Architecture for Gradually AI-Teamed Adaptation Rules in Learning

Management Systems

Niels Seidel

CATALPA, FernUniversit

at in Hagen, Universit

atsstr. 27, 58097 Hagen, Germany

Keywords:

Adaptive Learning, Learning Management System, Self-Regulated Learning.

Abstract:

Over the last two decades, there has been a substantial advance in developing adaptive learning environments.

However, current adaptive learning environments often face limitations, such as tailoring to speciﬁc contexts

or courses, relying on limited data sources, and focusing on single adaptation goals (e.g., knowledge level).

These systems commonly use a single data mining approach and are often tested in isolated studies, restricting

broader applicability. Integration with mainstream Learning Management Systems (LMS) also remains chal-

lenging, affecting accessibility and scalability in education systems.

In this paper, we present a system architecture for authoring and executing adaptation rules to support adap-

tive learning within Moodle, a widely used LMS that focuses on enhancing self-regulated learning. Using AI

methods like rule mining, clustering, reinforcement learning, and large language models can address some of

the known disadvantages of rule-based systems. In addition, the support of the adaptation rules can be quanti-

ﬁed and simulated using weekly user models from previous semesters.

Leveraging an active distance learning course, our investigation reveals an AI-teamed process for identifying,

deﬁning, and validating adaptation rules, ensuring the harmonized execution of personalized SRL feedback.

1 INTRODUCTION

Over the last two decades, there has been a substantial

advance in the development of adaptive learning envi-

ronments (Martin et al., 2020; Xie et al., 2019). This

trend can be attributed to several factors, including the

evolving availability of student data, the increasing

importance of online learning, the advances in AI, and

the awakening awareness of educational institutions

to address student diversity (De Clercq et al., 2020).

Consequently, numerous researchers have taken steps

to personalize learning activities and steer toward au-

tomating interventions based on learning analytics by

using large amounts of more and more ﬁne-grained,

multimodal, and multichannel data (Molenaar et al.,

2023).

Recent literature reviews have methodically ex-

plored the multifaceted aspects of adaptive learning

environments. These examinations have shed light on

how learner characteristics are harnessed within these

settings (Vandewaetere et al., 2011; Normadhi et al.,

2019), the artiﬁcial intelligence methods employed

(Almohammadi et al., 2017; Kabudi et al., 2021), and

https://orcid.org/0000-0003-1209-5038

the structure of learner models (Nakic et al., 2015).

Additionally, the conceptualization of learning ob-

jects (Apoki et al., 2020) and the approaches of adap-

tive feedback (Bimba et al., 2017) have been scruti-

nized. Concurrently, emerging trends within this do-

main have been identiﬁed and documented (Martin

et al., 2020; Xie et al., 2019), enriching the discourse

on adaptive learning.

Current adaptive learning environments, while pi-

oneering in their respective domains, exhibit a range

of limitations that merit consideration. Primarily,

these systems are tailored for particular didactic con-

texts, including speciﬁc courses, subjects, and target

audiences, which may restrict their broader applica-

bility. They typically draw insights from a ﬁnite ar-

ray of data sources incorporated as a ﬁxed feature set

within the trained model, potentially overlooking the

rich tapestry of learner characteristics, interactions,

and behaviors. Moreover, the focus of these envi-

ronments often narrows down to a single adaptation

goal (e.g., knowledge level), thereby limiting the mul-

tifaceted nature of learning processes. The predomi-

nant reliance on a single data mining approach fur-

ther conﬁnes the scope of these environments. Addi-

tionally, most of these systems are grounded in singu-

Seidel, N.

Architecture for Gradually AI-Teamed Adaptation Rules in Learning Management Systems.

DOI: 10.5220/0013215800003932

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Computer Supported Education (CSEDU 2025) - Volume 1, pages 243-254

ISBN: 978-989-758-746-7; ISSN: 2184-5026

243

lar, non-replicated studies, with a signiﬁcant portion

only testing models on a dataset (e.g., (Duraes et al.,

2019)). Finally, integrating these adaptive learn-

ing solutions into widely-used, off-the-shelf Learn-

ing Management Systems (LMS) remains a challenge

(Grubi

c et al., 2015; Kopeinik et al., 2014), hin-

dering their accessibility and scalability within the

broader educational landscape.

Another shortcoming observed in adaptive learn-

ing environments is their limited support for SRL in

practice. Empirical studies underscore the pivotal role

of self-regulated learning (SRL) skills in promoting

effective and efﬁcient learning, as highlighted in the

works of (Jivet et al., 2017) and (Sghir et al., 2023).

SRL, encompassing a diverse set of strategies and

processes such as goal setting, monitoring progress,

modifying behavior adaptively, assessing outcomes,

and engaging in reﬂection, is extensively described

in the study by (Wu et al., 2024). Students adept

at self-regulation are observed to gain a spectrum of

academic and extracurricular beneﬁts over their less

self-regulated peers. However, challenges in reﬂec-

tive thinking and effective progress monitoring in line

with their learning objectives are common obstacles,

as discussed by (Radovi

c et al., 2024).

Many tools have been developed to scaffold SRL,

aiming at various levels and through different means

such as goal setting, monitoring, and reﬂection (Perez

Alvarez et al., 2022). Approximately one-ﬁfth of

these tools offer some form of adaptive personaliza-

tion through recommendations. The corresponding

scientiﬁc literature on these tools often focuses on

conceptual frameworks (Yau and Joy, 2008; Nuss-

baumer et al., 2014) or the design phase (Yau, 2009)

and development (Kopeinik et al., 2014; Alario-

Hoyos et al., 2015; Renzel et al., 2015; Fruhmann

et al., 2010). Reports of smaller lab studies or appli-

cations in public sandboxes are rare (Gasevic et al.,

2012; Nussbaumer et al., 2015), and even recent liter-

ature includes few accounts of practical deployment

in real-world teaching scenarios (Wu et al., 2024)

(Seidel et al., 2021)

The gap between theoretical concepts, prototypes,

and tools educators can adapt and use for SRL sup-

port is noticeable. Educators’ role is vital, as they

are responsible for adaptive learning offerings and

must be equipped to guide adaptations effectively. In

addition, educators must be able to use tools to in-

struct their AI-based assistants as they desire. To

achieve widespread deployment of adaptive systems

across various universities and disciplines, these must

be built on existing learning environments, expanded

into adaptive learning environments, and developed in

close collaboration with educators to ensure meaning-

ful integration into educational practice.

