Techniques and applications of Machine Learning and Artificial Intelligence in education: a systematic review

Técnicas y aplicaciones del Machine Learning e Inteligencia Artificial en educación: una revisión sistemática

Wiston Forero-Corba

Universitat de les Illes Balears, UIB, España

Francisca Negre Bennasar

Universitat de les Illes Balears, UIB, España

Esta obra está bajo una Licencia Creative Commons Atribución-NoComercial 4.0 Internacional.

Machine Learning (ML) is a branch of artificial intelligence (AI) that has seen an exponential increase in recent years. The scientific community is paying increasing attention to educational tools enriched with smart technology, as they have the potential to revolutionize teaching-learning processes.

At present, ML research applied to education in areas such as teacher perception (Salas Rueda et al., 2022), student perception (Demir & Güraksın, 2022), academic performance (Ahajjam et al., 2022), school dropout (Alvarado Uribe et al., 2022) and computational thinking (Almeida Pereira Abar et al., 2021), among others, show in their results, the implication of the use of intelligent techniques in the solution of complex problems in the education sector.

Different types of research have been compiled in systematic reviews on AI (Zawacki-Richter et al., 2019; Zhai et al., 2021; Salas-Pilco & Yang, 2022; Su et al., 2022) and systematic reviews on ML (Sasmita & Mulyanti, 2020; Luan & Tsai, 2021; Mittal et al., 2022). Reviews on AI have mainly focused on the university sector, with the exception of Su et al. (2022) which studies the primary school and high school levels. ML reviews have identified common keywords in research, such as prediction, identification, performance, and recommendation, and have described the type of intelligent algorithms or techniques used. Although these systematic reviews were conducted during or after the pandemic, only the study by Mittal et al. (2022) addressed COVID-19.

In education, the difference between ML and AI is not always clear, even though both fields focus on applying the concept of prediction. ML is focused on systems learning from data (Luan & Tsai, 2021), while AI allows systems to perform tasks autonomously (Zhai et al., 2021). However, our systematic review departs from analyzing studies of both AI and ML applied to the education sector for the following reasons: AI and ML aim to create systems that can execute tasks that are normally considered human-like, both fields use mathematical and statistical techniques to analyze and process data, they have great potential to revolutionize the way we interact with the world, and finally, the period from 2021 to February 2023 has experienced an exponential growth in research related to this topic.

In recent years, ML has provided different techniques or algorithms to predict situations according to large amounts of information that, through good data processing and filtering, can generate very effective predictions. Different authors have developed ML algorithms to help educators (Duzhin & Gustafsson, 2018; Yu et al., 2022). This has allowed these intelligent techniques to be applied to the education sector and to help combat the dynamic problems that afflict all types of contexts.

AI in schools offers multiple possibilities for school administrators, teachers, and students. One example is ChatGPT, the latest version, GPT-4, is integrated into software such as Microsoft Office, Edge, and Bing, optimizing educational tasks. AI and ML have been oriented towards educational tasks (Zafari et al., 2021), which highlights the need to strengthen Teachers’ Digital Competence (TDC).

Continually, research in the education sector seeks to close educational gaps, and ML and AI emerge as an alternative means to achieve optimal results. A study of robotics with intelligent techniques aims to close the gap between educational and professional robotics by introducing ML techniques where differences in access, trajectory, progress and educational outcomes are best for students (Dietz et al., 2022). In addition to research in education, technological advancement is an important factor for the education gap. Technological development has opened the gap to challenges in understanding the use, application, and inner workings of technologies, especially emerging technologies such as AI and ML (Temitayo et al., 2022). This indicates its importance as an emerging technology based on its correct use and application for the benefit of quality and dignified education.

