Keep It Up: In-session Dropout Prediction to Support Blended

Classroom Scenarios

Nathalie Rzepka

, Katharina Simbeck

, Hans-Georg Müller

and Niels Pinkwart

University of Applied Sciences Berlin, Treskowallee 8, 10318 Berlin, Germany

Department of German Studies, University of Potsdam, Am neuen Palais 10, 14469 Potsdam, Germany

Department of Computer Science, Humboldt University of Berlin, Unter den Linden 6, 10099 Berlin, Germany

Keywords: Dropout Prediction, VLE, Blended Classroom.

Abstract: Dropout prediction models for Massive Open Online Courses (MOOCs) have shown high accuracy rates in

the past and make personalized interventions possible. While MOOCs have traditionally high dropout rates,

school homework and assignments are supposed to be completed by all learners. In the pandemic, online

learning platforms were used to support school teaching. In this setting, dropout predictions have to be de-

signed differently as a simple dropout from the (mandatory) class is not possible. The aim of our work is to

transfer traditional temporal dropout prediction models to in-session dropout prediction for school-supporting

learning platforms. For this purpose, we used data from more than 164,000 sessions by 52,000 users of the

online language learning platform orthografietrainer.net. We calculated time-progressive machine learning

models that predict dropout after each step (completed sentence) in the assignment using learning process

data. The multilayer perceptron is outperforming the baseline algorithms with up to 87% accuracy. By ex-

tending the binary prediction with dropout probabilities, we were able to design a personalized intervention

strategy that distinguishes between motivational and subject-specific interventions.

1 INTRODUCTION

With the onset of the COVID-19 pandemic, many

schools in Germany had to close on short notice and

teachers were forced to switch to online formats. Af-

ter more than a year of the pandemic, school closures

continued for weeks at a time, resulting in distance

formats becoming established in many schools. This

may have been the starting point for the use of digital

teaching methods, persisting even after the pandemic.

The use of digital instructional offerings can reduce

teacher workload and support internal differentiation

(Gerick et al. 2016, Kepser 2018).

In online learning platforms, dropout prediction

models are often used to provide early interventions

for at-risk users. Many studies are particularly con-

cerned with MOOCs, as dropout rates of up to 90%

are found (Kloft et al. 2014, Xing and Du 2019).

However, dropout prediction models in MOOCs have

quite different requirements to dropout prediction in

online learning platforms that support school teach-

ing:

(1) Voluntariness: most people in MOOCs partici-

pate voluntarily, while completing homework in

school is usually mandatory.

(2) Time Frame: MOOCs provide over several

weeks with a defined end. It is up to the users

how quickly they complete the course. The use

of a learning platform that accompanies school

lessons, on the other hand, depends primarily on

how the curriculum is defined as a whole.

(3) Drop Out: In the German school system you

cannot fail a single subject, but only a entire class

level. Thus, the definition of dropping out as it is

defined in MOOCs cannot be applied to school

settings.

(4) Integration: In school the subject matter is not

taught exclusively via an online course. Rather,

lessons are taught in the classroom, and accom-

panying exercises and homework take place on

the online platform.

Dropout prediction for a school assignment plat-

form is therefore in many respects different from

MOOC dropout prediction. The goal of this paper is

to translate existing research on dropout prediction to

Rzepka, N., Simbeck, K., Müller, H. and Pinkwart, N.

Keep It Up: In-session Dropout Prediction to Support Blended Classroom Scenarios.