This article tackles the effective implementation

of adaptive learning into teaching practice. The cor-

responding research question (RQ) is: How can a

scalable software architecture be designed and im-

plemented to support adaptive learning in an LMS

that dynamically personalizes the learning experience

while educators keep the learning environment under

control?

The contribution to research made here consists

of an expandable system for implementing, adjusting,

and monitoring adaptive learning scenarios within the

widely used open-source LMS Moodle. This adap-

tive system is based on preprocessed learner models

and facilitates dynamic, transparent adaptations, un-

like traditional expert systems or Intelligent Tutoring

Systems. Furthermore, we outline a methodology for

developing adaptation rules, involving various learn-

ing analytics methods and closely involving educa-

tors. We demonstrate the application of this system

and the developed adaptation rules in a currently ac-

tive course in which educators can monitor and mod-

ify the adaptations.

Section 2 introduces the architecture of a rule-

based adaptive learning environment tailored as a plu-

gin for the Moodle LMS. Section 3 presents a system-

atic approach that enables educators to identify, ad-

just, implement, and validate adaptation rules through

the assistance of AI methodologies. Following this,

we present an example of a course implemented with

comprehensive SRL support. The article concludes

with a discussion and critical analysis of the presented

AI-teamed rule-based systems, evaluating their prac-

tical relevance.

2 SYSTEM ARCHITECTURE

This section will present the architecture of a rule-

based adaptive learning environment tailored as a

Moodle plugin. The development goal was to harness

data from the learning process for modeling learners,

subsequently utilizing straightforward and transpar-

ent if-then rules to facilitate a range of personaliza-

tions at various levels within the LMS, including all

kinds of existing plugins. In addition to the execu-

tion of adaptation rules, the system is designed to em-

power educators, enabling them to create their own

rules and monitor and adjust them in real-time during

course operations. The source code is publicly avail-

able under the GPLv3 license:

https://github.com/CATALPAresearch/local ari.

CSEDU 2025 - 17th International Conference on Computer Supported Education

244

Server

Client

Moodle Database

Learner Model Recommender

Content Model

Moodle Web Service API

Indexed DB

Rule Set

Rule Manager

Sensors

Exams

Enrollments

Courses Catalog

Usenet Server

POS

SLO

feunews

text-based

learning objects

websockets

response

update

sync

load

load LM

send

calls

actor status

execute

actor

traces

feedback text

update

Another

Moodle

Plugin

Modeling

SOAP Interfaces

Examinations

SLO

OÜ

Data sources

Data warehouse

Extract, load, transform

Adaptation

ARI

User view in Moodle

University API

LLMRule Mining

Actors

Teacher View:

Adaptation Board

Student View

Figure 1: System architecture.

2.1 Overview

Describing the system architecture helps distinguish

between the ﬁve layers shown in Fig. 1.

The university’s data sources are the foundational

layer of our architecture. These encompass vari-

ous databases containing enrollment data, examina-

tion records, platforms for online quizzes, and news-

group servers.

Positioned above this is a layer dedicated to ex-

tracting, loading, and transforming data for subse-

quent processing—the ELT layer. Within this layer,

a component named University API undertakes these

tasks. It extracts and loads data via SOAP or Web-

Sockets from the underlying layer and transforms it

through preprocessing pipelines into appropriate data

structures. Some of these data structures are then

replicated in the Moodle database in the above layer.

Depending on the frequency of data changes, these

are updated daily, monthly, or at the commencement

of each semester and are stored in a processed form

for efﬁcient utilization.

The third layer of our system, the data and model

warehouse, is home to the Moodle database and the

content model. The Moodle database stores learning

outcomes (such as solutions to assignments) and, cru-

cially, processes data from the recorded interactions

between users and the learning objects within a Moo-

dle course.

The content model, on the other hand, harbors the

keywords extracted from the texts of the learning ob-

jects, providing a rich layer of information crucial

for understanding and enhancing the learning experi-

ence. This integration of detailed interaction data and

content-speciﬁc keywords forms the backbone of our

system’s ability to deliver a tailored and responsive

educational environment.

The fourth layer of our architecture is designated

for components that implement the adaptation of the

learning environment. The system presented here is

accomplished through the Adaptation Rule Interface

(ARI) and several supporting web services.

ARI initially provides interfaces to the Moodle

database via the learner model, to the content model

via the recommender, to the set of adaptation rules de-

ﬁned by educators in the adaptation rule board via the

rules stored in the ruleset, and ﬁnally to the learning

environment, where actors implement the rule-based

decisions. The central element of ARI is the rule

manager, which, with the help of data in the learner

model, checks whether the conditions of the adapta-

tion rules stored in the ruleset are met. Decisions on

whether and how a rule becomes effective for a stu-

dent in the learning environment are ultimately made

by incorporating sensor information from the learn-

ing environment and using a reinforcement learning

model.

In addition to ARI, this layer includes web ser-

vices that, for example, analyze speciﬁc data (e.g.,

clustering, sequence mining), determine recommen-

dations for learning objects (recommender), or trans-

form text prompts into well-formulated texts (e.g.,

Language Learning Models).

Architecture for Gradually AI-Teamed Adaptation Rules in Learning Management Systems

245

2.2 Learner Model

The learner model consolidates data from both uni-

versity administration systems and the learning en-

vironment, creating a comprehensive proﬁle of each

student. Sociodemographic information, enrollment

details, courses taken per semester, and academic per-

formance metrics are sourced and regularly updated

from university administrative systems. The learning

environment is facilitated by a research instance of

the Moodle LMS, which is enhanced with an array

of plugins. These plugins collect high-resolution data

on learner activities, such as scrolling actions on text

pages. Utilizing the browser’s Intersection Observer

API, the system discerns active reading by tracking

screen time for speciﬁc text sections, differentiating

it from scrolling and text searches, thereby estimat-

ing reading duration. Similarly, the video-watching

behavior is tracked for segments of two seconds. Fur-

thermore, interactions like mouse-overs are logged in

the learning analytics dashboard, offering insights

into the duration of content engagement and the in-

tensity of usage, painting a detailed picture of the

learner’s interaction with the educational content.

The learner models encapsulates a suite of met-

rics for individual activities within a course, such as

quizzes, assignments, videos, and extended text sec-

tions (referred to as longpage), as well as a compre-

hensive summary of all learning activities contained

in the course. For each learning activity and the

course as a whole, metrics are provided, including the

ﬁrst and last access, the number of sessions, the aver-

age session duration, the number of active days, and

the total time spent. Moreover, activity-speciﬁc met-

rics are included; for instance, in the case of reading

course texts (longpage), these encompass the propor-

tion of text read and the number of marked or com-

mented text passages. When a course is structured

into multiple units, the metrics for each activity are

recursively integrated into the learner model for the

respective course unit.

