The Use of Deep Learning in Open Learning: A Systematic Review (2019 to 2023)

bstract No records of systematic reviews focused on deep learning in open learning have been found, although there has been some focus on other areas of machine learning. Through a systematic review, this study aimed to determine the trends, applied computational techniques, and areas of educational use of deep learning in open learning. The PRISMA protocol was used, and the Web of Science Core Collection (2019– 2023) was searched. VOSviewer was used for networking and clustering, and in-depth analysis was employed to answer the research questions. Among the main results, it is worth noting that the scientific literature has focused on the following areas: (a) predicting student dropout, (b) automatic grading of short answers, and (c) recommending MOOC courses. It was concluded that pedagogical challenges have included the effective personalization of content for different learning styles and the need to address possible inherent biases in the datasets (e.g., socio-demographics, traces, competencies, learning objectives) used for training. Regarding deep learning, we observed an increase in the use of pre-trained models, the development of more efficient architectures, and the growing use of interpretability techniques. Technological challenges related to the use of large datasets, intensive computation, interpretability, knowledge transfer, ethics and bias, security, and cost of implementation were also evident.


The Use of Deep Learning in Open Learning: A Systematic Review (2019 to 2023)
Conceptions about open learning are vast and wide.In a general sense, it refers to educational modalities to broaden education and training by breaking the barriers of time and space.Open learning has provided both students and teachers with greater flexibility, professional development, productivity, culture and socialisation in learning communities (Tzeng et al., 2022;Uddin et al., 2021).This pedagogical conception has been complemented by distributed learning in which teachers, students, and the content to be taught and learned are not centralised but can occur at any time and place (Zakaria et al., 2022;Zheng et al., 2022).In both open and distributed learning, the teacher is a facilitator of learning, the learner assumes an increasingly active and interactive role, and teaching and learning are reinforced and mediated by the use of digital technologies promoting open, online, and ubiquitous distributed learning environments (Mardini et al., 2023;Zheng et al., 2022).
Distance education has gained momentum with the expansion of open educational offerings and online vocational training.This has led to the enrolment of large numbers of students and has reinforced and integrated the use of information and communication technology (ICT).Therefore, the generation of new models and patterns of teaching and learning has been closely linked to new ICT trends creating the inter-, multi-, and trans-disciplinary space of emerging technologies (Mrhar et al., 2021).
Over the last 10 years, common open learning tools have included open educational resources (OER), collaborative teaching platforms, virtual learning environments, and educational social networks.With the evolution of open and distributed learning, distributed learning ecosystems (DLE) have emerged.
These ecosystems utilize distributed learning infrastructures that bring together various tools and technologies related to OER, services, resources, and open learning environments (Otto et al., 2023).DLEs have promoted the effectiveness of open and online learning, although highly dependent on platforms, the Internet, interaction and interactivity, as well as teacher and learner empowerment.
Significant progress has been made in the design and development of DLEs; however, there is still a latent lack of theoretical and empirical analysis of how emerging technologies such as AI, virtual reality, and augmented reality have influenced open learning (Otto et al., 2023).
AI has been grounded in various disciplines, such as natural language processing (NLP), artificial neural networks, computer vision, robotics, knowledge engineering, and machine learning (ML), among others (Hassan et al., 2019).The development of DLE has enabled the use of artificial intelligence in education (AIEd), notably, as expressed by Chen, Feng, et al. (2020), in adapting content, designing virtual tutors, automated assessment, data analysis, the use of virtual and augmented reality, creating recommender systems, and developing specific skills.These have all sought to improve the accessibility, personalisation, and efficiency of learning (Alruwais, 2023;Goel & Goyal, 2020).
AIEd can be learner-centred (e.g., adaptive or personalised learning management systems), teachercentred (e.g., automating tasks such as administration, assessment, learning progress, and detecting plagiarism) or system-centred, providing administrators with decision-making information related to monitoring dropout patterns Chen, Feng, et al. (2020).
ML is one of the most widely used disciplines of AI.It enables, among other applications, the design of intelligent tutoring systems and performance prediction.However, in education, unstructured data such as images, text, and voice have often been handled (An et al., 2019).Conventional ML models may not be as effective in extracting useful features from these types of data, and therefore, it has been necessary to use more powerful models such as deep learning (DL).DL, a subset of ML, refers to the use of deep neural networks, configured with multiple successive layers of neurons (LeCun et al., 2015).It has represented the most advanced machine learning technique for solving problems with large sets of structured training data (Chassagnon et al., 2019) such as the analysis of traces and data from massive open online courses (MOOCs) for predicting school performance.While DL can address some of the limitations of ML, it is not a universal solution and also has its considerations, such as the need for large amounts of training data and computational resources.The choice between ML and DL depends on the specific nature of the problem and the data available in the educational context (El-Rashidy et al., 2023).
Recent studies on AI in open learning have focused on providing a learning experience for each learner by influencing motivation and online participation (Salas-Rueda, 2023).In a general sense, these systems must ensure the ability to provide feedback and structure adaptive learning content according to the individual capabilities of each learner.The usefulness of these tools depend on the design and development of more efficient intelligent tutoring systems, as shown in the teaching of mathematics, languages, and programming (Liang et al., 2023;Su et al., 2023).
In AI, DL as a subset of ML is based on the use of artificial neural networks (ANN), whose typology can be convolutional, recurrent, generative adversarial, deep, or modular neural networks.This area of AI has made inroads in education, mainly in analysing learning interactions in MOOCs, determining the chronological sequence of each student's interactions, predicting academic dropouts, and designing new and more efficient learning courses based on user experience, learning habits, and interactions (Verma et al., 2023).The most widely used technological models have been long short-term memory (LSTM) algorithms, sequential interaction rule mining process, and temporal interaction analysis (Yu et al., 2021).
Another recent application of DL has been automated online discussion message categorization based on convolutional neural network (CNN) and random forest classifiers.This advancement allowed for the analysis of interactions in online and open learning contexts.Utilizing the community of inquiry (CoI) framework, this application of DL has delineated the dimensions of cognitive presence (e.g., knowledge (re)construction, problem-solving), social presence (e.g., social interactions), and teaching presence (e.g., course design, interaction, interactivity) (Hu et al., 2021).
To enhance learning outcomes, DL has facilitated the creation of adaptive e-learning systems that analyze the behavior of individual learners in their interactions with learning objectives.Additionally, deep autoencoders have been utilized to learn and predict learner behaviors.DL has also supported the development of video analysis classification systems aimed at generating engaging video learning reports (Verma et al., 2023).
While there have been numerous studies on DL applications in education, few have focused on open learning.Consequently, there has been a lack of systematic reviews analyzing these applications.This research addresses this gap by examining the scientific literature.

