An adaptive learning model based on a machine learning approach

Abstract

There is a broad consensus that digital learning environments can facilitate personalized learning processes. For the past several years, adaptive learning has been promoted as a potential solution for transforming higher education. Adaptive learning in a digital environment is aimed at implementing intelligent tutoring systems, which provide students with personalized learning programs based on their own skill level. Computer algorithms are used to orchestrate the learning process and deliver customized learning resources to learners which meet their needs and expectation. Algorithms can also be developed to improve learners' retention, achieve higher course outcomes, provide a more precise measure of learning achievements, and stimulate learners' interest and motivation. This paper illustrates SALM (Smart Adaptive Learning Model), an adaptive learning model based on a machine learning approach. It results from an ongoing research carried out within the project DocTDLL whose aim is the application of Transformative Digital Learning to Ph.D. study programs. This model has been designed to support a transformative self-learning approach in a smart learning environment. It should enable students to access digital learning materials, providing them with guidance, hints, supportive tools, and suggestions according to their individual learning skills and styles.

Full text

An adaptive learning model based on a machine learning approach
Gilberto Marzano, Velta Lubkina
Rezekne Academy of Technologies, Latvia
Abstract
There is a broad consensus that digital learning environments can facilitate personalized learning
processes. For the past several years, adaptive learning has been promoted as a potential solution
for transforming higher education.
Adaptive learning in a digital environment is aimed at implementing intelligent tutoring systems,
which provide students with personalized learning programs based on their own skill level.
Computer algorithms are used to orchestrate the learning process and deliver customized
learning resources to learners which meet their needs and expectation. Algorithms can also be
developed to improve learners’ retention, achieve higher course outcomes, provide a more
precise measure of learning achievements, and stimulate learners’ interest and motivation.
This paper illustrates SALM (Smart Adaptive Learning Model), an adaptive learning model
based on a machine learning approach. It results from an ongoing research carried out within the
project DocTDLL whose aim is the application of Transformative Digital Learning to Ph.D.
study programs. This model has been designed to support a transformative self-learning approach
in a smart learning environment. It should enable students to access digital learning materials,
providing them with guidance, hints, supportive tools, and suggestions according to their
individual learning skills and styles.
Keywords: adaptive learning, intelligent tutoring systems, digital transformative learning, online
transformative learning, online self-learning, smart learning environments
Introduction
Nowadays, the labor market is evolving according to the advances of technology, and there is a
broad consensus that many workers will need help in retraining the competencies they already
have (Lent, 2018; Robertson, 2018). In the near future, the changes produced by the digital
revolution will require more steady efforts to be made in continuous learning (Collins &
Halverson, 2018).
In a rapidly changing society, higher education institutions should strive to support students,
providing them with the right skills and competencies to become lifelong learners (Joos &
Meijdam, 2019). Correspondingly, an increasing amount of research in the educational scope
nowadays focuses on technology that can be used to increase teaching-learning productivity and
efficiency. In this regard, various authors have emphasized the benefits that digital technology
can provide in education as well as the strategic importance that self-directed learning can have
in developing knowledge for professional and personal growth (Kaplan & Haenlein, 2016).
Undoubtedly, digital learning environments can extend the audience of learners and, at the same
time, facilitate personalized learning processes (Fullan, 2013).