Additionally, the learner model includes data on

the number of courses enrolled in and repeated and

a list of all courses previously undertaken, offering a

holistic view of the learner’s engagement and progres-

sion through the curriculum.

In operational terms, the learner model is

computed within milliseconds, utilizing optimized

database queries, and the resultant data is relayed via

a REST API in JSON format. The input parame-

ters for a request include a user identiﬁer, a course

number, and an optionally speciﬁed period. This ef-

ﬁcient infrastructure enables the learner models to be

accessed by various plugins within Moodle, contin-

gent upon the user’s authorization. Consequently, all

metrics can be retrieved for a deﬁned temporal seg-

ment, facilitating comparative analyses and extrapo-

lating temporal trends and learner progress.

2.3 Content Model and Recommender

A relatively straightforward recommendation system

has been constructed to facilitate recommendations

for subsequent learning steps. This system considers

the similarity of text-based learning objects and the

progression or learning success, as indicated in the

learner models.

Initially, texts from text-based learning objects are

extracted from the Moodle database, including details

like the type of learning object (e.g., quiz, assignment,

longpage) and an identiﬁer. Lengthier texts, such as

those in longpages, are subdivided into sections using

headings and subheadings. Following a preprocessing

step (which includes the removal of stopwords, num-

bers, and URLs), keywords are extracted using NLTK

Rake (based on (Rose et al., 2010)) and subsequently

reduced to their stems. In the case of the German lan-

guage, compound nouns are split apart. Furthermore,

keywords that primarily pertain to the context of the

learning setting (e.g., course, task) are removed from

the keyword list. The number of common keywords

determines the similarity of two learning objects.

Utilizing the learner model and additional infor-

mation from the Moodle database, the system can

generate a prioritized list of recommendations with

URLs to the relevant learning resources. These rec-

ommendations include quizzes related to already-read

sections, self-assessments corresponding to a partic-

ular section in the course text, and self-assessments

similar to submitted and pending assignments.

2.4 Adaptation Rule Interface

ARI is a so-called local plugin for Moodle, imple-

mented in TypeScript on the client and PHP on the

server side. Through the nature of local plugins, it is

loaded on every page in Moodle. Foremost, ARI is

an interface between the trace data and learner model

on the one side and the perceivable adaptations in the

browser view on the other. It enables adaptations that

a learner experiences visually and interactively. ARI

is used to check the conditions that must be met for

an adaptation rule to execute, using actual data from

the learner model (LM, Fig. 1), and to determine and

initiate speciﬁc actors (Actors, Fig. 1), taking into ac-

count contextual information from the user session

(Sensors, Fig. 1). Which input parameters are linked

to which outputs are deﬁned in adaptation rules (Rule-

CSEDU 2025 - 17th International Conference on Computer Supported Education

246

set, Fig. 1).

The adaptation rules were initially formulated

in natural language using ﬁll-in parameters, inde-

pendent of any speciﬁc subject or domain: In

a certain <situation> in a <period of time>

characterized by the <key indicator> which is

<like> a <value> in a <source context>, sup-

port the learner on the <target context> so that

in the <area> the <action> is performed providing

<information>.

The sensors determine, for example, the page cur-

rently being viewed in Moodle, the position on this

page, or the respective user activity or inactivity. Ac-

tors can access Moodle-speciﬁc actions such as sys-

tem notiﬁcations, messages, or even modal dialogues

and provide them with speciﬁc information. Actors

can also manipulate the page layout by specifying

CSS, for example, to highlight elements or change

their position and order. However, the so-called

stored prompts is the most ﬂexible actor. Prompts

to be executed or displayed in another Moodle plu-

gin (e.g., longpage) are stored in the browsers’ in-

dexedDB. Thus, a plugin can listen to changes in the

indexedDB store and execute actions in the deﬁned

way.

Sensors, actors, the rule set, and the learner model

are designed modularly and can therefore be ﬂexibly

extended. The reaction and response behavior of the

learners to the system’s interventions is continuously

captured and incorporated into the decision to trig-

ger subsequent actors using a reinforcement learning

model. In this way, the adaptations desired by the user

and provided by a so-called agent can be favored by

rewards, and the adaptations perceived as disturbing

in the respective situation can be avoided (loss) in the

future. The reinforcement learning agent is part of the

rule manager.

In reinforcement learning, an agent independently

learns a strategy to maximize rewards as deﬁned by

a reward function. In our context, the objective is to

elicit a positive user response to actions carried out by

the agent. Speciﬁcally, these actions are those iden-

tiﬁed by the rule manager from a set of adaptation

rules for execution. Positive reactions, or rewards,

are quantiﬁed based on various metrics depending on

the action. These may include reception (yes/no), du-

ration of engagement, utilization (such as clicking a

provided link), or explicit user feedback through a rat-

ing mechanism.

The agent interacts with the learning environment

through its actions at discrete time intervals, receiving

a reward for each interaction. The strategies available

to the agent pertain to both the nature of the action

and its urgency. These strategies regulate the attention

drawn and the intrusiveness of an adaptation rule’s ac-

tion as experienced by the user.

Implementing this reinforcement learning model

leverages the policy-gradient method in Tensor-

Flow.js, enabling a sophisticated and responsive

learning environment that adapts to the nuanced needs

of the users.

The adaptation rule board (Fig. 2) is conceptual-

ized as a cockpit for educators, offering a comprehen-

sive overview of the adaptation rules deﬁned for each

course. Within this interface, rules can be activated

or deactivated, and the total number of rule execu-

tions and the count of students affected by each rule

provide insightful metrics regarding the utilization of

these rules. The conditions and actions of each rule

are summarized in a format accessible to humans for

ease of understanding and management.

In the editing mode, educators can ﬁne-tune the

rules—for instance, by selecting variables from the

learner model, setting threshold values, and deﬁning

comparison operators. One or more actors can be se-

lected and conﬁgured on the actions side. For exam-

ple, suppose the aim is to issue feedback on SRL. In

that case, educators can incorporate variable values

from the learner model into the text using placehold-

ers or other placeholders to deﬁne a list of learning

resources recommended by the Recommender sys-

tem. Additionally, if the text is intended to serve as a

prompt for an LLM, this can be speciﬁed accordingly,

further enhancing the adaptability and functionality of

the system.

Figure 2: Adaptation rule board including actions as High

Information Feedback.