Gaps in the Analysis of Studies on DL in Open Learning
Table 1 shows that some reviews, mappings, and bibliometric studies on AIEd have been published in Scopus and the Web of Science (WoS).Only three of these explicitly included DL studies, (Chen, Xie, et al. (2020); Pan et al., 2023;Vanitha & Jayashree, 2023).The remainder focused on other branches of AI.Several points are worth highlighting regarding these studies.Crompton and Burke (2023) discussed how, in higher education, AIEd contributed to learning assessment and prediction from AI assistants, intelligent tutoring systems, and learning management.Along these lines, others have focused on (a) algorithms and systems for learning prediction and student retention (Alhothali et al., 2022;Shafiq et al., 2022); (b) MOOC recommender systems (Uddin et al., 2021); and (c) analysing the impact of deep learning on educational time series datasets (Vanitha & Jayashree, 2023).Su et al. (2023) focused on artificial intelligence digital literacy and AIEd challenges in the context of K-12 to higher education, while Koong Lin et al. ( 2023) focused on the unique use of ChatGPT in education.
In their systematic review, Liang et al. (2023) analysed research methods, and the role of AI in language teaching and its learning outcomes.Similarly, other mapping studies and reviews have explained the applications of engineering and computational techniques at certain levels of education, such as higher education (Hwang et al., 2021) and early childhood education (Su et al., 2023).
Although several authors (Chen, Xie, et al., 2020;Pan et al., 2023;Talan, 2021) conducted comprehensive analyses of more than 900 articles, they focused on scholarly output and its bibliometric analysis only, without delving into the advantages of AI in open learning.
In their extensive systematic review of influential AIEd studies, Chen, Xie, et al. (2020) stated that only two studies explicitly identified the use of DL, referring to the study of advanced neural network architecture and the achievements of optimising study strategies from parallel robot instruction.In their study, they concluded that this was understandable, as DL was a newer area of research compared to general AI and machine learning.
In general, the reviews argued that these AIEd studies focused on profiling and dropout prediction, content evaluation, adaptive system design, and intelligent tutoring systems; there is still a lack of studies that systematised the use of DL.This theoretical research reaffirmed the importance of analysing empirical studies on the unique use of DL and its relationship with open learning, as MOOC dropout prediction and course recommendation require the use of powerful computational models (Wang et al., 2024).
To fill this gap, this study analysed AIEd-related scientific articles focusing on DL and open learning published between 2019 and 2023 to explore the important questions that remain to be investigated.