DOI: 10.5220/0010969000003182

In Proceedings of the 14th International Conference on Computer Supported Education (CSEDU 2022) - Volume 2, pages 131-138

ISBN: 978-989-758-562-3; ISSN: 2184-5026

131

the reality of blended learning scenarios.We therefore

consider the early termination of a single session in-

stead of the dropout of the entire course. We define a

session as a limited time interval during which the

user is working on the completion of an assignment

consisting of several exercises. Leaving a session

early should be avoided, as the exercises of the as-

signments didactically build on each other. Predicting

this dropout provides the opportunity for interven-

tions in adaptive learning environments. These inter-

ventions can, for example, adjust task difficulty, dis-

play orthographic rules, or display motivational gam-

ified texts. This increases the motivation of the user

and leads to higher learning success. Therefore, the

research questions are as follows:

RQ1: How effective are machine learning models in

predicting dropout within a session?

RQ2: How can the predicted dropout probabilities be

used for in-session intervention to prevent early exit?

To this end, we first summarize the current re-

search on dropout prediction and learning analytics

intervention. Then we describe the underlying dataset

and the feature engineering of our study and the used

orthografietrainer.net platform. Next, the results of

the different models are compared using different

metrics for evaluation. Finally, the results are inter-

preted and discussed

2 RELATED WORK

2.1 Dropout Prediction in Online

Learning Environments

There are numerous studies on dropout prediction in

online learning environments. Many of these studies

examine dropouts in MOOCs (Dalipi et al. 2018,

Kloft et al. 2014, Xing and Du 2019). MOOCs do of-

fer many advantages, however, they have dropout

rates of up to 90% (Kloft et al. 2014, Xing and Du

2019). Dalipi et al. (2018) present several factors re-

sponsible for the high dropout rate. These include per-

son-related factors, such as a lack of motivation and

time, and course-related factors, such as poor course

design, too little interaction, and hidden costs. Drop-

out prediction models therefore show great potential

to define at-risk students and prevent dropout through

appropriate intervention measures (Xing and Du

2019).

Research on session dropout prediction, on the

other hand, is found less frequently. Lee et al. (2020)

investigated a session dropout prediction in an online

learning environment and proposed the Deep Atten-

tive Study Session Dropout Prediction (DAS) as a

new Transformer based encoder-decoder model. In

their work, Lee et al. (2021) combined knowledge

tracing with session dropout prediction and were able

to improve the area and receiving operator curve by

3.62% using Lee’s DAS model.

Xing and Du (2019) define three different investi-

gation paths that are often followed in prediction

models: fixed-term dropout prediction, temporal

dropout prediction, and dropout prediction model per-

formance optimization. Fixed-term dropout predic-

tion uses data from a defined period of time to per-

form the predictions. This comes with the disad-

vantage that interventions often cannot be applied

early enough. Other studies only use data from the

first week of the course to identify early dropouts.

This makes it impossible to distinguish in which of

the next weeks a person drops out. As a result, many

students are classified at-risk, and no accurate inter-

vention or personalized feedback is possible. Tem-

poral Dropout Prediction models use data from all the

previous weeks and are supposed to predict which

user will drop out in the next week. This leads to a

much smaller number of at-risk users, which makes it

possible to focus on those users. To be able to plan

personalized interventions, it would be appropriate to

calculate the dropout probability. Dropout prediction

model optimization deals with possibilities to im-

prove the model performances, for example with ap-

proaches using sentiment features in comments or us-

ing deep learning algorithms (Xing and Du 2019).

For prediction, most studies mainly use click-

stream data, which are calculated from log data in

online platforms (Sun et al. 2019, Xing and Du 2019).

They include all interactions a user has with the

course and in temporal dropout prediction models,

these must be considered individually for each week

(Hagedoorn and Spanakis 2017). Since this is a binary

classification problem, most models use supervised

learning algorithms to build the models for dropout

prediction (Liang et al. 2016). In a review by Dalipi

et al. (2018), logistic regression was the most used

machine learning model, followed by support vector

machines (SVM) and decision trees. Hidden Markov

models or survival analysis are used less frequently.

