBN:
detection of higher education students that need additional attention in 978-3-319-19607-7, 2015.
introductory programming courses”, In: International Journal of [27] Klempous R., Nikodem J., P. Baranyi, “Cognitive Infocommunications,
DIGITAL TECHNOLOGY & ECONOMY. Vol. 1, No. 1, pp: 1-11, Theory and Applications”, Springer International Publishing, ISBN 978-
2016. 3-319-95996-2, 2019
[6] S. Kotsiantis, C. Pierrakeas, P. Pintelas, “Predicting students [28] Bujdosó Gy., Boros K., Novac C. M., Novac C. O., “Developing
performance in distance learining using machine learning techniques”, cognitive processes as a major goal in designing e-health information
Applied Artificial Intelligence, Vol. 18, No 5, 411-pp: 426. 2004 provider VR environment in information science education”, 10th IEEE
[7] D. Kabakchieva, “Predicting Student Performance by Using Data International Conference on Cognitive Infocommunications –
Mining Methods for Classification”, CYBERNETICS AND CogInfoCom 2019 October 23-25, Naples, Italy, 2019.
INFORMATION TECHNOLOGIES, Vol 13, No 1, pp: 61- 72, 2013 [29] Kővári A., “CogInfoCom Supported Education : A review of
[8] S. Abu, I. Zaqout, M. Abu,, R. Atallah and E. Alajrami, “Predicting CogInfoCom based conference papers”, 9th IEEE International
Student Performance Using Artificial Neural Network: in the Faculty of Conference on Cognitive Infocommunications – CogInfoCom, August
Engineering and Information Technology”, In: International Journal of 22-24, Budapest, Hungary, 2018.
Hybrid Information Technology Vol.8, No.2, pp.221-228, 2015 [30] Lampert B., Pongracz A., Sipos J., Vehrer. A., Horvath I. “MaxWhere
[9] A. Daud, N. Radi, R. Abbasi, M. D Lytras, F. Abbas, J. S. Alowibdi, VR-Learning Improves Effectiveness over Clasiccal Tools of e-
“Predicting Student Performance using Advanced Learning Analytics”, learning”, Acta Polytechnica Hungarica, Vol. 15, No. 3, pp: 125-147,
Proceedings of the 26th International Conference on World Wide Web 2018,
Companion April pp: 415–421. 2017 [31] Kolnhofer –Derecskei A., Reicher R. Zs., Szeghegyi Á., “The X and Y
[10] Chicco, D., Jurman, G. “The advantages of the Matthews correlation Generations’ Characteristics Comparison”, Acta Polytechnica
coefficient (MCC) over F1 score and accuracy in binary classification Hungarica, Vol. 14, No. 8, pp: 107-125, 2017.
evaluation.”, BMC Genomics 21, 6. https://doi.org/10.1186/s12864-019-
6413-7 2020.
[11] Moodledocs: https://docs.moodle.org/dev/Analytics_API.

000409

Authorized licensed use limited to: Cornell University Library. Downloaded on May 22,2021 at 22:35:38 UTC from IEEE Xplore. Restrictions apply.
L. Bognár and T. Fauszt • Different learning predictors and their effects for Moodle Machine Learning models

000410

Authorized licensed use limited to: Cornell University Library. Downloaded on May 22,2021 at 22:35:38 UTC from IEEE Xplore. Restrictions apply.