A. Scope of the Survey
A.调查范围

This article revolves around the key applications of data-driven AI techniques in education and describes the most commonly used data-driven AI techniques in different application areas of education. It also describes the tools and platforms developed in the market as outcomes of the research work achieved in different applications including: 1) student grading and evaluations and retention and drops out prediction; 2) personalized learning; 3) sentiment analysis; 4) recommendation systems in education; and 5) classrooms' monitoring and visual analysis. We also analyze research trends in AI applications in education by providing a detailed bibliometric analysis of the domain. This article also advises on the current limitations, pitfalls, and future directions of research in the domain, and how it can fill the current gaps.
本文围绕数据驱动的人工智能技术在教育领域的主要应用，介绍了在教育的不同应用领域中最常用的数据驱动的人工智能技术。文章还介绍了市场上开发的工具和平台，这些工具和平台是研究工作在不同应用领域取得的成果，包括1) 学生评分和评价以及保留率和辍学率预测；2) 个性化学习；3) 情感分析；4) 教育推荐系统；5) 课堂监控和可视化分析。我们还通过提供该领域的详细文献计量分析，分析了人工智能在教育领域应用的研究趋势。本文还就该领域目前存在的局限、误区和未来的研究方向，以及如何填补目前的空白提出了建议。

B. Related Surveys B.相关调查

Literature reports several interesting articles analyzing different aspects of AI applications in education. There are also some surveys targeting different aspects and application areas of AI in education [13]. However, to the best of our knowledge, there is no recent detailed survey covering the domain from different perspectives. In previous works, Romero et al. [10] provided a general review of existing literature to analyze how EDM and learning analytics (LA) have been applied to educational data. Romero et al. [14] surveyed EDM literature for a decade (i.e., 1995 to 2005). Baker et al. [15] reported a detailed survey of data mining techniques used in the education sector. Charitopoulos et al. [16] provided a detailed overview of literature from 2010 to 2018 aiming at the use of soft computing methods, such as decision trees, random forests (RFs), artificial neural networks, fuzzy logic, support vector machine (SVMs), and genetic/evolutionary algorithms, in EDM and LA.
文献报道了几篇有趣的文章，分析了人工智能在教育中应用的不同方面。还有一些调查针对人工智能在教育中的不同方面和应用领域[13]。不过，据我们所知，最近还没有从不同角度对该领域进行详细调查。在以前的作品中，Romero 等人[10]对现有文献进行了总体回顾，分析了 EDM 和学习分析（LA）如何应用于教育数据。Romero 等人[14]调查了十年来（即 1995 年至 2005 年）的 EDM 文献。Baker 等人[15]详细调查了教育领域使用的数据挖掘技术。Charitopoulos等人[16]对2010年至2018年的文献进行了详细综述，旨在探讨决策树、随机森林（RFs）、人工神经网络、模糊逻辑、支持向量机（SVMs）和遗传/进化算法等软计算方法在教育数据挖掘和洛杉矶大学中的应用。

More recently, Fischer et al. [17] surveyed the existing data mining techniques in education with a particular focus on highlighting challenges in mining Big Data. Mduma et al. [18] focused on students' retention and dropout prediction techniques. Almasri et al. [19] provide a detailed survey of intelligent tutoring systems (ITS), another attractive application of AI in education, proposed from 2000 to 2018. Mousavinsab et al. [20] also provided a systematic review of literature on ITS with a particular focus on its key characteristics, applications, and evaluation methods. In another interesting survey on ITS, Malekzadeh et al. [21] provided a detailed overview of literature aiming at the development of different emotion regulation strategies to better engage learners in ITS. Tahiru [22] provided a systematic review of AI in certain applications of AI in education, such as ITS, automation of administrative tasks, and development of smart content. Al-Emran et al. [23] surveyed the Internet of Things (IoT)-based educational solutions. Matcha et al. [24] provided a systematic review of empirical studies on LA from a self-regulated learning perspective. Hooshyar et al. [25] also provided a systematic review of literature on self-regulated learning, however, they focused on the role of open learner models instead. Zaidi et al. [26], on the other hand, provided a detailed overview of online learning and the AI education market focusing on the scope, processes, and providers/competitors in the market. The authors also discussed the role and future potential of AI in online education. In another interesting relevant survey, Kim et al. [27] provided a detailed overview of pedagogical agents by analyzing their roles, current progress, and future potential in research-based pedagogical agent design.
最近，Fischer 等人[17] 对教育领域现有的数据挖掘技术进行了调查，重点强调了大数据挖掘所面临的挑战。Mduma 等人[18] 重点研究了学生保留率和辍学预测技术。Almasri 等人[19]详细调查了 2000 年至 2018 年期间提出的智能辅导系统（ITS），这是人工智能在教育领域的另一项有吸引力的应用。Mousavinsab 等人[20] 也对有关智能辅导系统的文献进行了系统综述，尤其关注其主要特点、应用和评估方法。在另一项关于 ITS 的有趣调查中，Malekzadeh 等人[21] 详细综述了旨在开发不同情绪调节策略的文献，以便让学习者更好地参与 ITS。Tahiru[22]系统回顾了人工智能在教育领域的某些应用，如智能学习系统、行政任务自动化和智能内容开发。Al-Emran 等人[23] 调查了基于物联网（IoT）的教育解决方案。Matcha 等人[24] 从自我调节学习的角度系统回顾了有关洛杉矶的实证研究。Hooshyar等人[25]也对有关自我调节学习的文献进行了系统综述，但他们关注的重点是开放式学习模式的作用。另一方面，Zaidi 等人[26] 详细概述了在线学习和人工智能教育市场，重点关注市场的范围、流程和提供商/竞争者。作者还讨论了人工智能在在线教育中的作用和未来潜力。在另一项有趣的相关调查中，Kim 等人[27]通过分析教学代理的作用、当前进展以及基于研究的教学代理设计的未来潜力，详细概述了教学代理。

