Using Process Mining For Learning Resource Recommendation A Moodle Case Study
Using Process Mining For Learning Resource Recommendation A Moodle Case Study
com
Available online at www.sciencedirect.com
Available online at www.sciencedirect.com
ScienceDirect
Procedia Computer Science 00 (2021) 000–000
Procedia
Procedia Computer
Computer Science
Science 19200 (2021)
(2021) 000–000
853–862 www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia
Abstract
Abstract
Nowadays, Learning Management Systems (LMS) play an intrinsic role in education. They gather traces about the learner (course
Nowadays, Learning
view, wiki view, quiz Management
attempt, etc.)Systems
in event (LMS) play an
logs. These logsintrinsic
offer therole in education.
opportunity They gather
to provide traces about
dashboards the learner
and analysis (course
on learners.
view, wiki view, quiz attempt, etc.) in event logs. These logs offer the opportunity to provide dashboards and
There are several techniques that analyze event logs for different purposes (adaptation, recommendation, performance detection, analysis on learners.
There are several
etc.). Within techniques our
this framework, that central
analyzefocus
eventislogs
uponfor different purposes
Educational (adaptation,
Process Mining recommendation,
technique performance
which generates detection,
process models for
etc.). Within
improving this framework,
learning our central focus is upon Educational Process Mining technique which generates process models for
resource recommendation.
improving learningan
We set forward resource recommendation.
architecture leading to discover process models and recommend to the learner not only learning resource
butWe setprocess
also forwardmodels,
an architecture leading
each of which is to discover
relative to aprocess
specificmodels
learning and recommend
resource. Thesetomodels
the learner
exertnot only learning
a certain influence resource
on the
but also
result process models,
of learning resourceeach of which is relative
recommendation. One of tothea reason
specificthat
learning
endowsresource.
our workThese models
with an exert
original a certain
aspect is thatinfluence on the
it automatically
result of learning resource recommendation. One of the reason that endows our work with an original aspect
analyses event logs based on multi-features extracted from the learner’s profiles. However, the state of the art works require ais that it automatically
analyses event logs
manual analysis stepbased
based on multi-features
on learning resultsextracted
uniquely.from the learner’s
We evaluated profiles. However,
the discovered the state
process models of the art
grounded works
on the require
event logs ofa
manual
Moodle analysis step based
LMS. These on learning
event logs containresults
42,438uniquely.
traces ofWe 100evaluated the discovered
students who processover
learned a course models
one grounded on the event
semester. Results logs of
corroborate
Moodle LMS. These event logs
the good performance of our work. contain 42,438 traces of 100 students who learned a course over one semester. Results corroborate
the good performance of our work.
© 2021 The
© 2021 The Authors.
Authors. Published
Published by
by Elsevier
Elsevier B.V.
B.V.
© 2021
This The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license
is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
(https://creativecommons.org/licenses/by-nc-nd/4.0)
This is an
Peer-reviewopen access
Peer-review under article under
underresponsibility the
responsibilityofofthe CC BY-NC-ND
thescientific license
scientificcommittee
committeeofof(http://creativecommons.org/licenses/by-nc-nd/4.0/)
the KESInternational.
KES International.
Peer-review under responsibility of the scientific committee of the KES International.
Keywords: Process mining ; Learning Management System; Event logs ; Learning resources; Clustering; Recommendation.
Keywords: Process mining ; Learning Management System; Event logs ; Learning resources; Clustering; Recommendation.
1. Introduction
1. Introduction
With the fast pace of modern life, E-learning is taking an increasingly important part from primary school to uni-
WithItthe
versity. fast pace
stands for a of modern life, E-learning
teaching-learning is taking
process where an increasingly
teachers important
and learners part from primary
are geographically school
distant and to uni-
where the
versity. It stands for a teaching-learning process where teachers and learners are geographically distant and where the
latter autonomously manage their activities (learning resources). It allows prevented public to continue their learning,
as we have witnessed with the health crisis associated with the new Coronavirus.
The amount of data produced and traces left by users on Learning Management Systems (LMS) offer the opportu-
nity to provide dashboards and analysis on learners.
These data illustrate: i) the learner’s features that may be stored in his/her profiles (learning style, interests, learning
results, etc.) and ii) the traces of the performed activities, namely event logs. There are currently multiple techniques
for analyzing data from LMS such as Educational Data Mining [2], Process Mining [19] and Educational Process
Mining [5]. In our work, we are basically interested in Educational Process Mining. It allows to discover, monitor, and
improve real process models by extracting knowledge about the learner based on several steps [18]: process discovery,
conformance checking, enhancement and evaluation.