3 IDENTIFICATION,

DEFINITION, AND

VALIDATION OF ADAPTION

RULES

This section presents a structured approach for iden-

tifying, deﬁning, and validating adaptation rules,

encapsulating these processes within three distinct

phases. Each phase comprises three steps, dur-

ing which the objectives are set, multiple ap-

Architecture for Gradually AI-Teamed Adaptation Rules in Learning Management Systems

247

proaches—often gradually employing a variety of

AI methodologies—are explored, and the anticipated

outcomes are articulated. It presents ﬁndings from

a user-centered design approach involving educators

and offers tangible examples to contextualize the pro-

cedure. Fig. 3 visually maps out the interplay of these

nine steps in conjunction with the ARI framework in-

side the Moodle LMS.

3.1 Identiﬁcation

3.1.1 Step 1: Determine Adaptation Level

The adaptation of a learning environment can be

aimed at varying targets (adaptation targets (Vande-

waetere et al., 2011)) across different levels. These

levels refer to frames of reference such as speciﬁc

tasks, classes of tasks, types of learning activities,

course units, entire courses, or even multiple courses

as part of a degree program.

It is pragmatic to initially focus on a single level

in the development of adaptation rules. Support for

SRL can fundamentally occur at different levels, each

offering unique opportunities and challenges in cus-

tomizing the learning experience. This stratiﬁed ap-

proach allows for more targeted and effective adap-

tation, ensuring that the needs of learners at various

stages and contexts are adequately met.

• Task level (T): At this level, feedback and hints

are provided to assist or guide students in improv-

ing speciﬁc tasks or multiple tasks of the same

type. This approach is tailored to address im-

mediate task-related challenges, enhancing task-

speciﬁc competencies.

• Course unit level (U): Adaptations on the course

unit level offer hints and recommendations related

to the learning resources and associated learning

activities within a particular course unit. This

level focuses on integrating and understanding re-

sources and activities conﬁned to a speciﬁc seg-

ment of the broader course structure.

• Course level (C): On the course level, hints

and recommendations encompass all learning

resources and associated activities within the

course. This broader perspective aims to provide

a cohesive learning experience, ensuring that all

course elements are aligned and contribute effec-

tively to the learning objectives.

• Programme level (P): At the program level, adap-

tations may include suggestions for the sequence

or combination of modules to be chosen within a

degree program. These adaptations are geared to-

wards long-term academic planning and progres-

sion, helping students navigate the curriculum that

aligns with their academic and career goals.

Beyond these four levels of adaptation, it is cru-

cial to consider scenarios where learners may become

disengaged or absent from the learning environment.

This is categorized as Inactivity (I), a state where

there has been no interaction with the learning offer-

ings to date or a signiﬁcant lapse in meaningful inter-

actions over time. Addressing inactivity is essential

to re-engage learners and ensure the learning envi-

ronment’s effectiveness, maintaining its relevance and

impact across diverse learner populations and circum-

stances.

3.1.2 Step 2: Outline Personas

To formulate adaptation rules, we must ﬁrst acknowl-

edge the high diversity among students in terms of

personal characteristics and learning behavior within

a particular course. To support SRL, individual

progress in the learning process and achievements are

key indicators for assessing learning status. However,

these indicators must be considered in light of tem-

poral patterns. Analyses from past semesters reveal

that alongside students who learn more or less con-

tinuously, some begin learning activities with several

weeks’ delay or discontinue them after only a few

weeks or months (Menze et al., 2022).

Drawing inspiration from the concept of personas

(Nielsen, 2014), these two indicators were used to de-

scribe three dominant temporal patterns, thus creat-

ing prototypical learner proﬁles. Learning progress

and learning success are differentiated on the nomi-

nal scales of low, medium, and high. The resulting 27

personas, represented in Tab. 1, serve as a framework

rather than personiﬁed characters. This ensures that

adaptation rules are devised for each persona, guaran-

teeing comprehensive coverage and support.

It is anticipated that all personas, and thereby the

different groups of learners they represent, should be

supported through adaptation rules. This approach al-

lows for the prioritization of groups that might bene-

ﬁt more from adaptive support, thereby enhancing the

effectiveness and inclusivity of the learning environ-

ment.

3.1.3 Step 3: Identify Situations Where Students

Require Support or Guidance

Adaptation rules within a learning system are de-

signed to specify situations where conﬁguration mod-

iﬁcation is warranted. For instance, if a learner un-

successfully attempts an assessment more than three

times, the system might suggest switching to a dif-

ferent type of feedback or assessment. These scenar-

CSEDU 2025 - 17th International Conference on Computer Supported Education

248

Identification Definition Valildation

ARI

Learner Model

Learning Analytics

Data Mining Generative AI

Step 1: Adaptation Level

Step 2: Outline Personas

Step 3: Identify Situations

Step 4: General Assumptions

Step 5: Rule Mining

Step 6: Define Rule Actions

Step 7: Adjust Thresholds

Step 8: Simulation

Step 9: Monitoring

Figure 3: Procedure with three steps each for the identiﬁcation, deﬁnition, and validation of adaptation rules. The arrows

point to AI-related research areas whose methods can be applied in the respective step.

Table 1: Personas considering learning progress, learning success, and temporal patterns. The task (T), course unit (U), course

(C), programme (P), and inactivity (I) levels relevant for adaptation are speciﬁed for each persona.

Progress Success Temporal patterns

Continuous learning Early drop-out Late-comer

low low T/U/C I I

low medium T/U/C I I

low high T/U/C I I

medium low T/U/C I I

medium medium T/U/C I I

medium high T/U/C I I

high low T/U/C I I

high medium T/U/C I I

high high T/U/C P P

ios are characterized using features from the learner

model.

To illustrate how situations triggering adaptations

are characterized, consider an analogy with self-

regulated walking. A system assisting a person in

walking would need to detect patterns such as short

strides (amplitude), slow leg movements (frequency),

hopping on one leg (variance), backward steps (se-

quence), pauses (continuity), and overall time taken

(duration) to provide relevant feedback. Similarly, in

SRL, amplitude refers, for instance, to deviations in

assessment scores from a learner’s average or peer

group; frequency captures how often students open

assignments without completing them; variance re-

ﬂects engagement differences, such as focusing on

reading while avoiding assessments; sequence ac-

counts for deviations from an intended task order;

continuity highlights learning interruptions that may

signal the need for intervention; and duration helps

estimate whether a student is genuinely engaging with

a task or merely guessing.