Literature also provides several interesting surveys on recommender systems for education [28]. For instance, Khanal et al. [29] provided a systematic review of literature on ML-based recommendation systems in e-learning. Similarly, Rahayu et al. [30] provided a systematic review of literature on learning path recommender systems by covering key topics, such as trends, ontology, recommendation processes and techniques, contributing factors, and evaluations.
文献还提供了一些关于教育推荐系统的有趣调查[28]。例如，Khanal 等人[29] 对电子学习中基于 ML 的推荐系统的文献进行了系统回顾。同样，Rahayu 等人[30]对学习路径推荐系统的文献进行了系统综述，涵盖了趋势、本体、推荐流程和技术、促成因素和评估等关键主题。

In contrast to the existing surveys, this article provides a broader picture of the domain by covering most of the key application areas of AI in education, such as student's grading and evaluations, students' retention and dropout prediction, students' performance, sentiment analysis, recommendation systems, classrooms' monitoring, and ITS. The article also highlights the key market players, tools, and platforms along with key challenges, potential market opportunities, future research directions, and pitfalls of AI in education. More importantly, we analyze the trends of the research in the domain from a different perspective by providing a detailed bibliometric analysis of the domain. The article also discusses the different types of AI algorithms including supervised, unsupervised, reinforcement learning (RL), and generative AI algorithms. Table I provides a comparison of this survey against the existing surveys on the topic.
与现有调查相比，本文涵盖了人工智能在教育领域的大多数关键应用领域，如学生评分和评价、学生保留率和辍学预测、学生表现、情感分析、推荐系统、教室监控和智能教育系统，从而为该领域提供了更广阔的视野。文章还重点介绍了人工智能在教育领域的主要市场参与者、工具和平台，以及主要挑战、潜在市场机遇、未来研究方向和陷阱。更重要的是，我们通过对该领域进行详细的文献计量分析，从不同角度分析了该领域的研究趋势。文章还讨论了不同类型的人工智能算法，包括有监督、无监督、强化学习（RL）和生成式人工智能算法。表 I 提供了本调查与现有相关调查的比较。

TABLE I Comparison of Our Article Against Existing Surveys
表 I 我们的文章与现有调查的比较

C. Contributions C.捐款

1. A detailed overview of existing literature in nine different application domains in which AI is deployed for education.
   对人工智能用于教育的九个不同应用领域的现有文献进行详细概述。

2. Describe and highlight recent works on the most commonly used ML algorithms adopted in literature over the years for these applications.
   描述并重点介绍近年来在这些应用中最常用的 ML 算法。

3. Explore and identify the future scope and market opportunities for AI researchers and developers in the education sector.
   探索并确定人工智能研究人员和开发人员在教育领域的未来发展空间和市场机遇。

4. Analyze the publication trends of research literature taking into account a total of 4447 articles published in various top subject venues through a detailed bibliometric analysis in terms of research productivity by authors, institutions, and country, and knowledge flow across various research venues.
   通过详细的文献计量分析，从作者、机构和国家的研究生产力以及不同研究领域的知识流动等方面，分析在各种顶级学科刊物上发表的共计 4447 篇文章的研究文献发表趋势。

5. Identify the research and development companies and corporations working in the domain along with the tools and platforms available for both educational institutions and researchers.
   确定该领域的研发公司和企业，以及可供教育机构和研究人员使用的工具和平台。

6. Highlight the limitations, pitfalls, and open research challenges in using data-driven AI in education.
   强调在教育领域使用数据驱动型人工智能的局限性、陷阱和公开研究挑战。

The rest of this article is organized as follows. Section II describes some key application domains of AI in education and different tools developed as part of the efforts. Section III details some key AI techniques employed in different applications of education. Section IV provides a comprehensive bibliometric analysis of existing literature. Section V provides basic insights into the domain based on our analysis of existing literature, and lists the key limitations and pitfalls of AI in education along with some potential directions of future research and open issues in the domain. Finally, Section VI concludes this article.
本文接下来的内容安排如下。第二节介绍了人工智能在教育领域的一些关键应用领域，以及在此过程中开发的不同工具。第三节详细介绍了在不同教育应用中采用的一些关键人工智能技术。第四节对现有文献进行了全面的文献计量分析。第五节根据我们对现有文献的分析，提供了对该领域的基本见解，并列出了教育领域人工智能的主要局限性和陷阱，以及该领域未来研究的一些潜在方向和未决问题。最后，第六节对本文进行了总结。