Basically, several research works have been particularly oriented towards enhancing algorithms used for process
discovery (fuzzy miner, inductive miner, etc.) rather than elaborating the quality of the input data. Most of the proposed
works analyse event logs based on learning results uniquely [13, 3, 7]. We judge that event logs and learning results
are insufficient as input data in process discovery. In fact, taking into account another source of data may be beneficial
in order to find the adequate process models for the learner and enhance the learning scenario relative to learning
resources. From this perspective, we propose to consider also the learner’s features since they reflect the learner’s
needs.
In fact, the learner has several activities available to him/her in order to acquire the necessary skills for graduation
in a specific learning resource. We aim to apply process mining techniques from event logs, learning results and
learner’s profiles in order to make useful recommendations to learners that would enable them to build up their own
learning scenario. For this reason, we set forward an architecture to discover process models and recommend to the
learner not only learning resources but also process models, each of which is relative to a specific learning resources.
The discovered process models were assessed based on the event logs of Moodle LMS. These event logs involve
42,438 traces of 100 students who learned a course over one semester. Results reveal that the generated process model
can promote learning resource recommendation for a specific learner.
The rest of this paper is organized as follows. In section 2, we present certain state of art works addressing process
mining and Educational Process Mining . In section 3, we provide an overview of the proposed architecture tackling
process mining and recommendation. In section 4, we focus on describing the process mining layer. In section 5,
experimental results are displayed. Section 6 wraps up the conclusion, exhibits some concluding remarks and offer
new perspectives for future works.
In our previous works, our central focus was upon the research area of recommendation systems. The latter aim
to recommend learning resources such as tags [12], web services [10] and events [17] based on learner’s interests
(expressed by keywords). In this paper, we basically consider not only learner’s interests but also his/her behavior,
expressed by process models, in order to recommend learning resources.
For this reason, we are concerned more with another research area of process mining so as to recommend not only
learning resources but also process models.
In this section, firstly, we discuss briefly process mining. Secondly, we report relevant research papers about Edu-
cational Process Mining. Finally, we compare these different works and we highlight their limitations.
Process mining is used in several areas, including information retrieval in a digital library [16], healthcare [14],
social media [11], etc. The purpose of process mining is to discover, monitor, and improve real processes by extracting
knowledge from event logs which is readily available in current information systems [20].
Each event in such a log refers to an activity (a well-defined step of a process) and is associated with a particular
case (a process instance). Process mining techniques use additional information such as the resource (a person or
device) that performs or initiates the activity, the timestamp of the event, or data items recorded with the event (e.g.,
the size of order) [20].
Wiem Hachicha et al. / Procedia Computer Science 192 (2021) 853–862 855
Wiem Hachicha et al. / Procedia Computer Science 00 (2021) 000–000 3
The process mining techniques are various (specific algorithms or using the notion of a region in graphs, genetic al-
gorithms, etc.) [18]. They aim to extract knowledge about processes by discovering them, checking their conformance
and improving current processes.
The first use of process mining is process discovery. A discovery algorithm takes an event log and generates
a model without using any prior information. Among process discovery algorithms, we cite inductive miner, fuzzy
miner, heuristic miner and alpha miner. These algorithms produce process models in various forms, such as Petri nets,
BPMN models, EPCs, etc.
The second use of process mining is conformance checking. At this level, an existing process model is compared
to an event log of the same process. The conformance checking can be used to check whether the reality, as recorded
in the log, conforms to the model or not.
The third use of process mining is enhancement. The idea lies in extending or improving an existing process
model, thereby using information about the actual process recorded in an event log.
In order to evaluate the results of process discovery algorithms, three main quality dimensions (metrics) are invested
[6]: i) Fitness (F) which determines how well the model allow the behavior present in the event log, ii) Precision (P)
that corresponds to the rate of activities in the event logs compared to the total of activities observed in the process
model and iii) Generalization (G) measures the ability of the model to generalize the behavior present in the event
log.
A suitable model has to find a balance between these metrics.
Educational Process Mining works allow to discover learner process models from event logs for different purposes
such as predicting student performance, adaptation and recommendation.