For the past several years, adaptive learning has been promoted as a potential solution for
transforming higher education (Dziuban, Moskal, Parker, Campbell, Howlin, & Johnson, 2018).
Moreover, in the last few years, digital technology and the advent of the internet has opened new
dimensions for adaptive learning. Currently, many adaptive learning platforms are available that
promise the application of adaptive learning for personalizing the learning process, whilst
numerous investigations are being carried out on the use of adaptive learning in the scope of
higher education (Dziuban, Moskal, Cassisi & Fawcett, 2016).
In this vein, this paper introduces SALM (Smart Adaptive Learning Model), an adaptive learning
model that includes machine learning functions. It is the result of an ongoing research being
carried out within the project DocTDLL, funded by the Latvian Council of Science, whose aim is
the application of Transformative Digital Learning to Ph.D. study programs.
Transformative learning can represent a powerful approach to support continuous learning,
providing trainees with the opportunity to learn, confront, engage, reflect, and explore new
learning modalities (Taylor & Cranton, 2012).
Research objectives and methodology
The objective of the DocTDLL project (Implementation of Transformative Digital Learning in
Doctoral Program of Pedagogical Science in Latvia) is to investigate how Transformative Digital
Learning (TDL) can support the Ph.D. study programs in Pedagogy as well as scientific and
academic capacity building in higher education.
The expected results of the project are:
Adapting the Personal-Cultural Orientations Scale and Digital Competency Profile to the
Latvian situation and cultural environment.
Conducting surveys in order to identify the attitudes, needs, and expectations of students
and lecturers/professors.
Developing new modules for the doctoral study program in Pedagogy, focusing on digital
learning practices and solutions.
Our research centers on the latter result. It is aimed at defining an intelligent tutoring system that
will provide students with personalized learning functions. The core idea of the research is
designing a smart learning environment (Chen, Cheng, & Chew, 2016) that is based on an
adaptive learning approach. This should include incremental learning, continual feedback,
regular assessment, benchmarking, and dynamic generation of learning paths according to the
learners’ skills and behavior.
According to Huang, Yang, & Hu (2012), a smart learning environment is a technology-
supported learning environment that provides appropriate support (e.g., guidance, feedback,
hints, or tools) in the right places and at the right time. It should meet individual learners’ needs,
which might be determined via analyzing their learning behaviors, performance, and the online
and real-world contexts in which they are situated.
For this purpose, we carried out a literature analysis of smart learning environments, including
commercial solutions such as Knewton’s Alta (https://www.knewton.com ) and ALEKS
(https://www.aleks.com).

We used a two-stage procedure. We first searched on online databases (Scopus, TR Web of
Sciences, Wiley Online Library, IEEE Xplore Digital Library, and Google Scholar), selecting a
pool of candidate articles relevant to the research aim. From this, we then used the selected
literature to support a needs analysis, performed involving a panel of Latvian Ph.D. students and
lectures/professors, as well as to model our smart learning environment, identifying its various
components and levels.
In the following paragraphs, we illustrate the functional structure of SALM, the structure of
learning, and how machine learning can support the action of virtual facilitators.
SALM basic principles
The SALM basic design is inspired by the principles of adaptive learning (Truong, 2016),
namely delivering custom training through just-in-time feedback, pathways, and resources. The
primary aim of SALM is to quickly and accurately determine what a student knows and does not
know in a learning context. It should answer the following two key questions:
1. What does the learner know?
2. What should the learner experience next?
An algorithm should select “the right item at the right time” for students as they learn. In SALM,
the algorithms that identify learners’ needs is based on Bayesian inference, also taking into
account the rate at which learning occurs. However, the SALM adaptive smart learning
environment also includes algorithms aimed at improving learners’ retention, as well as
stimulating the learners’ interest and motivation (Trimmer, Higginson, Fawcett, McNamara, &
Houston, 2015).

Structured learning units and learners’ profiles are the information base on which algorithms
operate. Figure 1 shows the functional structure of SALM.
Figure 1. SALM functional structure
The main components of SALM are:
1. Management, that encompasses the functions to support:
interaction between learners and teachers;
interaction among learners;
building of user profiles;
building of learning units;
organization of customized learning path;
filtering of resources according to learner profiles;
evaluation of learning retention;
analysis of learning resources to create ontologies.
2. Software library, which includes programs and algorithms to:
perform intelligent searches on the web according to learner profiles;
make inferences to select learning resources according to user profiles;
calculate the learning rate;
evaluate the learner status.
3. Profile subsystem, which contains the user profiles as the results of the analysis of the
user activity.
4. Resources subsystem, which is divided into learning units and web-based learning
resources. The former includes lectures and didactic materials such as course notes,

slideware, study guides, self-assessment questionnaires, etc. The latter includes web
searches, such as links to didactic objects, websites, articles, audio/video objects, etc.
In our model, the structure of learning units plays a crucial role, since a great part of adaptive
learning performance depends on how learning units have been created.
In the next paragraph, the basic structure of learning units is illustrated.
SALM structured learning units
Leaning units are didactic events that have one or more well-defined learning objectives, and
which cannot be broken down without losing their didactic function.
In SALM, learning units represent the nodes of a graph on three levels. The first level is formed
by the learning units belonging to domains (topics) that are essential to achieve the knowledge
and competences related to a given training program. The second level is composed by those
learning units containing events that are supplementary to one or more of the first-level learning
units, e.g., definitions, theories, techniques, author bios, etc. The third level encompasses those
learning units that represent didactic resources such as articles, videos, supplementary exercises,
etc.; they can be linked to the superior level nodes or selected as the result of a learner search.
All learning units are structured as follows:
1. General data: author, date of creation, history of modifications.
2. Classification: level, reference to the topic/s (first level units refer to only one topic),
learning requirements (e.g., which learning units must be learned previously), learner
status (e.g., knowledge, achievements, etc.).
3. External links: links to web resources.
4. Description: summary of each learning unit in terms of a concept or a list of concepts,
expressed by keyword/s, and the level of knowledge required.
The content creator provides the above information manually.
The graph of the learning units is organized by the person responsible for the training, according
to an evaluation of the learning difficulties made by them on the basis of their experience.
Successively, a program will analyze the learners’ behavior and reorganize the graph according
to the individual learner status, proposing repetitions or supplementary learning activities,
including searches on the didactic resources or on the web.
Whilst learning units are prepared by trainers, we hypothesized the possibility that learners can
add new learning resources gathered from their explorations on the web during searches. These
resources are listed under a see also section of a particular learning unit, and are also shared with
other learners who have a similar profile. Moreover, both students and trainers can express their
feedback on the added resources or include them within a particular category (topic).
Machine learning in SALM
The modern view of machine learning refers to a more general principle of learning as a
multilayer composition that can be applied in software applications (Goodfellow, Bengio, &
Courville, 2016).