A. Students' Evaluation and Performance, Retention, and Dropout Prediction
A.学生的评价和表现、保留率和辍学率预测

To be able to predict a student's likely future performance, retention, and dropout can provide very powerful platforms that facilitate educational interventions and remedial actions promptly. The development of AI models for the prediction of student performance and uncovering hidden insights and patterns are some of the most salient applications and research areas in EDM and learning analytics. Several studies have been conducted in the area of academic performance analysis and prediction, including by Adejo et al. [37], who conducted an empirical investigation and comparison of several data sources, classifiers, and ensembles of classification techniques to predict the academic performance of university students. In detail, they compared and analyzed the performance of ensemble techniques combining information from different data sources against the models trained on data from a single source. To this aim, several algorithms including DT, artificial neural networks (ANNs), and SVM were used and compared individually as well as in ensemble (combination) modes. Their findings support the premise that multiple data sources in combination with heterogeneous ensemble ML techniques provide efficient models for predicting student performance and also for identifying students at risk of attrition. Livieris et al. [38] also proposed an ensemble-based semisupervised approach for predicting student performance achieving sufficient accuracy in early prediction of student progress.
能够预测学生未来可能的表现、保留率和辍学率，可以提供非常强大的平台，便于及时采取教育干预和补救措施。开发用于预测学生成绩的人工智能模型，挖掘隐藏的洞察力和模式，是 EDM 和学习分析领域最突出的应用和研究领域。在学业成绩分析和预测领域已经开展了多项研究，其中包括 Adejo 等人[37]的研究，他们对若干数据源、分类器和分类技术组合进行了实证调查和比较，以预测大学生的学业成绩。具体而言，他们比较和分析了结合不同数据源信息的集合技术与根据单一数据源数据训练的模型的性能。为此，他们使用了包括 DT、人工神经网络（ANN）和 SVM 在内的几种算法，并对其单独和集合（组合）模式进行了比较。他们的研究结果支持这样一个前提，即多种数据源与异构集合 ML 技术相结合，可为预测学生成绩和识别有流失风险的学生提供高效模型。Livieris 等人[38] 也提出了一种基于集合的半监督方法来预测学生成绩，在早期预测学生进步方面达到了足够的准确性。

Deep learning techniques were also employed to tackle the challenging problem of forecasting the future performance of students. For instance, Kim et al. [39], proposed GritNet, a novel deep learning model, for the prediction of students' performance by treating it as a sequential prediction task. GritNet is mainly based on bidirectional long short-term memory (BLSTM). The authors applied the model to a group of Udacity students to predict their performance and were able to show favorable results over the logistic regression models with on-the-ground improvements in the early weeks of the course which are traditionally the most challenging to predict.
深度学习技术也被用于解决预测学生未来成绩这一具有挑战性的问题。例如，Kim 等人[39]提出了一种新颖的深度学习模型 GritNet，将学生成绩预测视为一项连续的预测任务。GritNet 主要基于双向长短期记忆（BLSTM）。作者将该模型应用于一组 Udacity 学生，对他们的成绩进行预测，结果表明，与逻辑回归模型相比，该模型在传统上最难预测的课程早期几周取得了良好的实际效果。

Student retention and dropout, which is a universal factor affecting both online and offline learning platforms, could be linked to student performance prediction. AI techniques have also been proven very effective in student retention and dropout prediction. As an initial effort, Aulck et al. [40] modeled student dropout based on a dataset of 32 500 demographics and transcript records at a large public institution. They conclude that early potential dropouts can be detected even with single-term transcripts thus opening the door for AI applications to predict and prevent some of the causes of dropouts.
学生保留率和辍学率是影响在线和离线学习平台的普遍因素，可与学生成绩预测联系起来。人工智能技术在学生保留率和辍学率预测方面也被证明非常有效。作为初步尝试，Aulck 等人[40] 基于一个大型公立院校 32 500 个人口统计数据和成绩单记录的数据集，对学生辍学情况进行了建模。他们得出的结论是，即使只有一个学期的成绩单，也能及早发现潜在的辍学现象，从而为人工智能应用预测和预防某些辍学原因打开了大门。