Authors in [3] explored the relationship between learning behavior and learning progress in MOOCs. They aimed
to better understand how successful and unsuccessful students distribute their activities differently over the weeks’
courses. To find the patterns in the learning behavior of students in the MOOC, an exploratory sequence analysis
using process mining and hierarchical clustering is applied as an excavation method.
The objective of [7] resides in discovering the self-regulated learning (SRL) processes of students during an e-
Learning course using the techniques of process mining. [7] applied the Inductive Miner algorithm to student inter-
action traces in a one-semester online course on the Moodle 2.0 platform. They divided the log file into two groups
resting on the students’ final grades: pass and fail. In addition to pass-fail files, they divided event files into subfiles
by a unit to further analyze students’ behavior and predict their academic performance [8].
In [15], a preprocessing task is performed to group users according to their type of course interactions. This study
allowed to discover the most specific browsing behaviors using the clustered data only rather than the full data set by
applying the Heuristic Miner.
Authors in [13] suggested applying process mining to understand learning processes based on student activity
traces from MOOC platform logs. They divided the students into separate groups in order to improve their analysis.
The grouping criteria are the type of certificate for which the students register and the level of success or the final
grade which they aspire to achieve. They applied the fuzzy miner process mining algorithm in order to visualize and
reproduce the real behavior of the students.
In [1], the authors attempted to identify the self-regulated learning patterns of students with the process mining
technique using data and information extracted from an online learning environment. One of the most frequently cited
self-regulated learning models is the Zimmerman. This model relies on the interaction between three phases of self-
regulated learning, i.e. forethought, performance and reflection. This study reveals that the process mining technique
offers new opportunities to better understand self-regulated learning behaviors in online learning. Using the Fuzzy
model and animation, [1] managed to trace how students adopt self-regulated learning throughout the course.
2.3. Synthesis
Table 1 plots a comparative study about Educational Process Mining works. The comparison rests upon the method
used for analysing event logs, features, the used process discovery algorithm, metrics of evaluation, the data proposed
in the process models and the objectives of the research work.
856 Wiem Hachicha et al. / Procedia Computer Science 192 (2021) 853–862
4 Wiem Hachicha et al. / Procedia Computer Science 00 (2021) 000–000
We notice that despite the efforts provided in certain works in order to brush up the results of process discovery
based on different algorithms, which are in turn evaluated based on the Fitness metric; these works exhibit certain
deficiencies in terms of input data. In fact, most of the proposed works require a step of manual analysis (clustering)
of event logs before process model discovery [15, 7]. Furthermore, the analysis of event logs rests only on learning
results such as certificate [13] and grade [7]. Moreover, some works focus on a specific type of activity (practice
[13, 3] or theoretical [15]). Thus, the discovered process may correspond to the learner’s features.
For this reason, we opt for analysing automatically event logs based on a clustering algorithm. The latter takes
into account not only learning results but also learner’s features that are stored in the learner’s profiles distributed in
different LMS.
This section introduces the architecture of applying process mining for recommendation. This architecture is based
on our previous architecture [9] for learner’s profiles interoperability. This architecture is confined to allowing: i)
the data exchange between heterogeneous learner’s profiles based on an interoperability layer and ii) adapting the
navigation over learning resources to learners based on adaptation layer. However, in this research study, the proposed
architecture is extended by two other layers allowing to apply process mining for recommendation. As depicted in
figure 1, only four layers are illustrated, namely client, recommendation, process mining and source. Additional details
about the interoperability and adaptation layers are recorded in [9].
Within this framework, the client layer allows the interaction between the learner and e-learning systems (LMS).
Thus, the learner can send a request by clicking on the provided links through different types of devices (PC, mo-
bile,...).
The source layer involves distributed databases of learning results, event logs,process models, global profiles and
learning resources.
The learning results include the knowledge or skills learners should acquire by the end of a particular course, activity,
etc. The event logs are files that contain a large amount of raw data about the interaction of learners with LMS. As
data in event logs are often noisy, this database necessitates a preprocessing step before applying the process mining
steps.