Acknowledging these six dimensions is instru-

mental in identifying characteristic situations where

learners may require support. These dimensions,

therefore, guide data analyses that could suggest en-

hancements to the learner model, expanding its ca-

pability to respond to diverse learning scenarios ef-

fectively. This approach underscores the importance

of a multidimensional analysis in adaptive learning

systems, ensuring a nuanced understanding of learner

needs and behaviors. However, a multidimensional

analysis requires a deep discussion with the educators

reponsible for a class.

3.2 Deﬁnition

3.2.1 Step 4: General Assumptions and Beliefs

Instructional strategies in the classroom are often

shaped by the educator’s perspectives and principles

on what constitutes effective learning.

In line with the six dimensions previously in-

troduced for identifying characteristic situations that

may necessitate adaptation, together with responsi-

ble educators, we have identiﬁed and discussed their

foundational assumptions in a long-standing distance

learning course. The deliberation on these founda-

tional assumptions of teaching practice is particularly

pertinent when considering which conditions should

be deﬁned as part of the adaptation rules. Further-

Architecture for Gradually AI-Teamed Adaptation Rules in Learning Management Systems

249

more, these assumptions form the basis for generaliz-

able adaptation rules that could be applicable across

multiple courses. Consequently, educators would not

have to start from scratch in deﬁning adaptation rules

but could instead create an adaptive, personalized

learning environment by adjusting these generalizable

rules. The following beliefes have been contributed

by the involved educators:

• Frequency: Frequent failures should be avoided.

• Sequence: Follow the educator’s intended order

of learning activities, e.g., from the ﬁrst to the last

course unit.

• Continuity: Continuous learning is better than

having learning breaks of several weeks. Those

who dropped out for some time should return.

• Duration: The more time you spend (actively) on

the course, the better you get.

• Variance: Learners should use receptive (e.g.,

reading) and productive (e.g., assessments) ac-

tivities. Learners should make use of all types

of provided learning material/activities. Learners

should mostly learn individually but also interact

with other students.

• Amplitude: Higher engagement is better than low

engagement. Higher progress is better than low

progress. Higher success is better than low suc-

cess.

As a result of this process, at least one condition

was articulated in natural language for each of the

six dimensions. Utilizing the adaptation rule board

(cf. 2.4), these conditions can be formally established

by selecting the relevant variables from the learner

model and comparing them to an initially estimated

threshold value using a comparison operator. These

threshold values are further reﬁned in Step 7. Ad-

ditionally, we posit that some of these assumptions

may apply to other courses, particularly within the

same discipline, thus presenting potential candidates

for general adaptation rules within a speciﬁc ﬁeld of

study. The general assumptions communicated by the

educators also help to consolidate the learner model

and, for example, to supplement missing indicators.

3.2.2 Step 5: Rule Mining

Adaptation rules in educational settings can be un-

derstood as the decisions an educator might make to

achieve speciﬁc goals. These goals are often rep-

resented as dependent variables, identiﬁable through

data analysis of past learner cohorts. Through rule-

mining processes, models in the form of decision lists

or sets can be created.

Sequential covering, a standard method in rule

learning algorithms like Ripper (Cohen, 1995), op-

erates by iteratively learning rules and excluding data

points already covered by new rules. Additionally,

Bayesian Rule Lists (Letham et al., 2015) can be uti-

lized to construct decision lists comprising sequences

of if-then statements. A more recent approach is Ex-

plainable Neural Rule Mining (Shi et al., 2022), em-

ployed for identifying causal patterns from neural net-

works.

The outcome of rule mining is typically a collec-

tion or list of conditions, enabling an assessment of

the support and conﬁdence of these conditions in vari-

ous scenarios. This process allows for a deeper under-

standing of the effectiveness and applicability of spe-

ciﬁc educational strategies, informed by data-driven

insights. We have applied the three mentioned rule

mining approaches to identify rule candidates to be

discussed with the educators.

3.2.3 Step 6: Deﬁne Rule Actions

Upon deﬁning, adjusting, and validating the condi-

tions of an adaptation rule, the subsequent step spec-

iﬁes the action the system should automatically exe-

cute when those conditions are met. Within the ARI,

various actions are supported (cf. 2.4). Literature pro-

vides examples of actions to emphasize, rearrange,

or conceal elements within the learning environment

(e.g., (Brusilovsky, 2007)). However, formulating ef-

fective feedback presents a more complex challenge

(Bimba et al., 2017).

Feedback is most impactful when it is rich in

information. Simple forms of reinforcement and

punishment are less effective compared to high-

information feedback (HIF), which is considerably

more beneﬁcial (Wisniewski et al., 2020). HIF details

tasks, processes, and, occasionally, self-regulation

levels. It is most effective when it helps students un-

derstand not just their mistakes but also the reasons

behind them and strategies to avoid similar errors in

the future.

There are three approaches to implementing this:

The simplest method involves crafting HIF as static

predeﬁned text. To better tailor the feedback to the

recipient, variables from the learner model can be in-

tegrated into the text using placeholders. For exam-

ple, the expression quiz.last access in the feedback

text could dynamically reﬂect the last date the stu-

dent accessed a quiz activity. Similarly, placehold-

ers can be used to insert feed-forward recommenda-

tions for subsequent learning steps, allowing for adap-

tive guidance without the educator needing to prede-

ﬁne speciﬁc learning objects. The third approach to

feedback formulation involves utilizing a Large Lan-

CSEDU 2025 - 17th International Conference on Computer Supported Education

250

guage Model (LLM) (e.g., (Wu et al., 2024)). This

method involves creating prompts by using the tem-

plate pattern (White et al., 2023) with a set of param-

eterized statements about the learner’s progress (HIF

feed-back) and recommendations for the next steps

(HIF feed-forward), which are then dynamically gen-

erated as a feedback message by the LLM at runtime.

3.3 Validation

3.3.1 Step 7: Adjust Thresholds

When manually deﬁning conditions, educators often

ﬁnd it challenging to set appropriate threshold values

for the variables used in these conditions.

To aid in determining a suitable threshold value,

historical data from learner models of previous years

are utilized. Two key visualizations are provided for a

given variable in the learner model. Firstly, the distri-

bution of values throughout the semester is depicted

in a histogram. Secondly, the changes in the vari-

able over time are represented, showcasing its tem-

poral evolution.

These visualizations serve a dual purpose. First,

they enable educators to identify meaningful value

ranges for the variables and the critical periods when

an adaptation rule should be triggered. Second, this

approach simpliﬁes the process of setting thresholds

and ensures that the adaptation rules are grounded in

empirical data, enhancing their relevance and effec-

tiveness.

3.3.2 Step 8: Simulation of Adaptation Rules

Simulating these rules using data from past semesters

is recommended to validate a set of adaptation rules

and ensure they are reliable and effective.