In literature, AI modeling techniques have been applied to predict dropout rates to calculate dropout probability as well as identify the contextual, demographic, and individual factors related to learning activities such that education administrators can design effective intervention and prevention remedies. For example, Solis et al. [41] considered several factors including the year of graduation, residence while attending the lectures, gender, grants/financial aid, program, program choice, and previous academic record. The authors also evaluated the performance of various ML algorithms for the prediction of student retention rates at university levels; using comparative experiments and analysis between the effectiveness of NNs, random forest, logistic regression, and SVM, the authors found that the combination of RF using ten randomly selected candidates per division was the optimum combination for predicting student dropouts. On validation of this work, their model was able to predict dropouts with over 90% accuracy combined with a sensitivity rate of 87%. The authors concluded that for better prediction complete information from all semesters must be used in training the algorithms. The authors also recommended the inclusion of other factors, such as the students' interest in the program, their level of interest in the university, earnings/work while studying, as well as the educational level of their relatives, family support, and their attitudes and personalities.
在文献中，人工智能建模技术已被用于预测辍学率，以计算辍学概率，并识别与学习活动相关的环境、人口和个人因素，从而使教育管理者能够设计有效的干预和预防补救措施。例如，索利斯等人[41]考虑了多个因素，包括毕业年份、上课时的居住地、性别、助学金/财政援助、课程、课程选择和以往的学业成绩。作者还评估了各种 ML 算法在预测大学学生保留率方面的性能；通过对 NN、随机森林、逻辑回归和 SVM 的有效性进行对比实验和分析，作者发现 RF 组合（每个分部随机选择 10 名候选人）是预测学生辍学率的最佳组合。在对这项工作进行验证时，他们的模型能够以超过 90% 的准确率和 87% 的灵敏度预测辍学率。作者总结说，为了更好地预测，在训练算法时必须使用所有学期的完整信息。作者还建议加入其他因素，如学生对课程的兴趣、对大学的兴趣程度、学习期间的收入/工作、亲属的教育水平、家庭支持以及学生的态度和性格。

From a different perspective, Pilkington et al. [42] conducted a qualitative study as part of funded research at a U.K. university with a sample of 75 researchers, tutors, and professors. They used a combination of “systematic, sequential, explanatory, and thematic” approaches to focus on findings from thematic analysis. They identified engagement, attendance, workload, family pressure, and mental health as factors that continue to contribute to dropout issues regardless of university engagement efforts. Apart from these factors, the sense of community, institutional social-environmental contribution, and academic integration are other critical factors contributing to students' retention and dropout [43]. It is therefore essential to go beyond basic AI modeling for predicting dropout and analyze the impact of ambient socioeconomic, psychological, demographic, and family factors to be able to conduct a determined analysis of the causes of dropout. For example, generally, the students do not have a fixed study approach in massive open online courses (MOOCs). Thus, tackling the temporal and diverse nature of MOOC data and considering the inconsistent learning/study activities of students, such as watching and rewatching a video, in dropout prediction is a very difficult task as these reasons are diverse and highly personalized. Chen et al. [44] applied visualization analytics methods and techniques (DropoutSeer) to analyze large datasets from MOOC systems to correlate ML predicted dropout rates with the learning activities of MOOC subscribers visually. The aim was to enable content designers to design more suitable engaging content and AI experts to design better predictive models [45]. This was shown to be more effective, for instance, than the process of feature identification as a critical step in the model-building process [46].
Pilkington 等人[42]从另一个角度进行了一项定性研究，作为英国一所大学受资助研究的一部分，样本包括 75 名研究人员、导师和教授。他们结合使用了 "系统性、顺序性、解释性和主题性 "方法，重点关注主题分析的结果。他们发现，参与、出勤率、工作量、家庭压力和心理健康是无论大学如何努力都会继续导致辍学问题的因素。除这些因素外，社区意识、机构的社会环境贡献和学术融合也是导致学生留校和辍学的其他关键因素[43]。因此，必须超越预测辍学的基本人工智能模型，分析周围社会经济、心理、人口和家庭因素的影响，才能对辍学原因进行确定的分析。例如，在大规模开放在线课程（MOOC）中，学生一般没有固定的学习方法。因此，要解决 MOOC 数据的时间性和多样性问题，并在辍学预测中考虑学生不一致的学习/学习活动，如观看和重看视频，是一项非常艰巨的任务，因为这些原因是多种多样的，而且是高度个性化的。Chen 等人[44]应用可视化分析方法和技术（DropoutSeer）分析 MOOC 系统的大型数据集，以可视化方式将 ML 预测的辍学率与 MOOC 用户的学习活动相关联。其目的是让内容设计者设计出更合适、更吸引人的内容，让人工智能专家设计出更好的预测模型[45]。例如，这比作为模型建立过程中关键步骤的特征识别过程更有效[46]。