Other studies use deep neural networks or recurrent

neural networks (RNNs) which can outperform clas-

sical ML model (Sun et al. 2019, Dalipi et al. 2018,

Xing and Du 2019). Deep learning models use neural

network architecture with multiple hidden layers and

CSEDU 2022 - 14th International Conference on Computer Supported Education

132

gain great results without the need for time consum-

ing feature engineering processes before (Hernández-

Blanco et al. 2019). In their review of deep learning

in educational data mining, Hernández-Blanco et al.

(2019) found that in 67% of the reviewed articles, the

deep learning model outperformed traditional ma-

chine learning approaches.

Research on dropout prediction is therefore well

advanced and has already been examined from vari-

ous sides. However, the study of courses in MOOCs

or universities, which are designed for longer periods

of time, plays a particularly important role. In-session

dropout prediction is less studied and especially

online environments that support and supplement

classroom learning have not been explored in studies

to date. Yet the use of these environments has in-

creased dramatically, partly due to the pandemic.

With this work, we hope to make a valuable contribu-

tion to in-session dropout prediction in blended learn-

ing scenarios.

2.2 Learning Analytics Intervention

Insights and predictions in the learning analytics field

allow strong conclusions to be drawn about student

learning behavior. However, to have a real impact, the

insights gained must also be meaningfully integrated

within learning analytics interventions. The goal is to

prevent academic failure at an early stage by monitor-

ing progress data and providing personalized and ap-

propriate support (Wong and Li 2020). Interventions

are considered as the “biggest challenge in learning

analytics” (Wong and Li 2018). Wise (2014) defines

learning analytics interventions as "the surrounding

frame of activity through which analytic tools, data,

and reports are taken up and used. It is a soft technol-

ogy in that it involves the orchestration of human pro-

cesses, and this does not necessarily require the crea-

tion of a material artifact.". The design of learning an-

alytics interventions thus plays a significant role in

how helpful the use of information obtained through

learning analytics is (Na and Tasir 2017).

Learning analytics interventions comes with dif-

ferent purposes: they can improve students’ success

and course performance, as well as retention, motiva-

tion, or participation (Na and Tasir 2017). Wong and

Li (2018) categorized learning analytics interventions

into four types: direct message, actionable feedback,

categorization of students, and course redesign. In

their review, most of the interventions used the

method of direct messages, which includes telephone

calls, emails or private messages to both, students at-

risk and their tutors. These messages contain, for ex-

ample, information about additional help or counsel-

ing (Dawson et al. 2017, Milne 2017), encouraging

messages to use online resources (Blumenstein 2017,

Smith et al. 2012), or information to notify the tutor

about students at-risk (Herodotou et al. 2017). The

method of actionable feedback provides the students

with their own performance data, for example via a

dashboard that describes their learning behavior

(Wong and Li 2018). Less common than direct mes-

sage and actionable feedback are the intervention

methods “categorization of students” and “course re-

design”. Categorization of students was done in vari-

ous degrees to indicate risk levels. Interventions were

then performed per category. Course redesign oc-

curred very infrequently in the review and involved

adjusting the course based on the information ob-

tained. Information on dropout rates and at-risk stu-

dents can also be used to evaluate instructors and

course design.

In our article, we propose the prediction of stu-

dents’ early termination to gain valuable insights that

will enable these interventions to be designed and ap-

plied in a personalized way.

3 METHODOLOGY

3.1 Orthografietrainer.net

The online platform orthografietrainer.net is a learn-

ing platform to acquire German spelling skills. It con-

tains exercise sets on various aspects of the German

language, such as capitalization, comma formation,

separated and combined spelling, and sounds and let-

ters. Furthermore, exercises on German grammar are

available. The platform is mainly used in blended

classroom scenarios. The teacher registers the entire

class on the platform and then assigns tasks as home-

work from the spelling area that was discussed in

class. This is the most common form of use of the

platform. During the COVID-19 pandemic, the ac-

cess rates have risen sharply as many new teachers

started using the platform. The platform addresses

mainly school classes from fifth grade to graduating

classes. The data set consists of 181,792 assignments,

164,580 sessions and 3,224,014 answered sentences.