In the simplest case, for a rule r, the support could

be calculated as the percentage of students s to which

the conditions c = c

, ..., c

of a rule r applied.

support(r) =

|s|

∑

s=1

(

1, if c = true, ∀ c ∈ r

0, otherwise

(1)

The necity metric indicates how many students

could not independently move out of the value range

covered by the condition of this rule in the following

week. This metric, therefore, reﬂects the number of

students who would not have improved independently

(e.g., without adaptive feedback).

necity(r, w) = 1 −

w+1

(r)|

(2)

Here, |a

w+1

(r)| represents the number of students af-

fected by the rule r in the subsequent week w+1, and

(r)| the number of students affected by the rule r

in the current week w. Smaller values for necity sug-

gest that the rule is redundant. If values are less than

0.5, the condition of a rule should be adjusted.

With support and necity, no differentiation has

been made yet for temporal changes and different

groups of people. Therefore, it may be useful to cal-

culate both metrics for each week and the personas

deﬁned in Step 2.

Finally, it is also important to check whether rules

overlap or contradict each other. For this purpose, the

conditions of all rules are mapped out in a tree. Con-

ditions that occur frequently appear near the root of

the tree.

3.3.3 Step 9: Monitor Implemented Rules

In the operational phase of a course, it is essential

to monitor regularly and, if necessary, adjust the au-

tomated components of learning support, as empha-

sized by (Tabebordbar and Beheshti, 2018).

The ARI facilitates monitoring adaptation rule us-

age within the adaptation rule board. Therefore, the

total number of rule executions and the number of

students affected by the rule are listed. The obser-

vation period can be deﬁned using ﬁlters. Following

a methodology similar to the work of (Tabebordbar

and Beheshti, 2018), user feedback (e.g., user rating

of feedback messages) is also considered.

This approach provides educators with an

overview of the adaptation processes, thus reﬂecting

the degree of personalization achieved in a course.

They can make ad hoc adjustments to modify rules

that demonstrate either too low or too high sup-

port, enhancing the effectiveness and relevance of the

learning support system.

4 ADAPTIVE SUPPORT FOR

SELF-REGULATED LEARNING

In this section, we describe the setting in which adap-

tive support for SRL is currently used in a course.

SRL Support Instruments. The course we stud-

ied was designed as a component of the complete

distance and online bachelor’s degree programs in

Computer Science. Students were studying a mod-

ule called ”Operating Systems and Computer Net-

works” composed of four learning units: Devices and

Processes (Unit 1), Memory and File Systems (Unit

2), Applications and Transport (Unit 3), and Medi-

ation and Transmission (Unit 4). During a period

Architecture for Gradually AI-Teamed Adaptation Rules in Learning Management Systems

251

of 11 weeks, students worked individually by study-

ing material and doing hands-on assignments, after

which students completed the course by doing the ﬁ-

nal exam. Students can use PDF material (with course

text and tasks), printed books, and an essential on-

line learning environment during their learning pro-

cess. For this study, four speciﬁc features were de-

veloped: a learner dashboard that included feedback

messages, self-assessment tasks with situated feed-

back, and reading support. Students were involved

as key stakeholders in the design and implementation

of such pedagogical features.

An overview page with a learner dashboard col-

lected all learning resources, including reading mate-

rials and various tasks, organized in rows by course

units. The learner dashboard allowed students to

track their progress and gain an overview of the avail-

able learning resources at a glance. The dashboard

has a tiled layout with a predeﬁned set of compo-

nents that the user can add, move, and remove. The

source code is publicly available under the GPLv3 li-

cense: https://github.com/CATALPAresearch/format

serial3. Currently, ﬁve components have been devel-

oped: a progress chart, learning goals, task list, due

dates, and feedback. The feedback component is used

to deliver adaptive feedback to support SRL learning

strategies as described in section 2. Examples of the

adaptation rules implemented for the SRL feedback

are listed in the documentation of ARI. The feedback

in the learner dashboard is presented as a list, which

includes a short title and the HIF. Below each list

item, students can rate the feedback: (i) This feed-

back is helpful for me, (ii) I want to put this feedback

into practice, (iii) This does not apply to me, and (vi)

Not now, maybe later. User interactions with the feed-

back items like scrolling into the display area, hov-

ering with the mouse, clicking links, and ratings are

stored to be used by the reinforcement learning agent

as rewards (or losses).

Evaluation Methods. The adaptive system was ini-

tially only evaluated technically. For this purpose,

it was implemented in the Moodle course described

above and conﬁgured with educators, including about

30 adaptation rules. Participation in the study was

based on informed consent following the GDPR.

Non-participating students did not suffer any disad-

vantages. 144 students took part in the study.

The following indicators were collected for the

evaluation: (1) frequency of rule execution, (2) dis-

tribution of executed rules per student, (3) reaction

of students to the display of feedback indicated by

the rules, (4) necessary adjustment of the adaptation

rules due to too frequent or too infrequent execution,

(5) time required to execute the rules and (6) com-

plaints/reports from educators and students.

Preliminary Evaluation Results. All deﬁned rules

have been executed. Per week, active students have

been confronted with up to 5 feedbacks caused by the

adaptation rules. We received only a few upvotes and

downvotes from the students. So far, educators have

not adjusted any rules. The execution time to compute

the rules did not affect the loading time of the Moodle

pages. The response time appeared to be indepen-

dent of the number of rules and the amount of data

collected in the learner model. We received no com-

plaints from the students so far. Educators requested

support in the deﬁnition of rules.

5 DISCUSSION AND

CONCLUSIONS

This article presented a system for implementing an

adaptive learning environment in Moodle based on

adaptation rules. It described a nine-step process for

ﬁnding, developing, and checking these rules. Thus,

the overall research question about how to design

and implement a scalable software architecture sup-

porting adaptive learning in an LMS could be an-

swered. However, using a rule-based approach may

raise questions, as expert systems of this type are seen

as outdated. We will discuss these arguments in terms

of the presented approach.

A key critique of rule-based systems is their rigid-

ity, as they rely on predeﬁned rules and may strug-

gle with unforeseen scenarios. However, in ARI, the

rule set functions as an educator-designed model, in-

formed by learning analytics, data mining, and LLMs.

Like any model, it simpliﬁes reality, but unlike com-

plex AI models, it remains transparent and easily ad-

justable.

Rule-based systems can become complex and dif-

ﬁcult to maintain as rules accumulate and interact

(Tabebordbar and Beheshti, 2018). In ARI, rules are

modeled in natural language, offering clarity despite

their number. While no strict limit exists, the six di-

mensions and personas provide structure, helping ed-

ucators start with a manageable rule set and expand it

effectively.