Literature shows that most of the initial efforts in the domain are based on classical ML algorithms, such as Bayesian classifiers, RFs, ANNs, and SVMs. However, similar to the other applications, recently the trend shifted toward more advanced techniques, such as DL and genetic algorithms-based techniques [47], [48]. Moreover, more recently, some explainable AI-based solutions have also been proposed for the tasks. For instance, Hasib et al. [49] proposed a local interpretable model-agnostic explanations (LIME)-based model for the interpretation of results obtained with a diversified set of AI algorithms deployed for the performance prediction of secondary school students. Moreover, the key features/factors considered in students' performance prediction in literature include students' academic record (e.g., previous grades), demographic features (e.g., age, gender, income, race), school, instructor, and parents-related features, and students' behavioral-related features [50]. A summary of some of the existing AI-based tools and platforms for grading, students' retention, drop out, and performance prediction has been provided in Table II.
文献显示，该领域最初的研究大多基于经典的 ML 算法，如贝叶斯分类器、RF、ANN 和 SVM。然而，与其他应用类似，最近的趋势转向了更先进的技术，如基于 DL 和遗传算法的技术 [47]、[48]。此外，最近还有人针对这些任务提出了一些基于人工智能的可解释解决方案。例如，Hasib 等人[49]提出了一种基于本地可解释模型的解释（LIME）模型，用于解释在中学生成绩预测中使用的一套多样化人工智能算法所获得的结果。此外，文献中考虑的学生成绩预测的关键特征/因素包括学生的学业记录（如以前的成绩）、人口特征（如年龄、性别、收入、种族）、学校、教师和家长相关特征以及学生的行为相关特征[50]。表 II 提供了一些现有的基于人工智能的评分、学生保留率、辍学率和成绩预测工具和平台的摘要。

TABLE II Summary of Tools/Platforms for Grading, Students' Retention, Dropout, and Performance Prediction
表 II 用于评分、学生保留率、辍学率和成绩预测的工具/平台汇总表

B. Personalized Learning B.个性化学习

Personalized learning has been subject to many simultaneous and fundamental transformations mainly due to growing students' needs, globalization, new challenges in education management, and technological developments [62]. Technology implementation and inclusion notwithstanding, the traditional learning approach with a static and unidirectional model including the teacher in front of students, reading text material, and written exam-based assessments that cover all sections of the classroom uniformly is being eroded. Contemporary learning directions converge to interactive, student-focused, tailored learning models that serve each student or student group much closer with better engagement, closer interaction, improved comprehension, and wider scope coverage of learning outcomes.
主要由于学生需求的增长、全球化、教育管理的新挑战以及技术的发展，个性化学习同时经历了许多根本性的变革[62]。尽管有技术的实施和包容，但传统的静态和单向的学习方法，包括教师面对学生、阅读课文材料和统一覆盖课堂所有部分的基于考试的书面评估，正在受到侵蚀。当代的学习方向趋向于互动的、以学生为中心的、量身定做的学习模式，这种模式更贴近每个学生或学生群体，参与度更高，互动更密切，理解能力更强，学习成果的覆盖范围更广。

The adaptability of the learning model to cater to multiple learning settings would not have been feasible just a few years ago before the wide availability and accessibility of technology, such as AI, which leverages the power of cognitive computing, which makes use of reasoning, language processing, ML, and human capabilities/input allowing better solutions and data analysis, in the support of education [63]. Shawky and Badawi [64] explored RL, which represents a branch of AI algorithms, as a cognitive computing catalyst to provide adaptive learning materials and paths in support of bespoke, learner-centered requirements. To this aim, a myriad of personal, social, and environmental factors determining and affecting learners' experience have been monitored and analyzed in different learning settings to update and suggest new learning paths. The objective is to adapt to the “most influential” of these learning factors per learner needs and learning settings. The designed smart learning platform also recommends appropriate learning material in a connected, continuous way that adapts to the changing needs of the learners. The personalized learning approach is composed of three steps. First, a learning path/setting is suggested for learners based on their state. Subsequently, the learner's state and the reward received by the suggested learning path/settings are updated using RL, which ultimately leads to effective personalized learning settings based on the learner's needs. The approach is evaluated in an extensive simulation setup using 100 states and actions.
在人工智能等技术广泛应用和普及之前，学习模型适应多种学习环境的能力在几年前是不可行的，因为人工智能利用了认知计算的力量，利用推理、语言处理、ML 和人的能力/输入，为教育提供更好的解决方案和数据分析[63]。Shawky 和 Badawi [64]探讨了代表人工智能算法分支的 RL，将其作为认知计算的催化剂，提供自适应学习材料和路径，以支持定制的、以学习者为中心的要求。为此，在不同的学习环境中，对决定和影响学习者体验的无数个人、社会和环境因素进行了监测和分析，以更新和建议新的学习路径。目的是根据学习者的需求和学习环境，调整这些学习因素中 "最具影响力 "的因素。所设计的智能学习平台还以连接、持续的方式推荐适当的学习材料，以适应学习者不断变化的需求。个性化学习方法由三个步骤组成。首先，根据学习者的状态为其推荐学习路径/设置。随后，利用 RL 更新学习者的状态和建议的学习路径/设置所获得的奖励，最终根据学习者的需求进行有效的个性化学习设置。我们使用 100 种状态和操作对该方法进行了广泛的模拟评估。

AI can be used for a diversified set of tasks in personalized learning, however, in this section, we will mainly focus on ITS and recommender systems. In the next subsections, we provide an overview of these applications of AI in personalized learning.
人工智能可用于个性化学习中的各种任务，但在本节中，我们将主要关注智能学习系统和推荐系统。在接下来的小节中，我们将概述人工智能在个性化学习中的这些应用。