These were completed by 52,032 users in the period

03-01-2020 to 04-31-2020. There are different terms

to be defined before explaining the didactic structure

of the platform:

(1) Exercise Set: A set of sentences, consisting of 10

sentences. Further sentences are added automati-

cally if a sentence is not solved correctly.

Keep It Up: In-session Dropout Prediction to Support Blended Classroom Scenarios

133

(2) Sentence: One sentence containing one or sev-

eral gaps to be filled in by the user.

(3) Assignment: Assigned exercise set which is dis-

played to the user as pending; generally as home-

work.

(4) Session: Period of time a user is working on the

exercise set.

(5) Dropout: Termination of the session without

completing the exercise set; after dropout, the as-

signment is still pending.

A peculiarity of the platform is the didactic struc-

ture of the exercise sets. The platform focusses less

on learning German spelling rules and more on re-

peated practice. The typical exercise process is there-

fore as follows: the teacher teaches a sub-area of Ger-

man spelling in traditional lessons and assigns exer-

cise sets on the online platform to students as home-

work. The homework is displayed to the students as

pending assignments. Each exercise contains ten dif-

ferent sentences. Each training sentence is available

in at least three versions representing the same ortho-

graphical spelling problem in different verbal con-

texts and words. The program starts in the training

mode (1) which shows sentence 1 of the exercise set.

If sentence 1 is solved correctly, sentence 2 is dis-

played, then sentence 3 is displayed as well. The ex-

ercise set is successfully finished after 10 sentences if

the user makes no mistakes.

If the learner has made a mistake, the program

switches to version mode (2). Here the previously in-

correct sentence is displayed again until the user an-

swers it correctly. The solution is displayed and the

previous incorrect sentence is tested again immedi-

ately after. If it has now been answered correctly, the

other two versions of the sentence are displayed. If

the versions have been answered correctly, the incor-

rect sentence from the beginning is displayed again

and - if the sentence is now answered correctly - the

version mode is terminated. The program then

switches back to training mode.

At the end of the training mode the test mode fol-

lows (3), in which all sentences that were once an-

swered incorrectly are displayed again. As a result,

the assignment takes longer the more mistakes are

made by a user. Figure 1 shows the process of an as-

signment at orthografietrainer.net.

The students have the possibility to leave the ses-

sion at any time and continue working on the assign-

ment at another time. After 45 minutes of inactivity,

a user is automatically logged out. As described be-

fore, the task process is automatically extended if a

user makes mistakes: before moving on with the next

sentence, versions of the mistaken sentence are dis-

played, as well as the wrong sentence is tested again.

A session is thus considered terminated if it is left

without completing the exercise set. This can either

be the case when the required 10 sentences have not

been answered, or if there are more than 45 minutes

between two sentences. If the latter is the case it is

possible that the assignment is resumed later and fin-

ished successfully. In our data set, this is then defined

as one assignment which is worked on in two ses-

sions. The first session is considered dropped out, the

second is considered as finished successfully.

3.2 Feature Selection & Engineering

Our analysis is intended to predict whether a user will

interrupt the session before completing the exercise

set on the basis of the learning process data. Although

these are text tasks, no text analysis is carried out. The

tasks are either correct or incorrect, no further analy-

sis (e.g. of the type of error) is done. The information

as to whether the user terminates the exercise prema-

turely has not yet been stored in the database and is

therefore calculated separately. In addition to the fea-

tures of the learning process shown in table 1, features

of the assignment and session are calculated (table 2).

Figure 1: Task Process at orthografietrainer.net.

CSEDU 2022 - 14th International Conference on Computer Supported Education

134

Table 1: User Features.