The global profile database provides a profile for each learner containg a global view about his/her data that is dis-
tributed in different e-learning systems. Building up the global profile is a preliminary step in the interoperability
layer, which is detailed in [9]. It is represented by the FOAF1 language and involves several features: personal data,
demographic data and social data that characterize the known persons (friends) and interests (the learner’s interests
1 http://www.foaf-project.org/
Wiem Hachicha et al. / Procedia Computer Science 192 (2021) 853–862 857
Wiem Hachicha et al. / Procedia Computer Science 00 (2021) 000–000 5
and those of his/her friends). The interest corresponds to the topic of the learning resources which the learner finds
pleasure in spending time learning about. In this paper, we extend the global profile by new features in order to fur-
ther improve the result of process mining and recommend pertinent learning resources. These features describe the
location of the learner (distant, face-to-face or hybrid), the social aspect (individual or collaborative work) and the
learning style (theoretical or practice aspect).
The learning resources database encompasses any element involved in the learning process such as course, quiz, web
pages, images, videos, etc. These elements can be represented by different standards such as Learning Object Meta-
data2 (LOM) , Information Message Service Simple Sequencing3 (IMS SS) and IMS Learning Design4 (IMS LD).
The process mining layer allows the discovery of process models based on event logs and learning results. Each
process model demonstrates the most common usage behavior by learners in an LMS. It consists of the number of
occurring activities and their frequencies. The analysis of process model allows to visualize and reproduce the real
2 https://standards.ieee.org/findstds/standard/1484.12.1-2002.html
3 https://www.imsglobal.org/simplesequencing/index.html
4 https://www.imsglobal.org/learningdesign/index.html
858 Wiem Hachicha et al. / Procedia Computer Science 192 (2021) 853–862
6 Wiem Hachicha et al. / Procedia Computer Science 00 (2021) 000–000
behavior of the learner, so as to find the patterns in the learning behavior of learners, and even to recommend learning
resources that seem to play an intrinsic role in learning.
For this reason, we propose in the recommendation layer recommending not only learning resources but also
process models based on a process for similarity computing between the process models and the global profile features.
In order to generate process models relative to a specific learner, we propose to cluster event log based on learn-
ing results and learner’s features. From this perspective, the process mining layer branches out into two processes:
clustering and knowledge extraction.
Clustering is applied in order to group learners with similar learning results and features. Learners are grouped
using the k-means clustering algorithm.
First, event logs are split into two clusters based on learners’ learning results: Pass (containing only events of learners
who passed the learning resource) and Fail (containing only events of learners who failed the learning resource).
Second, from the Pass-Fail clusters, we increase the granularity by dividing the event clusters into sub-clusters based
on learner’s features that are stored in the global profile (cf. section 3). The sub-clusters allow to analyze learner’s
behavior more thoroughly.
The algorithm of event log clustering is portrayed in Figure 2. It takes as input the event log, the number of
clusters “kc”, the number of sub-clusters “ksc” and the “clustering-param” which is initialized by “Learning-Results
(pass/fail)”. It applies k-means to generate, for each event log, “kc” clusters according to the parameter “Learning-
Results (pass/fail)”. Afterwards, it applies the k-means algorithm to generate, for each cluster, “ksc” sub-clusters
according to the learners’ features (learning style, location and social behaviour).
Each sub-cluster contains a set of events performed by learners who are very close in their location (face-to-face
study, online learning or hybrid learning), learning style (theoretical or practice activities), social behaviour (individual
or collaborative learning).
Wiem Hachicha et al. / Procedia Computer Science 192 (2021) 853–862 859
Wiem Hachicha et al. / Procedia Computer Science 00 (2021) 000–000 7
The knowledge extraction process allows the extraction of learning scenarios that are commonly used by users
designed in process models. The originality of this process refers to the fact that it rests on the result of event log
clustering. The knowledge extraction takes place according to four steps. The first step stands for a process discovery
by applying a set of process discovery algorithms (cf. section 2.1). This step is mandatory for knowledge discovery
resting on event logs provided by an LMS. Event logs contain all of learners’ events recorded during their interactions
characterized by attributes such as timestamp, learner full name (ID), event context (the course and activities to which
the user have accessed), event name and Description.
As far as this research is concerned, only three attributes are invested corresponding to the most required by pro-
cess discovery algorithms: Learner ID, Event name and timestamp. Each student identified in LMS left a trace of
an event(s). Each event represents an activity that was performed. An activity may be initiated to specific learners
depending on their learning style, location and their social behaviour. Timestamp is the time at which the event was
performed. It is used to identify the order of activities, delay, and bottleneck.
The second step is conformance checking, which is performed in order to check whether the data recorded in
the event log correspond really to the process model or not. The third step is enhancement. It allows to improve an
existing process model. The fourth step allows the assessment of the process model based on evaluation metrics. This
step gives much importance to all metrics by assigning weights. Depending on the evaluation result, the third step
may be re-executed.