2) Recommendation Systems in Education
2) 教育领域的推荐系统

Recommendation systems have proven highly effective in various domains, such as business, food, tourism, and entertainment [80]. Similarly, these systems have found wide application in the education sector. In education, recommendation systems serve multiple purposes, aiding various stakeholders. For instance, they recommend relevant learning materials to students. In addition, they assist teachers in professional development by identifying useful documents and resources based on input from other teachers [81]. Moreover, recommendation systems can suggest remedial actions to enhance learning quality, supporting the operational side of academic teaching [82]. In this environment, assessment data collected over several academic semesters are analyzed based on learning outcomes at the course and program levels. Historical student attainment shortcomings, typically addressed by domain experts and course coordinators, are utilized to build a pool of remedial actions (recommendations) over a span of 3 to 5 years [83]. Training data/features, such as course domain, course level, section size, and lab options, guide experts in selecting recommended remedial actions for subsequent assessments, whether formative or summative. A multilabel classification algorithm is employed to choose suitable actions for each rubric line, representing performance per group of students, from the master pool. AI provides notable strengths in this domain, offering efficiency, consistency, and fairness in the application of remedial actions. While this setup is particularly suitable for large colleges with extensive student populations and archives of structured, outcome-based assessment data spanning several years, the approach has demonstrated success and reasonable accuracy even with a smaller number of learning instances [83].
事实证明，推荐系统在商业、食品、旅游和娱乐等各个领域都非常有效[80]。同样，这些系统在教育领域也得到了广泛应用。在教育领域，推荐系统有多种用途，可为不同的利益相关者提供帮助。例如，它们向学生推荐相关的学习材料。此外，它们还能根据其他教师的意见，识别有用的文件和资源，从而帮助教师进行专业发展[81]。此外，推荐系统还能建议采取补救措施以提高学习质量，从而为学术教学的运作提供支持[82]。在这种环境下，根据课程和项目层面的学习成果，对几个学期收集的评估数据进行分析。通常由领域专家和课程协调员处理的历史学生成绩缺陷，被用来建立一个跨度为 3 至 5 年的补救行动（建议）库[83]。训练数据/特征，如课程领域、课程级别、章节规模和实验选项，指导专家为后续评估（无论是形成性评估还是总结性评估）选择推荐的补救措施。采用多标签分类算法，从主库中为代表每组学生成绩的每条评分标准线选择合适的行动。人工智能在这一领域具有显著的优势，它在应用补救措施方面具有高效性、一致性和公平性。虽然这种设置特别适用于拥有大量学生和结构化、基于结果的多年评估数据档案的大型学院，但这种方法即使在学习实例数量较少的情况下也取得了成功，并具有合理的准确性[83]。

Table IV summarizes some existing recommendation systems in education for both students and teachers. Most of the existing recommendation systems aim to select/recommend relevant educational resources, learning activities, and finding peers [80]. Similar to other application domains, the majority of educational recommendation systems are based on content-based, collaborative filtering, or hybrid recommendation methods. Moreover, a diversified set of metrics is used for the evaluation of recommendation systems in the domain, such as students' grades and learning outcomes [84].
表 IV 总结了一些现有的面向学生和教师的教育推荐系统。大多数现有的推荐系统旨在选择/推荐相关的教育资源、学习活动和寻找同伴[80]。与其他应用领域类似，大多数教育推荐系统都是基于内容、协同过滤或混合推荐方法。此外，该领域中的推荐系统还采用了多样化的评估指标，如学生的成绩和学习成果[84]。

TABLE III Summary of AI-Based Tools/Platforms for Personalized Learning
表 III 基于人工智能的个性化学习工具/平台概述

TABLE IV Summary of AI-Based Tools/Platforms for Recommendation Systems in Education
表 IV 基于人工智能的教育推荐系统工具/平台概述

C. Sentiment Analysis in Education
C.教育领域的情感分析

Sentiment analysis involves analyzing and extracting people's opinions about a service or an entity. Often referred to as opinion mining, it is a challenging task that generally involves different phases, such as collection and storage of data as well as analysis of the data using a combination of knowledge-based and ML techniques [88]. In the context of education, sentiment analysis attempts to improve the learning process by analyzing students' feedback to better understand their opinions, emotions, and concerns and make adjustments to the content or delivery of the learning material accordingly [89]. Sentiment analysis is generally associated with the extraction of people's emotions, which in the learning context could be very beneficial as emotions may affect students' motivation and learning outcomes [90]. Thus, early identification and proper handling of students' emotions and concerns may result in a better learning experience/process.
情感分析涉及分析和提取人们对某项服务或实体的看法。它通常被称为意见挖掘，是一项具有挑战性的任务，一般涉及不同的阶段，如收集和存储数据，以及使用基于知识的技术和 ML 技术对数据进行分析 [88]。在教育领域，情感分析试图通过分析学生的反馈来改善学习过程，从而更好地了解他们的意见、情感和担忧，并相应地调整学习材料的内容或交付[89]。情感分析通常与提取人的情感有关，这在学习环境中可能非常有益，因为情感可能会影响学生的学习动机和学习效果[90]。因此，及早识别和妥善处理学生的情绪和担忧可能会带来更好的学习体验/过程。