Name Description

Field Field of grammar

Class level Class level of the user

Gender Gender of the user

Test position Mode in which the sentence is

displayed

User Attribute Group of users

First Reading

Describes, if the sentence is dis-

played for the first time to the

user

Distracted

Describes, if the user submitted

a task while missing a field

Success Describes, if the answer is cor-

rect

Difficulty Difficulty of the sentence

Years regis-

tered

Count of years the user is regis-

tered at the platform

Pending Count of pending tasks

School

Describes whether the sentence

was processed in school time

Multiple false Describes, if the same sentence

was answered incorrectly sev-

eral times

Datetime

Date and time when the sen-

tence was processed

Table 2: Session Features.

Name Description

Break Describes, if an assignment has

more than one session

Session No.

Session number of the assign-

ment

Order No. Order number of the sentence

Break

Describes, if this assignment was

interrupted earlier

Steps Describes the difference be-

tween the next possibility to fin-

ish an assignment and the current

sentence number

The "Break" feature describes an exercise that has

been interrupted. The exercise is then divided into

two sessions, which are defined by the "Session No."

feature. Each session contains N records of answered

sentences, which are ordered by time by the "Order

No." feature. The feature "Previous Break" shows if

and how often the assignment has been interrupted.

Table 2 shows the session features.

The structure of the exercise process results in

many exercises being completed at sentence numbers

10, 14, 18, or

22 sentences. A sum of 10 sentences

occurs when the user does not make a single mistake.

A sum of 14 sentences occurs when the user makes

one mistake: Thus, each error adds 4 extra sentences

that must be answered. While this is true for many

assignments, it is not true for all. For example, if a

user makes a second error in version mode, the loop

becomes one level deeper. To explain this specific

task structure in the model as well, the feature “Steps”

is added. This feature describes the difference be-

tween the next possibility to finish an assignment and

the current sentence number.

After preprocessing and feature engineering, the

features were one-hot encoded. This led to a number

of 24 input variables. As the successfully completed

assignments outweighed the unsuccessful ones, the

ratio of successfully / unsuccessfully completed as-

signments was balanced.

3.3 Matrices & Machine Learning

Models

The goal of this research is to predict the dropout

probability within a session. MOOC dropout predic-

tions are usually estimated after several weeks of part

icipation. In this case however, we re-estimate drop-

out prediction after every single sentence. Every time

a user submits a sentence, the prediction model is up-

dated using all sentences answered so far. Thus, the

matrix used by the prediction model grows over the

course of the session. The sentence position can take

values from 1 (beginning of exercise) to 300 (if many

mistakes were made). Since few sessions span more

than 60 sentences, up to 60 matrices are created for

every user. A matrix is defined by:

𝐴



1, … , 𝑚





1, … , 𝑛



→𝐾,



𝑖,

𝑗



⟼ 𝑎



(1)

𝑖=1,…,𝑚𝑎𝑛𝑑

𝑗

=1,…𝑛

(2)

The variables m and n define the number of lines

and columns of the matrix. Here, i describes the lines

of the matrix which represents one session. Further, j

is defined as the columns of the matrix where each

column is representing one feature. Each entry a

thus representing a feature j in session i. Each matrix

represents a sentence position and contains the an-

swered sentence of the respective sentence position

and all previous ones. Matrix 1 thus contains all the

first sentences of the sessions, and matrix 2 contains

the first two sentences.

Keep It Up: In-session Dropout Prediction to Support Blended Classroom Scenarios

135

Figure 2: Accuracy and F1-Score per Sentence and Model (DTE=Decision Tree Classifier, KNN=k-Nearest Neighbor,

logreg=Logistic Regression, MLP= multilayer perceptron).

In this study, three popular machine learning mod-

els were implemented as baseline algorithms to com-

pare to the multilayer perceptron (MLP): logistic re-

gression (logreg), decision tree classifier (DTE) and

k-nearest-neighbor (KNN). We follow the selection

of models by Xing and Du (2019) and the most fre-

quent models found in the review by Dalipi et al.