In this regard, SALM aims to utilize the users’ experience in order to create intelligent agents
(virtual facilitators) that support learners in a smart adaptive learning environment according to
learners’ profiles. Initially these profiles are created by the training creator, then they are
improved by analyzing the experience and actions of users (e.g., opening and mastering a
learning unit, reading an article, watching a training video, performing an exercise, etc.). Profiles
are used for generating personalized learning paths and prompting didactic resources to learners,
as well as for suggesting searches on the web.
An intelligent agent analyzes what type of resources a user prefers, and what kind of learning
units are assimilated faster.
Our idea is to track the visited learner units and resources, recording paths that can subsequently
be suggested to other learners who share a similar profile. The assumption is that an effective
learning path can be dynamically generated and suggested to learners who have a similar profile,
since it is built by processing and classifying their knowledge, learning rate, interests,
motivation, and other factors related to the learning process.
The correspondence between a learner with their learner profiles is obtained by evaluating the
learner status.
The learning status of a user is obtained by tracking their progress through a learning unit and the
resources related to a specific topic. The learning status is expressed in terms of a five-point
Likert scale (poor, fair, good, very good, excellent), and is interpreted by another intelligent
agent.
Learner profiles are used to organize the initial learning graphs, whilst the learning status is used
to set the navigation path on the graph according to the level of knowledge of a learner, as well
as for refining the learners’ profiles. It is expected that a more efficient learner will go through
fewer learning units and resources in order to achieve their learning goal, whilst a less effective
learner will need more support in order to finalize their knowledge acquisition.
Ontologies are automatically generated to organize the web-learning resources, as well as to
predict which part of knowledge should be learned by a user next, whilst another agent provides
personalized access to learning resources. A proactive algorithm provides help to users who are
experiencing difficulties in attaining a given learning goal. This is based on the evidence that a
learner needs to absorb certain notions or perform specific exercises in order to reinforce their
achievements.
The SALM model has been conceived to accumulate data from the learners’ activity and the
various information resources. Machine learning algorithms are used to analyze and determine
various dependencies between data during the learning process.
A Bayesian machine learning algorithm should enable the encoding of the learners’ profiles
(prior probabilities). If the learner status corresponds to the likelihood of a particular profile, and
there is other evidence, e.g., learners’ actions, supporting that probability, one can encode a prior
that pushes the algorithm to estimate that probability.

Indeed, the Bayes theorem describes the probability of an event, based on prior knowledge of
conditions that might be related to the event. It is expressed as:
where P(A) and P(B) are independent probabilities of event A (learner profile) and B (learner
status) respectively, and P(A/B) is the probability of observing A (learner profile), given the event
B (learner status) has happened.
In SALM, we try to infer the parameters (θ) of our model as:
P(θ / D) = [P (D/ θ) . P(θ)] /P(D)
where P(θ) is the prior belief; P(D/θ) is the likelihood of data D, given that θ is observed or true;
and P(D) is a normalizing constant.
Conclusion
Transformative learning can support the learning process in many disciplines, whilst
Transformative Digital Learning can support intelligent incremental learning, automatic
feedback, and assessment, as well as the availability of different paths for the attainment of a
knowledge.
Nevertheless, TDL deviates considerably from traditional learning formats, including face-to-
face, online, and blended learning. Furthermore, most of the popular digital platforms are
generally conceived as content distribution systems, with little concern for the interests and the
immediate reactions of singular learners in the virtual classroom (Gjermeni & Percinkova, 2018).
This paper has, instead, presented SALM, a preliminary model that combines adaptive learning
and DTL. It is the result of an analysis of research into smart learning environments (Huang,
Yang, & Hu, 2012) that can support easy, engaged, and productive learning.
The core concept of SALM is in realizing personalized adaptive learning functions that are
constructed with four main aspects, namely learner profiles, competency-based
accomplishments, personal learning, and a flexible smart learning environment. The possibility
to use Bayesian inference in order to implement machine learning algorithms has been explored,
taking into account current research results (Akhrif, Benfares, & Hmina, 2019; Daniel, 2016;
Kim & Ritter, 2019).
The next step of our research will be to implement the SALM structural design and to develop
the algorithms necessary to support the generation and refinement of the personalized learning
path.