Sentiment analysis also plays a significant role in extracting students' opinions on learning materials from social media. According to Chauhan et al. [91], students' views are continuously exchanged using different social media platforms in response to their experiences in the classroom or online learning; these provide a rich pool of data for evaluating students learning. From an MOOCs perspective, Kastrati et al. [92] demonstrated the effectiveness of sentiment analysis in automatically analyzing students' feedback. The authors also highlight the labor involved in manually assessing and annotating students' feedback. To overcome this limitation, they proposed a weakly supervised framework for aspect-level sentiment analysis specifically aiming to highlight polarity in student feedback about MOOCs. Liu et al. [93] argued that the continuous feedback of students in MOOC environments stipulates that students' emotions and learning activities be tracked for understanding learning requirements. To analyze students' emotions (i.e., positivity, negativity, and confusion) and concerned aspects (e.g., teaching styles and learning activities), the authors proposed the temporal emotion-aspect model (TEAM) which tracks students' emotions toward the concerned aspects and characterizes their joint distribution over time. In other words, the model associates emotions with different aspects and their evolution over a period of time. The results indicated that: 1) content-related aspects were the main emphasis with a higher likelihood of confused or negative emotions; 2) there were higher likelihoods of emotional expressions at the start and end of a semester; and 3) underachieving students were less active in emotional engagement and tended to express more confusion toward the end of a semester when compared to high-achieving and medium-achieving students. Such observations could be useful to teachers for timely instructional guidance or psychological intervention that could help the learners in the future. Applications of this technology in education and other fields are already established with favorable results in education, healthcare, social media, and natural language processing domains [89].
情感分析在从社交媒体中提取学生对学习材料的意见方面也发挥着重要作用。Chauhan 等人[91]指出，学生在不同的社交媒体平台上针对自己在课堂或在线学习中的体验不断交换意见；这些都为评估学生的学习情况提供了丰富的数据。从 MOOCs 的角度来看，Kastrati 等人[92] 证明了情感分析在自动分析学生反馈方面的有效性。作者还强调了人工评估和注释学生反馈所涉及的工作量。为了克服这一局限性，他们提出了一个弱监督的方面级情感分析框架，专门用于突出学生对 MOOCs 的反馈中的极性。Liu 等人[93]认为，MOOC 环境中学生的持续反馈要求跟踪学生的情绪和学习活动，以了解学习需求。为了分析学生的情绪（即积极、消极和困惑）和相关方面（如教学风格和学习活动），作者提出了时态情绪-方面模型（TEAM），该模型跟踪学生对相关方面的情绪，并描述它们随时间的共同分布。换言之，该模型将情感与不同方面及其在一段时间内的演变联系起来。研究结果表明1）与内容相关的方面是主要重点，出现困惑或消极情绪的可能性较高；2）在学期开始和结束时，出现情绪表达的可能性较高；3）与成绩优秀和成绩中等的学生相比，成绩较差的学生在情绪参与方面不太积极，在学期结束时倾向于表达更多的困惑。这些观察结果有助于教师及时进行教学指导或心理干预，从而对学习者的未来有所帮助。这项技术在教育和其他领域的应用已经确立，并在教育、医疗保健、社交媒体和自然语言处理领域取得了良好的效果[89]。

The other key role of opinion mining/sentiment analysis in education research includes the investigation of learners' satisfaction with the available resources, their attitude toward learning, their concerns, and evaluation of teachers' performance, and teaching methods [94]. For instance, Munezero et al. [95] analyzed students' learning diaries to predict students' sentiments, emotions, and opinions about their learning experience. According to Kechaou et al. [96], knowledge and evaluation of user opinions is an essential prerequisite for the effective development of e-learning systems. To this end, an opinion mining method has been applied in their research to support e-Learning content developers to enhance the quality of provided services using three feature selection methods, namely, mutual information (MI), information gain (IG), and CHI statistics (CHI) in conjunction with hidden Markov models (HMM) and SVM-based hybrid learning methods. Experimental results indicate that opinion mining is more challenging in e-learning blogs in the presence of noise. Although this work has shown that IG constitutes the optimal potential for sentimental term selection and produced optimum accuracy in sentiment classification. More recently, Mostafa et al. [97] reviewed work in sentiment analysis related to gamification in learning, the author proposed a classifier that will analyze the sentiments of students while using gamification tools for learning in Egypt.
意见挖掘/情感分析在教育研究中的另一个关键作用包括调查学习者对可用资源的满意度、他们的学习态度、他们的担忧以及对教师表现和教学方法的评价[94]。例如，Munezero 等人[95] 通过分析学生的学习日记来预测学生对学习经历的情绪、情感和观点。Kechaou 等人[96]认为，了解和评估用户意见是有效开发电子学习系统的必要前提。为此，他们在研究中采用了一种意见挖掘方法，利用三种特征选择方法，即互信息（MI）、信息增益（IG）和CHI统计（CHI），结合基于隐马尔可夫模型（HMM）和SVM的混合学习方法，支持电子学习内容开发人员提高所提供服务的质量。实验结果表明，在存在噪声的电子学习博客中，意见挖掘更具挑战性。尽管这项工作表明，IG 构成了情感术语选择的最佳潜力，并在情感分类中产生了最佳准确性。最近，Mostafa 等人[97]回顾了与学习游戏化相关的情感分析工作，作者提出了一种分类器，可以分析埃及学生在使用游戏化工具学习时的情感。