The current curricula are constantly updated and with that in mind, curriculum development, which must provide answers to the demands imposed by the knowledge society, must include topics and activities based on ML and AI at all school levels, allowing to dynamize the teaching-learning processes. However, the complexity and dynamics of AI teaching highlight the need for a detailed examination of the curriculum development process in a given context (Dai et al., 2022), showing the relevance of curriculum assessment in all instructional areas and how to approach them according to the context.

Educational processes along with these intelligent techniques and tools applied in and out of the classroom have led to their implementation being treated with restraint due to the ethical considerations involved (Bogina et al., 2022). So much so, that teachers need to be trained and updated to cope with the teaching processes, improving their competences in communication, research, pedagogy, technology, and management, among others. As referred to by UNESCO (2019) in the Beijing Council on AI and education, education sectors must address the integration of the TDC on AI in ICT competency frameworks, to support the teachers training in educational environments with a strong presence of AI.

The inclusion of ML in education has made digital transformation of great benefit to all educational actors, making the education system more convenient for both teachers and students (Nafea, 2018). However, it would also be of great benefit to school administrators and families, who are an important reference point in any educational community and are closely involved in the benefits that these new technologies can generate.

The training of teachers in AI and ML is a challenge for educational institutions. For digital transformation in the classroom to become a reality, teachers must be prepared to adapt technology to their teaching practices (Almeida Pereira Abar et al., 2021), which requires solid knowledge in these areas. Lack of such knowledge limits the optimal implementation of AI and ML technologies in education. As such, school administrators need to take on the challenge of leading the training of the TDC.

The aim of this research is to identify opportunities for improving teaching-learning processes and educational management at all levels of the educational context through the application of machine learning and artificial intelligence.

On this basis, this paper answers the following research questions (RQ):

1. RQ1: What levels of education have ML or AI studies been conducted in education?

2. RQ2: In which countries has ML or AI research in Education been conducted and which country has the most influence?

3. RQ3: What are the key issues and the most frequent words used in the studies?

4. RQ4: What ML techniques have been used in research?

5. RQ5: What were the results of implementing ML or AI as an emerging technology in education?

The methodology considered appropriate for ascertaining the current status of all types of research is the systematic review (Marín, 2022), following the PRISMA 2020 protocol (Yepes-Nuñez et al., 2021). The search equation (Table 1) was applied to obtain the studies in the Web of Science (WoS) and Scopus databases. From the inclusion and exclusion criteria for the filtering and narrowing of studies applied (Table 2), a group of 55 articles could be systematically obtained (Table 3).

Table 1 shows the search equation according to subject, educational approach, context, and level. For the document search in both databases, this equation is applied to the title, abstract and keywords. In WoS, "TS" is applied to the equivalent formula (title, abstract, and keywords) and in Scopus, the equivalent of "TITLE-ABS- KEY". The design of the search terms as well as the inclusion and exclusion criteria (Figure 1) are based on the recommendations by Zawacki-Richter et al. (2020), for systematic reviews focused on educational research, as well as the indications from Marín (2022) for educational technology research.

The search formula was as follows:

Table 1
Search equation

Source: own elaboration.

The inclusion and exclusion criteria are as follows:

Table 2
Inclusion and exclusion criteria

Source: own elaboration.

Considering Table 2, the studies were taken between 2021 and 2023 to reflect the latest advances in scientific knowledge. This research was done during and after the pandemic. In previous systematic reviews (Sasmita & Mulyanti, 2020; Su et al., 2022), the selection of studies was limited to the English language. This is because most high-impact journals publish their articles in English, which is why we selected studies in English for our review. This allowed us to obtain studies relevant to our research. Databases are limited to WoS and Scopus as they are valued as the two most relevant bibliometric tools, being considered the two leading databases of academic articles in the world ranking (Zhu & Liu, 2020), allowing the identification of quality studies. To identify the latest research in the area, follow trends and research relevance, the WoS Core Collection database was used.

Figure 1 shows the entire procedure with all inclusion and exclusion criteria.

Figure 1
PRISMA Flowchart of the study
Source: own elaboration.