Objectives and Research Questions
This study aimed to identify, through a systematic review, the trends, applied computational techniques, and areas of educational use of deep learning in open learning.For this purpose, we examined the scientific literature from 2019 to 2023, present in one of the main bibliographic reference database collections, namely the Web of Science (WoS).
Aligned with this objective and addressing theoretical gaps, we sought to answer the following questions:

Method
We utilized the updated PRISMA statement guidelines to search for and select scientific information (Page et al., 2021).A quantitative procedure was employed for coding to ensure the validity of the study (Zawacki-Ritcher et al., 2020).

Search Strategy and Criteria
Only articles from peer-reviewed journals were selected to ensure a high level of quality.The search parameters were narrowed to the period from 2019 to 2023 to ensure the currency of the literature, which is essential in the AIEd area.Mendeley was used to extract articles and eliminate duplicates.
The electronic search protocol included the WoS databases, namely the Social Sciences Citation Index The inclusion and exclusion criteria are detailed in Table 2.

Screening and Validation Strategy
After carrying out the search of papers, two independent researchers used a standard checklist form to exclude irrelevant articles and determine their eligibility.In the process, any discrepancies between reviewers were resolved.
Subsequently, bibliographic data were extracted, and key findings and results were synthesised and recorded.The PRISMA diagram and its checklist for determining study quality were used to assess and carry out the selection process (Page et al., 2021).Finally, the two independent researchers read the selected articles to extract relevant information and answer the research questions.Any inconsistencies between the two researchers' results were resolved by a third reviewer.The process revealed a few articles on the use of DL in open learning (Figure 1 Figure 1 Selecting Studies (PRISMA Protocol)

Coding and Visualisation Tools
The selected studies were coded for deductive aspects and internal validity with a set of three criteria: keywords, authors, and first authors' countries.Concerning inductive coding for conclusion validity, the focus was on the influence of DL on open learning to identify trends.Finally, grounded coding for external validity focused the findings on how DL was used rather than how it could be used.VOSviewer version 1.6.19 was used for bibliometric data visualisation and analysis.

Bibliometric Output Information
In the literature analysis, 23 articles were finally selected (Figure 2) of which two were highly cited (Onan, 2021;Xing & Du, 2019).The analysis showed a trend of three to six manuscripts published annually.The low scientific output regarding the use of DL in open learning was highlighted.A co-authorship network of the 147 authors was created.Of these, collaboration was shown in 83 (Figure 4).The most cited names were those with the highest presence, represented by 23 clusters.

Figure 4
Co-Authorship Network Of the 83 authors, only seven showed strong collaboration (Figure 5).

Co-Authorship Network with Consistent Connections
A total of 81 author keywords were identified.After debugging and standardising common keywords and abbreviations, these were reduced to 68 (Figure 6).These aspects are further explored in the discussion of the pedagogical dimension.

Pedagogical Dimension (Question 2)
Each study was analyzed in-depth, and the results are presented in Table 3.The main findings focused on the analysis and prediction of learner behaviour, interaction, and interactivity in MOOCs, although some specific cases evaluated other open learning systems (Hamal & El Faddouli, 2022;Mardini et al., 2023).

Technological Dimension (Question 3)
The DL techniques and algorithms used, and their level of accuracy, are shown in Table 4.As can be seen, the most commonly used DL techniques or algorithms were related to the classical use or modern variants of BERT, LSTM, DNN, and NLP.Their levels of accuracy are adequate to measure the dependent variables (refer to Table 3).