The extracted process models are ranked in the recommendation layer in order to recommend to the learner learning
resources with the closest process model (relative to his/her features).
5. Experimental results
In this paper, the experimental results are basically grounded upon the evaluation of the process mining layer (dis-
covered process models). This section reveals, at a first stage, the methodology undertaken for event log preprocessing
in order to prepare for Knowledge extraction. At the second stage, it exhibits the results.
In order to assess the discovered process models, we extracted the event logs of learners who studied the course
“Introduction to human-machine interfaces (HMI)” created on the Moodle platform of the University of La Rochelle
(France). A pedagogical scenario has been established allowing the learner to achieve the final goal. These event logs
include 42,438 traces of 100 students that learned a course over one semester. Each trace corresponds to an activity
performed by a learner.
The log file provided by Moodle involves each learner’s events recorded during his/her interactions with Moodle
summarized in nine attributes: Time, User full name, Affected user, Event context, Component, Event name, Descrip-
tion, Origin, and IP address. We used only three attributes relative to the process discovery algorithm requirement (cf.
table 2).
It was necessary to preprocess and filter the log file. This stands for an essential step when we are using real event
logs and the data is often noisy. For this reason, we convert the time into a format that matches the process discovery
algorithm and filter the activities that are not useful.
860 Wiem Hachicha et al. / Procedia Computer Science 192 (2021) 853–862
8 Wiem Hachicha et al. / Procedia Computer Science 00 (2021) 000–000
Time format conversion: Departing from the collected data, there was a problem with the Time attribute whose
value could not be used as a timestamp. The format of the value did not match any timestamp pattern. Therefore, in
the preprocessing phase, we implemented an algorithm in order to change the format of the Time attribute to conform
to the predefined timestamp.
Filter by instances: The total activities for the students on Moodle before filtering by instances were 42,438. This
filter let us reduce the log file from 42,438 to 36,816 records. The reduction in the number of events led to the reduction
of the number of activities from 64 activities to 28 activities.
5.2. Results
In order to implement our approach, we attempted to explore software and libraries that support process mining.
We found several commercial software and open-source libraries such as Disco5 , ProM6 and PM4Py7 . We chose the
library PM4Py [4] founded by process mining group of Fraunhofer FIT headed by W.M.P. Aalst. This library displays
a relatively wide range of functionalities. It contains basic algorithms for process discovery as Alpha miner, Inductive
miner and Heuristic miner.
As previously mentioned, the process mining layer of the proposed architecture rests on event log clustering and
process discovery. Firstly, we fixed the number of clusters after successive experiments. We divided the event log into
two clusters of learners based on learners’ results (pass and fail). Afterwards, each cluster was divided into 3 clusters
based on the learners’ features (cf. section 4.1). For example, for the pass learners we detected in the first cluster
32% of learners who preferred face-to-face study, collaborative learning and theoretical activities. The second cluster
contains 32% of learners who opted for online learning, collaborative and practice activities. The third cluster contains
34% of learners who would prefer rather Hybrid Learning , collaborative learning and practice activities.
Secondly, we evaluated process mining algorithms resting on the result of event log clustering. In fact, we used
three process discovery algorithms to obtain the models representing the event log’s behavior. These algorithms are:
alpha miner, inductive miner and heuristic miner. Subsequently, the Fitness (F), Precision (P) and Generalisation (G)
of discovered models were measured (cf. Table 3). These metrics are foregrounded in section 2.1.