The two researchers were involved in the screening, jointly reviewing the studies up to the results. For the systematic review, the Rayyan tool was used, which allowed coding data on the year of publication, journal name, countries of authorship, sample, methodology and results. The socialization of the sample, methodology and results of each study was necessary to unify criteria and guarantee the quality of the research.

The documentary analysis was carried out using descriptive statistics and systematic content analysis. Orange Data Mining 3.35.0 software was used to perform the geographical location of the studies, the word cloud was used to analyze the top 20 most frequent words in the selected full papers. In addition, VOSviewer 1.6.19 was used for the network map, Microsoft Excel for the statistical graphs and app.diagrams.net for the classification of ML techniques.

The results of the 55 articles below were drawn from 45 high-impact journals, as shown in Figure 2. The number of journals analyzed is an indicator that the study was comprehensive, covering a wide range of perspectives, trends, and patterns.

Figure 2
Journals vs Number of studies/journal
Source: own elaboration.

The journals with the highest number of studies in the review were Applied Sciences and Education and Information Technologies with 3 articles each. The significance of having 45 different journals out of 55 in the review increases the likelihood that a wider range of studies will be included and therefore be more representative of the available evidence.

Table 3 presents the studies selected in this review, specifying the title, central subject, context of application, country, or countries in which the research was implemented, whether it covered the COVID-19 topic, educational level, or levels at which the study was applied and the year of publication. (P: Primary), 2. (P, S: Primary, Secondary), 3. (S: Secondary), 4. (S, U: Secondary, University) 5. (U: University).

Table 3
Selected studies

Note: Primary school, S: Secondary school, U: University, ASD: Autism Spectrum Disorder, FATE: fairness, accountability, transparency and ethics, STEM: Science, Technology, Engineering and Mathematics

To answer the first research question, based on Table 3, Figure 3 shows the level of education applied in the studies.

RQ1: What levels of education have ML or AI studies been conducted in education?

Figure 3
Educational level applied in studies
Source: own elaboration.

To answer the second research question, it is noted that studies in English often do not reflect the diversity of global research. Therefore, the choice was made to select research papers in English and to analyze how non-English speaking countries can base their studies in English to have a wider research reach. Figure 4 shows the geographical location (countries) in which the research was conducted and/or applied.

RQ2: In which countries has ML or AI research in Education been conducted and which country has the most influence?

Figure 4
Geographical location of the studies
Source: own elaboration.

Figure 4 shows that the United States (USA) has the largest number of studies. Therefore, Figure 5 estimates the states with the highest research influence in the articles.

Figure 5
Influence of articles on USA
Source: own elaboration.

To answer the third research question, Figure 6 shows a network map depicting the relationships between the key subjects of the selected studies, and Figure 7 shows a word cloud highlighting the 90 most frequent and relevant words in these studies.

RQ3: What are the key issues and the most frequent words used in the studies?

Figure 6 shows a network map representing the key subjects from the titles and abstracts of the 55 research papers. The network map shows four subclusters of interrelated key subjects, identified by colors: green for machine learning (ML), yellow for artificial intelligence (AI), blue for education and red for prediction.

Figure 6
Network map of the 55 research projects on ML and AI in Education
Source: own elaboration.

The green ML sub-cluster is connected to the yellow sub-cluster representing the AI theme because it is a key piece of technology for creating intelligent tools from data recognition and learning. On the other hand, the red sub-cluster representing the key subject of prediction is connected to the ML sub-cluster because ML techniques or algorithms are based on prediction for decision making. Finally, both the ML and AI sub-cluster are connected to the blue sub-cluster representing the education subject, because it has the potential to improve the teaching-learning process in several ways, such as focusing on improving teacher skills, predicting, and identifying students' strengths and weaknesses to estimate their academic progress, supporting fields such as educational robotics and augmented reality, among others.

Figure 7
Word cloud representing the most frequent words
Source: own elaboration.