(2018). In our decision tree model (DTE), we use en-

tropy as a criterion to measure the quality of a split

and define the maximum depths of the tree as 5. The

multilayer perceptron (MLP)

consists of an input

layer with 24 input variables. The three fully con-

nected hidden layers (48, 24, and 12 nodes), as well

as the input layer, apply the rectified linear unit func-

tion (ReLu). As we are facing a binary classification

problem, the output layer applied the sigmoid func-

tion and as a loss function, we use binary cross-en-

tropy. To reduce bias, we additionally apply 5-fold

cross-validation.

The models described above are designed to pre-

dict whether a user will terminate the session prema-

turely. Most of the models are not only capable of

making a binary statement, but also of estimating the

probability of termination. Accordingly, we calculate

the probability of early exit for each session in order

to take appropriate interventions with this infor-

mation. This way, interventions can be personalized:

the higher the probability of termination, the stronger

the interventions will be.

4 RESULTS

Figure 2 shows the performance of the different ma-

chine learning algorithms. Specifically, all models

show an increasing accuracy over time. While the

curve initially rises sharply up to sentence 10, it flat-

tens out in the further course and approaches a peak

value. The MLP has the highest overall accuracy, up

to 87%. The decision tree classifier is also very good,

but is still slightly worse than the MLP at almost all

points in time and reaches a maximum value of about

85%. The other models are closer together and reach

up to 80% accuracy. The KNN model performs worst,

remaining below 80%.

We furthermore calculated the importance of the

features on the logistic regression model. Here we see

that especially success and task difficulty have an in-

fluence. Other important features are previous break,

i.e. whether the assignment has already been worked

on in a previous session, the class level, the assign-

ment type, capitalization and the user attribute. The

most important feature in the Decision Tree Classifier

is count wrong, i.e. the total number of errors. In the

KNN model, it is as well count wrong, additionally to

first reading, success and the three features indicating

the testposition.

In addition to the binary statements as to whether

a session is terminated prematurely or not, many

models also calculate a termination probability. This

does not apply to the KNN model and it is therefore

not included in the further discussion. Figure 3 shows

Figure 3: Probability distribution per model (DTE=Deci-

sion Tree Classifier, logreg=Logistic Regression, MLP=

multilayer perceptron).

CSEDU 2022 - 14th International Conference on Computer Supported Education

136

the distribution of the probabilities per model. Here

we can see that the distributions vary between the

models. The distribution of the logistic regression

model is denser and spreads mainly between 0.05 and

0.5. Few very low and few very high probabilities are

predicted. The MLP model on the other hand predicts

many low probabilities up to 0.1 and many high ones

up to 1.0, but few in the middle of the distribution.

The DTE model shows single high values and no

smooth distribution.

5 DISCUSSION

Following our two research questions, we have

shown that dropout prediction models can be applied

not only in the context of MOOCs over multiple

weeks, but also in the context of learning platforms

that accompany classroom teaching. Instead of pre-

dicting over a period of several weeks, we were able

to build models for local prediction, predicting early

termination within a session and providing the oppor-

tunity to offer in-session interventions.

Previous studies on dropout prediction in MOOCs

showed that the MPL outperformed the other

models.

This was also the case when transferring the research

to our blended learning scenario. The accuracy of the

in-session models was worse than that of the MOOC

prediction models. Xing and Du (2019), for example,

achieved an accuracy of 96% while our model could

only reach up to 87%. This can be explained by the

different data basis: in MOOCs, all clickstream data

is used over several weeks, whereas in our models

much less interaction takes place (namely only within

one session). Our analysis also calculated feature im-

portance and showed that the most important features

for in-session dropout prediction are success, task dif-

ficulty, and count of wrong tasks.

In our study, we have also gone one step further

and extended the mere prediction of early termination

with specific termination probabilities to design per-

sonalized learning analytics interventions. The drop-

out probability allows us to better differentiate be-

tween different interventions, which vary depending

on the level of dropout probability.