Critics may argue that rule-based systems offer

only superﬁcial personalization. However, clear cri-

teria enhance transparency, while the learner model

allows for detailed condition combinations. Personal-

ized SRL feedback and LLM-generated texts further

enhance individualization.

CSEDU 2025 - 17th International Conference on Computer Supported Education

252

Standard rule-based systems don’t learn from

data, relying on predeﬁned rules. In contrast, ARI’s

learner model integrates the latest data, enables au-

tomatic analysis of Moodle logs, and adapts through

reinforcement learning.

Rule-based systems may offer generic feedback

without deep process analysis. While learning se-

quences aren’t yet mapped in rules, the learner model

tracks recent activity sequences, allowing conditions

based on contains and similar operators.

A common critique of rule-based systems is their

slower real-time adaptability due to static rules. Moo-

dle, often sluggish, requires frequent page reloads

(e.g., in quizzes). However, with the learner model

stored in the browser’s IndexedDB and accessible

across tabs, updates occur at least on each reload, en-

suring near real-time visibility. Yet, threshold values

remain unchanged without educator intervention, pro-

viding stability and predictability—both essential for

consistency in learning.

From an ethical standpoint, key challenges must

be addressed (e.g., (Prinsloo and Slade, 2017)). In

line with the SHEILA framework (Tsai et al., 2018)

and consultations with our ethics committee and data

protection ofﬁcers, we ensure: (1) voluntary partic-

ipation (opt-in with informed consent, opt-out pos-

sible), (2) automated decisions serve only as rec-

ommendations without restricting access, (3) trans-

parency through explicit adaptation rules and learner

model data, (4) continuous monitoring, and (5) edu-

cator involvement in activating and conﬁguring adap-

tation rules.

ACKNOWLEDGEMENTS

This research was supported by the Center for Center

of Advanced Technology for Assisted Learning and

Predictive Analytics (CATALPA) of the FernUniver-

sit

at in Hagen, Germany.

REFERENCES

Alario-Hoyos, C., Est

evez-Ayres, I., P

erez-Sanagust

ın, M.,

Leony, D., and Delgado-Kloos, C. (2015). MyLearn-

ingMentor: A Mobile App to Support Learners Par-

ticipating in MOOCs. Journal of Universal Computer

Science, 21:735–753.

Almohammadi, K., Hagras, H., and Alghazzawi, D. (2017).

A survey of artiﬁcial intelligence techniques em-

ployed for adaptive educational systems within E-

learning platforms. Journal of Artiﬁcial Intelligence

and Soft Computing Research, 7(1):47–64.

Apoki, U. C., Al-Chalabi, H. K. M., and Crisan, G. C.

(2020). From Digital Learning Resources to Adap-

tive Learning Objects: An Overview. In Simian, D.

and Stoica, L. F., editors, Modelling and Development

of Intelligent Systems, pages 18–32, Cham. Springer

International Publishing.

Bimba, A. T., Idris, N., Al-Hunaiyyan, A., Binti Mahmud,

R., and Liyana Bt Mohd Shuib, N. (2017). Adaptive

feedback in computer-based learning environments: a

review. Adaptive Behavior, 25(5):217–234.

Brusilovsky, P. (2007). Adaptive Navigation Support. In

The Adaptive Web: Methods and Strategies of Web

Personalization, pages 263–290.

Cohen, W. W. (1995). Fast Effective Rule Induction. In

Prieditis, A. and Russell, S., editors, Machine Learn-

ing Proceedings 1995, pages 115–123. Morgan Kauf-

mann, San Francisco (CA).

De Clercq, M., Galand, B., and Frenay, M. (2020). One

goal, different pathways: Capturing diversity in pro-

cesses leading to ﬁrst-year students’ achievement.

Learning and Individual Differences, 81:101908.

Duraes, D., Toala, R., Goncalves, F., and Novais, P. (2019).

Intelligent tutoring system to improve learning out-

comes. AI Communications, 32(3):161–174.

Fruhmann, K., Nussbaumer, A., and Albert, D. (2010).

A psycho-pedagogical framework for self-regulated

learning in a responsive open learning environment.

In eLearning Baltics 2010: Proceedings of the 3rd

International eLBa Science Conference, pages 1–11.

Fraunhofer Verlag. International eLBa Science Con-

ference ; Conference date: 01-07-2010 Through 02-

07-2010.

Gasevic, D., Jovanovic, J., Pata, K., Milikic, N., Holocher-

Ertl, T., Jeremic, Z., Ali, L., Giljanovic, A., and

Hatala, M. (2012). Self-regulated Workplace Learn-

ing: A Pedagogical Framework and Semantic Web-

based Environment. Educational Technology & Soci-

ety, 14:75–88.

Grubi

c, A., Stankov, S., and

Zitko, B. (2015). Adaptive

courseware: A literature review. Journal of Universal

Computer Science, 21(9):1168–1209.

Jivet, I., Scheffel, M., Drachsler, H., and Specht, M. (2017).

Awareness Is Not Enough: Pitfalls of Learning An-

alytics Dashboards in the Educational Practice. In

Lavou

E., Drachsler, H., Verbert, K., Broisin, J.,

and P

erez-Sanagust

ın, M., editors, Data Driven Ap-

proaches in Digital Education, pages 82–96, Cham.

Springer International Publishing.

Kabudi, T., Pappas, I., and Olsen, D. H. (2021). AI-enabled

adaptive learning systems: A systematic mapping of

the literature. Computers and Education: Artiﬁcial

Intelligence, 2:100017.

Kopeinik, S., Nussbaumer, A., Winter, L.-C., Albert, D.,

Dimache, A., and Roche, T. (2014). Combining Self-

Regulation and Competence-Based Guidance to Per-

sonalise the Learning Experience in Moodle. In 2014

IEEE 14th International Conference on Advanced

Learning Technologies, pages 62–64.

Letham, B., Rudin, C., McCormick, T. H., and Madigan,

D. (2015). Interpretable classiﬁers using rules and

Bayesian analysis: Building a better stroke prediction

Architecture for Gradually AI-Teamed Adaptation Rules in Learning Management Systems

253

model. The Annals of Applied Statistics, 9(3):1350–

1371.

Martin, F., Chen, Y., Moore, R. L., and Westine, C. D.

(2020). Systematic review of adaptive learning re-

search designs, context, strategies, and technologies

from 2009 to 2018. Educational Technology Research

and Development, 68(4):1903–1929.