Table 3. Evaluation metrics obtained on graphs mined from the used data set
Inductive Miner Heuristic Miner Alpha Miner
F P G F P G F P G
All learners 1 0.0367 0.7760 0.9861 0.2284 0.8661 0.7655 0.1052 0.7892
ClusterP0 1 0.0504 0.7928 0.9917 0.1587 0.7717 0.7721 0.1031 0.8067
ClusterP1 1 0.0439 0.7145 0.9874 0.1410 0.7513 0.5552 0.0923 0.7482
ClusterP2 1 0.0505 0.8191 0.9928 0.1520 0.7784 0.5847 0.1183 0.8333
ClusterF0 1 0.0424 0.7508 0.9926 0.1665 0.7825 0.4553 0.1236 0.7873
ClusterF1 1 0.0533 0.8135 0.9889 0.1786 0.7330 0.4539 0.1245 0.8237
ClusterF2 1 0.0391 0.6802 0.9886 0.1613 0.7498 0.4677 0.0998 0.6880
Table 3 displays the obtained values of metrics for the application of each algorithm on the event log of : i) all
learners and ii) each generated cluster (clusters of pass learner: clusterP0, clusterP1 and clusterP2, clusters of fail
learner: clusterF0, clusterF1 and clusterF2). As plotted in the table, the inductive miner proves its high performance
in the fitness value (1) and generalisation values that vary between 0.6802 and 0.8191. The alpha miner confirms
its efficiency in generalisation values (between 0.6880 and 0.8333). However, the heuristic miner generates the best
values of fitness (between 0.9861 and 0.9928), precision (between 0.1410 and 0.2284) and generalisation (between
0.7330 and 0.8661). Consequently, the heuristic miner proves to be the best algorithm as it sets a balance between the
three metrics. Indeed, the discovered process models are indicative of the actual process occurring in Moodle for each
5 https://fluxicon.com/disco/
6 https://www.promtools.org/doku.php
7 https://pm4py.fit.fraunhofer.de/
Wiem Hachicha et al. / Procedia Computer Science 192 (2021) 853–862 861
Wiem Hachicha et al. / Procedia Computer Science 00 (2021) 000–000 9
cluster of learners. Figure 3 presents an extract of the process model generated from Heuristic Miner algorithm. The
output of this algorithm is a Heuristics Net that can be then converted into a Petri net.
Fig. 3. Extract of the process model generated from Heuristic Miner algorithm
This figure illustrates a Petri Net example of the process model discoverd for clusterP1 representing learners who
opted for online learning, collaborative and practice activities. As depicted in the Petri net, the most frequently adopted
activities are performed online, collaborative (Forum add post) and practice (“quiz attempt started”, “quiz attempt
viewed”, “quiz attempt reviewed”, etc.). From this perspective, the type of activity reflects the learner’s features for
this cluster and each new learner belonging to clusterP1, will be guided with these activities.
6. Conclusion
The central objective of this paper is to enhance learning resource recommendation based not only on learner’s
interests [12, 10, 17] but also on learner’s behaviour described by process models. The latter are generated by Ed-
ucational Process Mining techniques. As a matter of fact, we have extended our previously elaborated architecture
proposed in [9] through the addition of a process mining layer. This layer allows to extract process models resting
on event log clustering and knowledge extraction based on multi-features of the learner which are related to his/her
learning results, location, learning style and social behavior. The resulting process models are invested in the recom-
mendation layer in order to suggest to the learner pertinent learning resources. Each learning resource follows the
most adequate process model compared to the learner’s features.
862 Wiem Hachicha et al. / Procedia Computer Science 192 (2021) 853–862
10 Wiem Hachicha et al. / Procedia Computer Science 00 (2021) 000–000
We experimented the proposed architecture, particularly, the process mining layer based on a data set extracted
from the Moodle LMS. The generated results corroborate the effectiveness of the extracted process models and prove
that they can enhance the result of recommendations as they depend mainly on the learner’s features.
The most prominent implications of the research findings reside in identifying learners at risk of dropping out
and potential failing students at an early stage as well as increasing significantly learners’ academic outcomes. These
implications are suggestive that we need to improve the proposed approach. Indeed, we can consider other learners’
features in order to enhance the result of the process mining and recommendation layer.
At this stage of analysis, we would assert that our research is a step that may be built upon and taken further as
it offers fruitful lines of investigation and opens promising research directions. Indeed, in future works, we aspire to
improve the event log clustering by adding other learner’s features in order to extract model processes that simulate
more the learner. Moreover, we shall address the evaluation of the recommendation layer.
References
[1] Bakar, M.H.b.A., Ismail, S., Ali, S.H.S., 2019. A Process Mining Approach to Understand Self Regulated- Learning In Moodle Environment.
International Journal of Advanced Trends in Computer Science and Engineering 8, 195–200.
[2] Baker, R.S., Inventado, P.S., 2014. Educational data mining and learning analytics, in: Learning analytics. Springer, pp. 61–75.
[3] van den Beemt, A., Buijs, J., van der Aalst, W., 2018. Analysing structured learning behaviour in Massive Open Online Courses (MOOCs):
An approach based on process mining and clustering. International Review of Research in Open and Distance Learning 19, 38–60. doi:10.
19173/irrodl.v19i5.3748.