To quantitatively establish the ranking of the word cloud, Table 4 shows the 20 most frequent words in our word cloud. According to the ranking, the words "students", "learning" and "ai" (abbreviation for artificial intelligence) are the three most frequent words, indicating that the selected studies have a high ratio of application of smart tools in education.

Table 4
Most frequent words in studies

Source: own elaboration.

RQ4: What ML techniques have been used in research?

Figures 7 and 8 provide information in response to the fourth question.

ML techniques are classified according to the type of learning:

• Supervised learning: Learning from labelled data (Segura et al., 2022).

• Unsupervised learning: Learning from unlabelled data (Taha et al., 2018).

• Semi-supervised learning: Learning from labelled and unlabelled data (Chrysafiadi et al., 2022).

• Reinforcement learning: Learning from interactions with their environment (Dietz et al., 2022).

Figure 8 classifies the techniques according to the type of learning: supervised, unsupervised and reinforcement learning. The names of the techniques in English and their initials are maintained in relation to other scientific papers.

Figure 8
ML techniques found in ML studies (Red: Supervised learning, Blue: Unsupervised learning, Red and Blue: Semi-supervised learning, Grey: Reinforcement learning)
Source: own elaboration.

Figure 9 represents the frequency of supervised learning techniques found in the studies. The most commonly used techniques in the studies are as follows Random Forest (RF), Decision Tree (DT) and K-nearest neighbors (KNN), being the least usedBoruta Algoritm, Causal Forest (CF), Convolution Neural Networks (CNN),Back Propagation Network (BPN), Logistic Model Trees (LMT),Penalized Multinomial Regression (PMR), Graph Network Block (GNB),Multilayer Logistic regression (MLR), Gaussian Process Regression (GPR),Least Square Regression (LSR), Leasts Absolute Shrinkage and Selection Operator (LASSO), Artificial Neural Network (ANN), Stacking Emsemble Learning, Multilayer feed-fodward neural network (MFFNN), and Multilayer Linear Regression (MLR).

Figure 9
Supervised learning techniques in studies
Source: own elaboration.

RQ5: What were the results of implementing ML or AI as an emerging technology in education?

In response to the fifth research question, Table 5 can be found in the main section of the Annex. This table presents the research study, the sample, methodology and results. The order is established according to the order criteria in Table 3.

According to the results in Table 5, the opportunities for improvement in teaching-learning processes and educational management can be grouped into the following categories: prediction of academic performance and school dropout, analysis of student and teacher perception, development of virtual robotics, learning on generative models, implementation of AI and ML, insertion of computational thinking at all levels, strengthening the legal framework in education, efficiency of school management, social robotics intervention, computer security training, incorporation of AI in clinical education, STEM for forensic analysis and AI support in students with special educational needs (SEN), among others. These enhancement opportunities can help improve student academic performance, reduce dropout rates, strengthen educational equity, and improve the overall quality of education.

The studies highlight predictions at the institutional level; however, classroom-level predictions are also recommended because they are more accurate and are based on more specific data on individual students. Nevertheless, institution-level predictions can provide a more general view of academic performance since they are based on institution-wide data, such as grade point averages, attendance rates, graduation rates, dropout rates, etc.

The methodologies used in the studies were defined at two levels: research and teaching. At the research level, the aim was to find new knowledge and test hypotheses using quantitative, qualitative, or mixed methods. At the teaching level, the aim is to strengthen the TDC necessary for personal and professional development for student learning.

This systematic literature review analyzed 55 references on the use of ML and AI in education conducted in 38 countries, with the United States leading the way, from primary school through university levels. The results show that the 33 intelligent techniques extracted from the studies can be applied in the education sector to:

• Detect students' academic performance early.

• Improve the educational skills of teachers.

• Facilitate the learning of students with autism spectrum disorders (ASD).

• Predict school dropout and make decisions about it.

• Improve and generate educational content.