Menze, D., Seidel, N., and Kasakowskij, R. (2022). Interac-

tion of reading and assessment behavior. In Henning,

P. A., Striewe, M., and W

olfel, M., editors, DELFI

2022 – Die 21. Fachtagung Bildungstechnologien der

Gesellschaft f

ur Informatik e.V., pages 27–38, Bonn.

Gesellschaft f

ur Informatik.

Molenaar, I., de Mooij, S., Azevedo, R., Bannert, M.,

arvel

a, S., and Ga

sevi

c, D. (2023). Measuring self-

regulated learning and the role of AI: Five years of

research using multimodal multichannel data. Com-

puters in Human Behavior, 139:107540.

Nakic, J., Granic, A., and Glavinic, V. (2015). Anatomy

of Student Models in Adaptive Learning Systems:

A Systematic Literature Review of Individual Differ-

ences from 2001 to 2013. Journal of Educational

Computing Research, 51(4):459–489.

Nielsen, L. (2014). Personas. In Soegaard, M. and Dam,

R. F., editors, The Encyclopedia of Human-Computer

Interaction. Interaction Design Foundation.

Normadhi, N. B., Shuib, L., Md Nasir, H. N., Bimba, A.,

Idris, N., and Balakrishnan, V. (2019). Identiﬁcation

of personal traits in adaptive learning environment:

Systematic literature review. Computers & Education,

130:168–190.

Nussbaumer, A., Dahn, I., Kroop, S., Mikroyannidis, A.,

and Albert, D. (2015). Supporting Self-Regulated

Learning, pages 17–48. Springer International Pub-

lishing, Cham.

Nussbaumer, A., Kravcik, M., Renzel, D., Klamma, R.,

Berthold, M., and Albert, D. (2014). A Framework

for Facilitating Self-Regulation in Responsive Open

Learning Environments.

Perez Alvarez, R., Jivet, I., Perez-sanagustin, M., Scheffel,

M., and Verbert, K. (2022). Tools Designed to Support

Self-Regulated Learning in Online Learning Environ-

ments: A Systematic Review. IEEE Transactions on

Learning Technologies, 15(4):508–522.

Prinsloo, P. and Slade, S. (2017). Ethics and Learning

Analytics: Charting the (Un)Charted. In Lang, C.,

Siemens, G., Wise, A., and Ga

sevi

c, D., editors,

Handbook of Learning Analytics, pages 49–57. SO-

LAR.

Radovi

c, S., Seidel, N., Menze, D., and Kasakowskij, R.

(2024). Investigating the effects of different levels of

students’ regulation support on learning process and

outcome: In search of the optimal level of support

for self-regulated learning. Computers & Education,

215(105041).

Renzel, D., Klamma, R., Kravcik, M., and Nussbaumer, A.

(2015). Tracing self-regulated learning in responsive

open learning environments. In Li, F. W., Klamma,

R., Laanpere, M., Zhang, J., Manj

on, B. F., and Lau,

R. W., editors, Advances in Web-Based Learning –

ICWL 2015, pages 155–164, Cham. Springer Interna-

tional Publishing.

Rose, S., Engel, D., Cramer, N., and Cowley, W. (2010).

Automatic Keyword Extraction from Individual Docu-

ments, pages 1–20.

Seidel, N., Karolyi, H., Burchart, M., and de Witt, C.

(2021). Approaching Adaptive Support for Self-

regulated Learning. In Guralnick, D., Auer, M. E., and

Poce, A., editors, Innovations in Learning and Tech-

nology for the Workplace and Higher Education. TLIC

2021. Lecture Notes in Networks and Systems, pages

409–424, Cham. Springer International Publishing.

Sghir, N., Adadi, A., and Lahmer, M. (2023). Recent ad-

vances in Predictive Learning Analytics: A decade

systematic review (2012–2022). Education and In-

formation Technologies, 28(7):8299–8333.

Shi, S., Xie, Y., Wang, Z., Ding, B., Li, Y., and Zhang, M.

(2022). Explainable Neural Rule Learning. In Pro-

ceedings of the ACM Web Conference 2022, WWW

’22, pages 3031–3041, New York, NY, USA. Associ-

ation for Computing Machinery.

Tabebordbar, A. and Beheshti, A. (2018). Adaptive Rule

Monitoring System. In 2018 IEEE/ACM 1st Interna-

tional Workshop on Software Engineering for Cogni-

tive Services (SE4COG). Proceedings, pages 45–51.

Tsai, Y.-S., Moreno-Marcos, P. M., Jivet, I., Scheffel, M.,

Tammets, K., Kollom, K., and Ga

sevi

c, D. (2018).

The sheila framework: Informing institutional strate-

gies and policy processes of learning analytics. Jour-

nal of Learning Analytics, 5(3):5–20.

Vandewaetere, M., Desmet, P., and Clarebout, G. (2011).

The contribution of learner characteristics in devel-

opment of computer-based adaptive learning environ-

ments. Computers in Human Behaviors, 27:118–130.

White, J., Fu, Q., Hays, S., Sandborn, M., Olea, C., Gilbert,

H., Elnashar, A., Spencer-Smith, J., and Schmidt,

D. C. (2023). A prompt pattern catalog to enhance

prompt engineering with chatgpt. arXiv:2302.11382.

Wisniewski, B., Zierer, K., and Hattie, J. (2020). The Power

of Feedback Revisited: A Meta-Analysis of Educa-

tional Feedback Research. Frontiers in Psychology,

10.

Wu, T. T., Lee, H. Y., Li, P. H., Huang, C. N., and Huang,

Y. M. (2024). Promoting Self-Regulation Progress

and Knowledge Construction in Blended Learning via

ChatGPT-Based Learning Aid. Journal of Educa-

tional Computing Research, 61(8):3–31.

Xie, H., Chu, H.-C., Hwang, G.-J., and Wang, C.-C.

(2019). Trends and development in technology-

enhanced adaptive/personalized learning: A system-

atic review of journal publications from 2007 to 2017.

Computers & Education, 140:103599.

Yau, J. (2009). A MOBILE CONTEXT-AWARE FRAME-

WORK FOR SUPPORTING SELF-REGULATED

LEARNERS. In International Conference on Cog-

nition and Exploratory Learning in Digital Age

(CELDA, pages 415–419. IADIS.

Yau, J. Y.-K. and Joy, M. (2008). A Self-Regulated Learn-

ing Approach: A Mobile Context-aware and Adap-

tive Learning Schedule (mCALS) Tool. International

Journal of Interactive Mobile Technologies (iJIM), 2.

CSEDU 2025 - 17th International Conference on Computer Supported Education

254