Discussion and Conclusions
This study aimed to identify, through a systematic review, the trends, applied computational techniques, learning (Pan et al., 2023;Vanitha & Jayashree, 2023).
For the second and third research questions, dealing with pedagogical and technological dimensions, respectively, we identified trends in the use of DL in open learning in several key directions.
As an interesting note, we agreed with Goel and Goyal (2020) that AI studies in general have identified that there is a possible relationship between personal or friendship relationships and the likelihood of dropout-in other words, virtually all likely friends among all enrolled students showed the same behaviour in both video interaction and dropout likelihood.
A relationship has been established between the course content, the characteristics of open educational resources, their complexity, and student cognitive fatigue (Jiang, 2022;Zakaria et al., 2022).In addition, analysing these factors as well as student interactivity and interaction, through logs and traces, has established the probability of weekly dropout (Yin et al., 2020).Similarly, analysis of interaction in forums (An et al., 2019;El-Rashidy et al., 2023) and quizzes (Xing & Du, 2019) in a MOOC course can help predict the probability of weekly dropout.
Analysis of students' cognitive presence in sustained discourse in a virtual community (e.g., integration, problem-solving, and intuition) is a high predictor of academic performance in MOOCs (Liu et al., 2023a).
Sentiment analysis or opinion mining of individual student responses allows for some assessment and prediction of retention or dropout in a MOOC course.In this regard, research has been conducted on mass assessments (Onan, 2021) and forum interactions (Chen, Feng, et al., 2020;Mrhar et al., 2021).
There has been some hesitation from students regarding mass assessments related to quality and veracity, although the computational results provided good reliability rates (Onan, 2021).A hybrid procedure between automated and teacher-led assessments was suggested in several reviews (Mrhar et al., 2021).
Furthermore, the analysis of interaction and interactivity in activities posted in forums within MOOCs showed that there was a correlation between learner sentiment and engagement, and the likelihood of dropping out.As the number of students dropping out of the MOOC course decreased, the feelings and motivation towards the course increased (Chen, Feng, et al., 2020;Mrhar et al., 2021).
In this area (MOOC dropout), the results can be grouped into the following clusters: • Cluster 1: Dropout prediction using artificial neural networks, association rules mining, data analytics, ML, and personalisation.
• Cluster 3: The identification of cognitive presence through the community of inquiry model, online discussions, pre-trained language model, and text analysis.
• Cluster 4: Students' dropout through deep-neural networks and the deepfm model.
• Safety: DL models can be vulnerable to adversarial attacks, where carefully designed inputs can mislead the model.Ensuring the robustness of the model is a constant challenge.
• Cost of implementation: Developing and implementing DL solutions can be costly in terms of human resources, hardware, and time.This can limit its adoption in resource-constrained contexts.
Despite these challenges, research advances have continued to address these concerns and improve the applicability of DL in open learning environments.In this educational domain, various DL algorithms have been employed, such as convolutional neural networks (CNNs) for processing visual data, recurrent neural networks (RNNs) for temporal sequences, and transformers for natural language processing (NLP) tasks (Hamal & El Faddouli, 2022).Techniques have included transfer learning, data augmentation, and personalized optimization.In terms of trends, there has been an increase in the application of pre-trained models, the development of more efficient architectures, and the growing use of interpretability techniques.In this sense, the integration of Artificial Intelligence (AI) in the personalization of the learning experience is an emerging trend reiterated in the Horizon Reports (EDUCAUSE, 2023).