• Close educational gaps.

• Implement AI teaching at all educational levels.

• Strengthen the information security of the educational community.

• Motivate learning through mobile devices.

• Strengthen the field of robotics.

• Improve academic and career guidance for students.

• Prevent the spread of fake news on social networks.

• Understand and reflect on the relationship between humans and machines.

• Develop critical thinking based on computational thinking.

The distribution of studies on the application of intelligent techniques in education is analyzed. The studies analyzed focused on the use of AI and ML techniques. The results show that the application of intelligent techniques in education is gaining ground at all educational levels. In the past, most of this research focused on the university sector (Forero & Negre, 2022), but 74.6% of the analyzed studies were applied at the primary school and secondary school level. Our review is more comprehensive than other systematic reviews, as it analyzes studies at all primary, secondary and university levels.

Table 3 shows that 20% of the selected studies addressed COVID-19 in some way. This significant increase compared to other systematic reviews is since the studies were conducted between 2021 and February 2023, when much of this research was still ongoing during the pandemic. The COVID-19 pandemic has been a major global event that has had a significant impact on all aspects of life. Consequently, it is not surprising that many scientific studies have focused their attention on this issue. From our review it can be inferred that one in five studies focused on the COVID-19 disease and its consequences.

In recent years, there has been an increase in the publication of research from non- English-speaking countries in high-impact English-language scientific journals. This is due to the growing importance of science and technology in the globalized world, increased investment in research and development in emerging countries, and the need to share knowledge and results with a wider scope. In Latin America, Brazil, Mexico, Chile, and Ecuador are the countries that have experienced the greatest growth because they have a strong science and technology base and are increasingly investing in research. In Europe, Spain, the Netherlands, Portugal, Italy, Greece, Switzerland, Lithuania, Finland, Slovenia, Russia, and Turkey are some of the countries leading the growth as they have a long scientific tradition and are committed to international research. In Africa, Benin, Cape Verde, Angola, Morocco, Nigeria, Ghana, Tanzania, Kenya, South Africa, and Namibia have experienced significant growth because they are increasingly investing in research to address the development challenges they face. In Asia, Japan, China, Saudi Arabia, Iran, Vietnam, South Korea, Qatar, India, Israel, and Malaysia are always stepping up as they have strong economies and are increasingly investing in research to drive economic growth. This increase is a positive trend that is contributing to the globalization of science.

Figure 5 shows that California, Massachusetts, and Texas are the states with the highest concentration of ML and AI research in education. This is because institutions such as the University of Southern California, the Massachusetts Institute of Technology and the University of North Texas are putting a lot of effort into this field. The authors of this research are mainly engineers, which highlights the need to involve education professionals in the research process.

As can be seen in Figure 8, this study found 33 different ML techniques, which are classified into the four main categories of learning: supervised (28), semi-supervised (1), unsupervised (3) and reinforcement learning (1), some techniques are subgroups of others, such as Artificial Neural Network ANN (Zafari et al., 2021) and Neural Network NN (Oskotsky et al., 2022), but are not grouped together as a single technique, to respect the full name that appears in the research. This indicates that experts are increasingly convinced that ML techniques are appropriate and very important for educational research as they are recognizing the potential to improve educational understanding and practice through new models and methods of teaching-learning.

In line with the above, institutions can use smart techniques and tools to help their students. Grade prediction is a high-impact tool that can considerably benefit both students and institutions (Gerlache et al., 2022), for example, they can provide students with insight into their current performance and potential for success, helping students to identify areas or subjects in which they need to improve and to take steps to improve their results. In addition, grade predictors can help students make decisions about their future careers, knowing how they will do in a particular career, students can make more robust decisions about what to study and where they want to work in the future.