5.1 Outlook & Limitations

In this paper, we have shown that dropout prediction

models can be applied to classroom supporting online

learning. Using didactic and data-driven thresholds,

we are now able to propose interventions for several

probability domains that better match user needs.

Moreover, session dropout prediction can be used not

only to provide adequate interventions to learners but

also to improve learning paths in the learning envi-

ronment and to inform course creators about frequent

obstacles to exercise completion. Learning analytics

interventions can be designed on different bases (de-

scribed in related work). However, the actual inter-

ventions are then highly dependent on the specific

learning platform. In the case of the ortho-

grafietrainer.net learning platform, which is used to

acquire spelling and grammar skills, different inter-

ventions make sense based on the subject-specific

tasks and the structure of the assignments:

• To increase motivation, basic motivational mes-

sages or a display of the number of sentences still

pending can be implemented

• If users have problems with a certain sentence, they

sometimes cannot get out of the loop of versions.

In this case, it makes sense to release the user from

the loop and continue with the next sentence once

a certain probability has been reached.

• The difficulty of the sentences can be adjusted so

that the user has a higher probability of solving the

sentences correctly and thus completing the exer-

cise.

Basically, interventions can be divided into moti-

vational and subject-specific interventions. Motiva-

tional interventions only serve to further encourage

users and to keep them engaged by means of displays

or gamification elements. Subject-specific interven-

tions, on the other hand, adapt the tasks, for example,

the sentence sequencing or the difficulty. They thus

have an impact on what is learned and therefore offer

a stronger intervention method. The distinction be-

tween intervention types can be used to apply differ-

ent interventions starting at different thresholds de-

pending on dropout probabilities. Future research

should implement and evaluate these very interven-

tions in blended classroom scenarios to validate the

effectiveness of this approach.

An important limitation in our study is the fact

that the students do the homework at home and are

confronted with other confounding factors there.

Therefore, it can never be ruled out whether a drop-

out from the session is actually related to the learning

experience or has resulted from the surrounding cir-

cumstances. Furthermore, we have no information

about the way in which the platform is integrated into

the school lessons. For example, it can make an enor-

mous difference whether homework is graded or not.

This information could be included in the future by,

as an example, a survey for teachers using the plat-

form.

Keep It Up: In-session Dropout Prediction to Support Blended Classroom Scenarios

137

ACKNOWLEDGEMENTS

This research was funded by the Federal Ministry of

Education and Research of Germany in the frame-

work “Digitalisierung im Bildungsbereich” (project

number 01JD1812A).

REFERENCES

Blumenstein, M. (2017). The Student Relationship Engage-

ment System (SRES). In Building an Evidence Base

For Teaching and Learning Design Using Learning An-

alytics, 31.

Sun, D.; Mao, Y.; Du, Y.; Xu, P.; Zheng, Q.; and Sun, H.

(2019). Deep Learning for Dropout Prediction in

MOOCs. In 2019 Eighth International Conference on

Educational Innovation through Technology (EITT),

87–90.

Dalipi, F.; Imran, A. S.; and Kastrati, Z. (2018). MOOC

Dropout Prediction Using Machine Learning Tech-

niques: Review and Research Challenges. In 2018

IEEE Global Engineering Education Conference

(EDUCON), 1007-1014. IEEE 2018.

Dawson, S.; Jovanovic, J.; Gašević, D.; and Pardo, A.

(2017). From prediction to impact: Evaluation of a

learning analytics retention program. In Proceedings of

the Seventh International Learning Analytics &

Knowledge Conference. ACM, 474–478.

DOI=10.1145/3027385.3027405.

Gerick, J.; Eickelmann, B.; and Bos, W. (2016). Zum

Stellenwert neuer Technologien für die individuelle

Förderung im Deutschunterricht in der Grundschule. In

Individualisierung im Grundschulunterricht. Anspruch,

Realisierung und Risiken, 131–136.