Conclusion
This study aimed to contribute to the educational community by identifying the pedagogical potential of Deep Learning (DL) in open learning, such as: • Content recommendation and automatic grading of short answers with feedback as key tools for meaningful personalized learning, achieved through constructive alignment between objectives, activities, and learning assessment.
• Predicting student dropout based on levels of student interactivity and interaction with digital educational resources, and the quality of responses to self-assessment as well as formative and summative assessment activities.This would be linked to an adequate multidirectional synchronous and asynchronous pedagogical communication and interaction process, providing support and tutoring services to motivate the student to learn in an autonomous, personalized, and collaborative way.
The results obtained in the application of DL in open learning have influenced the efficiency of higher education administrations, early counseling, and mentoring, as well as the design and implementation of educational interventions.However, there is agreement on the need to delve deeper into the ethical and moral issues of artificial intelligence (AI) concerning cultural differences, inclusion, and student emotions, as well as the pedagogical use of AI by teachers (Mouta et al., 2023).
The studies reviewed demonstrated their contribution to education; however, there was a lack of research to follow up on these results by answering questions such as: Have Massive Open Online Courses (MOOCs) been redesigned based on the results of DL application in open learning?(See Table 3); and Is the dropout rate maintained?
Learning as an educational, cultural, and psychosocial process depends on a variety of cognitive, motivational, affective, communicative, sociological, pedagogical, didactic, and technological factors.In terms of technology, DL and AI in education have brought us closer to identifying some criteria for approaching success in open learning.The algorithms and methods used have offered a high cognitivist weight of pedagogical value but could be enriched with other pedagogical foundations.It is interesting that the description of the DL methods used has hardly described the pedagogical basis, which to some extent obscures educational assessment.

Limitations of the Study
This study was limited to consulting only the Web of Science (WoS) Core Collection, specifically SCIE and SSCI.Therefore, it is possible that interesting results published in journals indexed in Scopus or other databases were omitted.Additionally, only articles in English were analyzed, overlooking articles published in other languages that could have enriched the results obtained in this research.

Future Lines of Research
There has been a lack of studies that have analyzed and compared the results obtained in the use of DL and machine learning (ML) in open learning, which would help to make decisions and consequently define the most efficient techniques and algorithms.In this sense, a systematic review and meta-analysis are recommended.

From
an educational and open learning point of view, DLEs have focused on the diversity and interactions of actors and (re)use activities, allowing the creation of solutions such as resource aggregation mechanisms and open learning repositories.In this context, open pedagogy, as well as software advances based on artificial intelligence (AI) and the standardisation of OER metadata have all developed.
How does scientific collaboration relate to the study of DL in open learning, highlighting its application areas?(This question was addressed through the bibliometric dimension, specifically co-authorship networks, keyword networks, and main clusters.)What are the dependent variables studied and their main findings?(This was answered through the pedagogical dimension by analyzing the content of each study.)What are the DL techniques or algorithms used in open learning (independent variable), and what is their level of accuracy?(This was addressed through the technological dimension, identifying, for each study, DL techniques or algorithms, data sources used, and levels of accuracy.)

(
SSCI) and the Science Citation Index Expanded (SCIE), housing more than 12,000 journals.A full-text search was conducted in line with the research objectives and questions.The Boolean search included terms related to DL, open learning, and distributed learning.Since open education, distance education, online education, and distributed learning are related but distinct, they were included as keywords to verify later whether each result was related to open education.Similarly, DL and ML are different, but as some authors mention them interchangeably (Chen, Xie, et al. (2020) they were used as keywords, and subsequently, only works referring to DL were included in the analysis.The techniques used were checked to identify whether they referred to ML (e.g., supervised learning, unsupervised learning, and reinforcement learning) or DL (e.g., convolutional, recurrent, generative adversarial, deep, or modular artificial neural networks).The initial Boolean search string focused on "deep learning OR DL OR machine learning OR ML," which yielded generic AIEd results.Subsequently, these were filtered according to the search string "open learning OR OL OR distributed learning OR DL" to obtain documents related to the research topic.In summary, the final search string focused on "(deep learning OR machine learning) AND (open education OR distance education OR online education OR distributed learning) AND (educational technologies OR artificial intelligence in education)."

Figure 6
Figure 6 areas of educational use of deep learning in open learning.To this end, the WoS Core Collection, including both SCIE and SSCI, was searched.The PRISMA protocol guided the selection of 23 articles, reflecting a low academic production related to the research objective, which indicates that deep learning (DL) in education is relatively recent.Concerning the first research question, the application of DL has mainly focused on predicting student dropout, automatic grading of short answers, and recommending MOOC courses.Technological underpinnings were based on (a) the flow of student clicks on videos, (b) student interaction and interactivity, (c) the quality of responses in interactive activities such as forums, and (d) the length of time spent on activities.The low academic research output reflects that there is still insufficient scientific collaboration in this research area, which may be a consequence of how recent the use of DL is in open

Table 2
Inclusion and Exclusion Criteria

Table 4
Techniques, Level of Accuracy, and Data Sources