The most frequent applications using ML techniques focus on the prediction of academic performance, in particular, the Random Forest algorithm is the most frequently used in these investigations, which is a supervised learning technique with high prediction probability. An example of the effectiveness of Random Forest for academic performance prediction is a study by Houngue et al. (2022) where it achieved 99% prediction accuracy. However, where there are other techniques with good probability, the choice will depend on the type of data available and the specific objective of the study. In general, ML algorithms have achieved a higher predictive level compared to classical models (Costa Mendes et al., 2021), this is because ML algorithms can learn complex patterns in the data, which allows them to generate more accurate predictions.

The research sample handles a large amount of information, since, for ML techniques to be effective in their predictive capacity, it is necessary for the data to be correctly labelled. Therefore, Big Data comes to play an important role, where the role of the data must maintain those aspects of ethical and moral integrity regarding the information of both the participants (Blease et al., 2021) as well as the curriculum (Eguchi et al., 2021). Both studies agree that the data used for ML must be ethical and moral, as biases in the data can negatively affect the accuracy of the models.

ML studies are developing new techniques that can improve the prediction system in the education sector. For example, the study by Suzuki et al. (2022) used an ML model to predict the academic performance of primary school students in Japan with an error of 10%. The study by Tarik et al. (2021) used an ML model to predict class attendance of university students in Malaysia with an error of 5%. Thus, being able to effectively predict educational management by minimizing errors at the technical level and at the institutional level enables problem solving in dynamic educational contexts.

The TDC is indispensable for teachers, as it assesses their skills in the knowledge and use of digital technologies. Therefore, identifying these opportunities for improvement in teaching-learning processes helps to mainstream ML and AI concepts in all subject areas and levels of knowledge. The teachers from different subject areas and with different levels of computer science training may have different conceptions of how to integrate ML concepts in schools (Temitayo et al., 2022). Therefore, this paper also seeks to raise awareness about the importance of teachers, regardless of their background, having the necessary skills and competences to apply ML and AI in the classroom. The integration of smart technologies is a crucial educational innovation across all subject areas and educational levels, as it has the potential to bridge the digital and school divide that has become a challenge for education experts.

This review, due to its current subject matter, broad scope and to limit the proposed methodology, only studies in English have been analyzed. However, it is possible that there is research in ML and AI applied to education in other languages that have not been considered and that could be of interest.

Although the WoS and Scopus databases have been used to narrow down the study, the research could be extended by consulting other databases, as they could yield interesting results on ML and AI.

It is necessary to strengthen education systems in Latin America, Africa, and Oceania with the implementation of AI and ML experiences and research, improving the provision of human and physical resources and quality teacher training, especially in the TDC.

The equation used allows the implicit integration of certain studies with the keyword “K-12” (Ali, DiPaola, Lee, Sindato et al., 2021; An et al., 2022; Duncan et al., 2022; Eguchi et al., 2021; Sanusi et al., 2022). However, it does not include references to the keywords “high* education*”, “ungraduate*”, “vocational training*”, “vocational education”, “adult education” or “corporate training*”. This could be of interest for future work, as these keywords could broaden the scope of the research.

The curricula of subjects must incorporate the concepts of new intelligent technologies in a cross-cutting manner. To this end, it is necessary for educational and institutional management to strengthen the competences of teachers and students in these new educational fields.

Despite the scarcity of research related to diversity, learners with special educational needs, disability and illness, there is a need to deepen and strengthen these fields to close gaps and have a positive impact on education and society.

Finally, it is hoped that this research will contribute to the knowledge and understanding of educational practices with ML and AI and how these can be implemented to strengthen teaching-learning and educational management processes in all types of contexts.

How to cite: Forero-Corba, W., & Negre Bennasar, F. (2024). Techniques and applications of Machine Learning and Artificial Intelligence in education: a systematic review. [Técnicas y aplicaciones del Machine Learning e Inteligencia Artificial en educación: una revisión sistemática]. RIED-Revista Iberoamericana de Educación a Distancia, 27(1). https://doi.org/10.5944/ried.27.1.37491