[4] Berti, A., van Zelst, S.J., van der Aalst, W., 2019. Process mining for python (pm4py): bridging the gap between process-and data science.
arXiv preprint arXiv:1905.06169 .
[5] Bogarı́n, A., Cerezo, R., Romero, C., 2018. A survey on educational process mining. Wiley Interdisciplinary Reviews: Data Mining and
Knowledge Discovery 8, 1230–1247.
[6] Buijs, J.C., Van Dongen, B.F., van Der Aalst, W.M., 2012. On the role of fitness, precision, generalization and simplicity in process discovery,
in: OTM Confederated International Conferences” On the Move to Meaningful Internet Systems”, Springer. pp. 305–322.
[7] Cerezo, R., Bogarı́n, A., Esteban, M., Romero, C., 2020. Process mining for self-regulated learning assessment in e-learning. Journal of
Computing in Higher Education doi:10.1007/s12528-019-09225-y.
[8] Chango, W., Cerezo, R., Romero, C., 2021. Multi-source and multimodal data fusion for predicting academic performance in blended learning
university courses. Computers & Electrical Engineering 89, 106908. URL: https://doi.org/10.1016/j.compeleceng.2020.106908,
doi:10.1016/j.compeleceng.2020.106908.
[9] Ghorbel, L., Zayani, C.A., Amous, I., 2015. Improve the adaptation navigation in educational cross-systems. 19th International Conference
on Knowledge Based and Intelligent Information and Engineering Systems,Procedia Computer Science 60, 662–670.
[10] Kalaı̈, A., Zayani, C.A., Amous, I., Abdelghani, W., Sèdes, F., 2018. Social collaborative service recommendation approach based on user’s
trust and domain-specific expertise. Future Generation Computer Systems 80, 355–367.
[11] Li, G., De Carvalho, R.M., 2019. Process Mining in Social Media: Applying Object-Centric Behavioral Constraint Models. IEEE Access 7,
84360–84373. doi:10.1109/ACCESS.2019.2925105.
[12] Mezghani, M., Péninou, A., Zayani, C.A., Amous, I., Sèdes, F., 2017. Producing relevant interests from social networks by mining users’
tagging behaviour: A first step towards adapting social information. Data & Knowledge Engineering 108, 15–29.
[13] Mukala, P., Buijs, J., Leemans, M., van der Aalst, W., 2015. Exploring Students’ Learning Behaviour in MOOCs using Process Mining
Techniques. Computing Conference , 1–12URL: http://bpmcenter.org/wp-content/uploads/reports/2015/BPM-15-10.pdf.
[14] Pika, A., Wynn, M.T., Budiono, S., ter Hofstede, A.H., van der Aalst, W.M., Reijers, H.A., 2019. Towards privacy-preserving process mining
in healthcare, in: International Conference on Business Process Management, Springer. pp. 483–495.
[15] Romero, C., Cerezo, R., Bogarı́n, A., Sánchez-Santillán, M., 2016. EDUCATIONAL PROCESS MINING: A TUTORIAL AND CASE
STUDY USING MOODLE DATA SETS, in: Data Mining And Learning Analytics: Applications in Educational Research. chapter 1, pp.
3–28. doi:10.1002/9781118998205.
[16] Trabelsi, M., Suire, C., Morcos, J., Champagnat, R., 2019. User’s behavior in digital libraries: Process mining exploration, in: International
Conference on Theory and Practice of Digital Libraries, Springer. pp. 388–392.
[17] Troudi, A., Ghorbel, L., Amel Zayani, C., Jamoussi, S., Amous, I., 2020. MDER: Multi-Dimensional Event Recommendation in Social Media
Context. The Computer Journal 64, 369–382. URL: https://doi.org/10.1093/comjnl/bxaa126, doi:10.1093/comjnl/bxaa126,
arXiv:https://academic.oup.com/comjnl/article-pdf/64/3/369/36927838/bxaa126.pdf.
[18] Van Der Aalst, W., 2012. Process Mining: Overview and Opportunities. ACM Transactions on Management Information Systems (TMIS) 3,
1–17. doi:10.1145/0000000.0000000.
[19] Van Der Aalst, W., 2016. Process mining: Data science in action. Springer , 1–477.
[20] Van Der Aalst, W., Adriansyah, A., De, A.K.A., Medeiros, Arcieri, F., 2012. Process Mining Manifesto. Business Process Management
Workshops , 169–194.