Hagedoorn, T. R.; and Spanakis, G. (2017). Massive Open

Online Courses Temporal Profiling for Dropout Predic-

tion. In 2017 IEEE 29th International Conference on

Tools with Artificial Intelligence (ICTAI), 231–238.

Hernández-Blanco, A.; Herrera-Flores, B.; Tomás, D.; and

Navarro-Colorado, B. (2019). A Systematic Review of

Deep Learning Approaches to Educational Data Min-

ing. In Complexity 2019, 1–22.

Herodotou, C.; Heiser, S.; and Rienties, B. (2017). Imple-

menting randomised control trials in open and distance

learning: a feasibility study. In Open Learning: The

Journal of Open, Distance and e-Learning 32, 2, 147–

162.

Liang, J.; Yang, J.; Wu, Y.; Li, C.; and Zheng, L. (2016).

Big Data Application in Education: Dropout Prediction

in Edx MOOCs. In 2016 IEEE Second International

Conference on Multimedia Big Data (BigMM), 440–

443.

Milne, J. (2017). Early alerts to encourage students to use

Moodle. In Building an evidence base for teaching and

learning design using learning analytics, 24–30.

Kepser, M. (2018). Digitalisierung im Deutschunterricht

der Sekundarstufen. Ein Blick zurück und Einblicke in

die Zukunft. In Mitteilungen des Deutschen

Germanistenverbandes 65, 3, 247–268.

Lee, S.; Kim, K. S.; Shin, J.; and Park, J. (2021). Tracing

Knowledge for Tracing Dropouts: Multi-Task Training

for Study Session Dropout Prediction. In Educational

Data Mining Conference 2021.

Lee, Y.; Shin, D.; Loh, H.; Lee, J.; Chae, P.; Cho, J.; Park,

S.; Lee, J.; Baek, J.; Kim, B.; and Choi, Y. (2020).

Deep Attentive Study Session Dropout Prediction in

Mobile Learning Environment. arXiv preprint

arXiv:2002.11624.

Kloft, M.; Stiehler, F.; Zheng, Z.; and Pinkwart, N. (2014).

Predicting MOOC Dropout over Weeks Using Machine

Learning Methods. In Proceedings of the 2014 Confer-

ence on Empirical Methods in Natural Language Pro-

cessing (EMNLP), 60–65.

Na, K. S.; and Tasir, Z. (2017). A Systematic Review of

Learning Analytics Intervention Contributing to Stu-

dent Success in Online Learning. In Fifth International

Conference on Learning and Teaching in Computing

and Engineering - LaTiCE 2017. 62–68.

DOI=10.1109/LaTiCE.2017.18.

Smith, V. C.; Lange, A.; and Huston, D. R. (2012). Predic-

tive Modeling to Forecast Student Outcomes and Drive

Effective Interventions in Online Community College

Courses. In Journal of Asynchronous Learning Net-

works 16, 3, 51–61.

Wise, A. F. (2014). Designing pedagogical interventions to

support student use of learning analytics. In Proceed-

ings of the Fourth International Conference on Learn-

ing Analytics And Knowledge - LAK '14. 203–211.

DOI=10.1145/2567574.2567588.

Wong, B. T.; and Li, K. C. (2018). Learning Analytics In-

tervention: A Review of Case Studies. In 2018 Interna-

tional Symposium on Educational Technology (ISET).

178–182. DOI=10.1109/ISET.2018.00047.

Wong, B. T.; and Li, K. C. (2020). A review of learning

analytics intervention in higher education (2011–2018).

Journal of Computers in Education 7, 1, 7–28.

Xing, W.; and Du, D. (2019). Dropout Prediction in

MOOCs: Using Deep Learning for Personalized Inter-

vention. Journal of Educational Computing Research

57, 3, 547–570.

CSEDU 2022 - 14th International Conference on Computer Supported Education

138