References
Akhrif, O., Benfares, C., & Hmina, N. (2019, October). Collaborative learning services in the
smart university environment. In Proceedings of the 4th International Conference on Smart City
Applications. Article no. 42.
Chen, N.S., Cheng, I.L., & Chew, S.W. (2016). Evolution is not enough: Revolutionizing current
learning environments to smart learning environments. International Journal of Artificial
Intelligence in Education, 26(2), pp. 561-581.
Collins, A., & Halverson, R. (2018). Rethinking education in the age of technology: The digital
revolution and schooling in America. Teachers College Press.
Daniel, B.K. (2016). Big data and learning analytics in higher education. Springer International
Publishing, Switzerland.
Dziuban, C.D., Moskal, P.D., Cassisi, J., & Fawcett, A. (2016). Adaptive Learning in
Psychology: Wayfinding in the Digital Age. Online Learning, 20(3), pp. 74-96.
Dziuban, C., Moskal, P., Parker, L., Campbell, M., Howlin, C., & Johnson, C. (2018). Adaptive
Learning: A Stabilizing Influence across Disciplines and Universities. Online Learning, 22(3),
pp. 7-39.
Fullan, M. (2013). Stratosphere: Integrating technology, pedagogy, and change knowledge.
Pearson Canada.
Gjermeni, F., & Percinkova, B. (2018). Combining Intelligent Algorithms and E-Learning Styles
to Create an Improved Intelligent System in Evaluating an E-Learning Student’s Profile.
ANGLISTICUM. Journal of the Association-Institute for English Language and American
Studies, 7(2), pp. 11-21.
Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning. MIT press.
Huang, R., Yang, J., & Hu, Y. (2012). From digital to smart: The evolution and trends of learning
environment. Open Education Research, 1(1), pp. 75-84.
Hwang, Gwo-Jen. “Definition, framework and research issues of smart learning environments-a
context-aware ubiquitous learning perspective”. Smart Learning Environments 1, no. 1 (2014): 4.
Available at: https://slejournal.springeropen.com/articles/10.1186/s40561-014-0004-5, last
accessed on 01.03.2020.
Kaplan, A.M., & Haenlein, M. (2016). Higher education and the digital revolution: About
MOOCs, SPOCs, social media, and the Cookie Monster. Business Horizons, 59(4), pp. 441-450.
Kim, J.W., & Ritter, F.E. (2019, July). Consideration of a Bayesian Hierarchical Model for
Assessment and Adaptive Instructions. In International Conference on Human-Computer
Interaction, pp. 521-531.
Joos, P.P.M., & Meijdam, L. (2019). Innovation in higher education: needed and feasible. Tilburg
Series in Academic Education, 2. Available at:

https://pure.uvt.nl/ws/portalfiles/portal/30073297/UVT0144_ESSAY_NR2_2018_EN_V2_def.p
df; last accessed on 12.15.2019.
Robertson, J. (2018, March 6). Digital Revolution Demands Changes to Education, in
Forbes.com. Available at: https://www.forbes.com/sites/gradsoflife/2018/03/06/digital-
revolution-demands-changes-to-education/#1cfdf53e522a
; last accessed on 12.15.2019.
Lent, R.W. (2018). Future of work in the digital world: Preparing for instability and
opportunity. The Career Development Quarterly, 66(3), pp. 205-219.
Trimmer, P.C., Higginson, A.D., Fawcett, T.W., McNamara, J.M., & Houston, A.I. (2015).
Adaptive learning can result in a failure to profit from good conditions: implications for
understanding depression. Evolution, medicine, and public health, 2015(1), pp. 123-135.
Truong, H.M. (2016). Integrating learning styles and adaptive e-learning system: Current
developments, problems and opportunities. Computers in human behavior, 55, pp. 1185-1193.