The majority of the recent efforts in the domain are based in higher education, where several important topics, such as the collection of relevant and meaningful data for sentiment analysis, designing sentiment analysis methodologies, and exploring the relationship between sentiment and behavioral analysis with the learners' performance and achievement, are explored [94]. The majority of the state-of-the-art solutions are based on textual information, where mostly DL techniques, such as LSTM, biLSTM, and transformers, such as bidirectional encoder representations from transformers (BERT) and robustly optimized BERT (RoBERTa) are used. However, literature also hints at the use of visual sentiment analysis techniques for analyzing and extracting students' emotions and expressions [98].
该领域的最新成果大多基于高等教育，其中探讨了几个重要课题，如收集相关和有意义的数据用于情感分析、设计情感分析方法以及探索情感和行为分析与学习者的表现和成绩之间的关系[94]。大多数最先进的解决方案都是基于文本信息的，其中大多使用了 DL 技术，如 LSTM、biLSTM 和变换器，如来自变换器的双向编码器表示法（BERT）和鲁棒性优化 BERT（RoBERTa）。不过，也有文献提示使用视觉情感分析技术来分析和提取学生的情绪和表情[98]。

The research efforts in the domain resulted in several interesting AI-based sentiment analysis tools. Some of the existing sentiment analysis tools developed or customized for education are provided in Table V.
该领域的研究工作产生了几种有趣的基于人工智能的情感分析工具。表 V 列出了一些为教育开发或定制的现有情感分析工具。

TABLE V Summary of Tools/Platforms for Sentiment Analysis in Education
表 V 教育领域情感分析工具/平台概述

D. Classroom Monitoring and Visual Analysis
D.课堂监控和视觉分析

Classroom monitoring and visual analysis play a key role in classroom management. It can be used for a diversified list of tasks to support teachers in different ways [104]. For instance, classroom monitoring and visual analysis of the classrooms allow for analyzing students' reactions to course contents and other learning materials and activities. It also allows for investigating the students' behavior in the classroom in terms of the time they keep and lose focus during lectures. In addition, it could also be used to analyze teachers' spatial behavior in classrooms, which has a significant impact on students' motivation and engagement in learning [105]. However, monitoring and analyzing individuals in a classroom for a long time is not an easy task. Thanks to advanced video analytics and AI algorithms, it is possible to automatically analyze students' responses, reactions, and levels of attentiveness during lectures. For instance, Narendra et al. [106] proposed a video analytic framework to monitor and investigate how long and when students keep and lose focus during lectures. Similarly, Martinez et al. [107] proposed an interactive teacher's dashboard enabling teachers to keep an eye on groups' learning activities and collaborations in a multitabletop learning environment. This real-time monitoring and analysis could overcome the limitations of conventional supervision, where the teachers can only see the final product of the students/group.
课堂监控和可视分析在课堂管理中发挥着关键作用。它可用于多种任务，以不同方式为教师提供支持[104]。例如，通过课堂监控和可视化分析，可以分析学生对课程内容、其他学习材料和活动的反应。它还可以调查学生在课堂上的行为，如他们在讲课过程中保持和失去注意力的时间。此外，它还可用于分析教师在课堂上的空间行为，这对学生的学习动机和参与度有重大影响[105]。然而，长时间监控和分析教室中的个人并非易事。借助先进的视频分析和人工智能算法，可以自动分析学生在授课过程中的反应、回应和专注程度。例如，Narendra 等人[106] 提出了一个视频分析框架，用于监测和研究学生在讲课过程中保持和失去注意力的时间长短。同样，Martinez 等人[107] 提出了一种交互式教师仪表板，使教师能够在多桌面学习环境中密切关注各小组的学习活动和协作情况。这种实时监控和分析可以克服传统监督的局限性，即教师只能看到学生/小组的最终成果。

Classroom utilization and occupancy calculations, which are part of budgetary planning and strategic planning of higher education institutions, especially where real estate is a premium asset [108], is another key application of classroom monitoring and visual analysis. Students in modern offline or online degrees have many technology-driven advantages at their fingertips but equally, suffer excessive demands which often cause dropouts and classroom underutilization. Although predicting room occupancy/utilization is an age-old problem [109], the use of modern AI technology as an instrument in measuring or increasing the efficiency of room utilization is a new topic. Sutjaritthamet et al. [110] used on-campus sensor instruments to monitor classroom attendance while respecting student privacy. Several measurement approaches were evaluated in a lab experiment to identify the best sensor technology in terms of cost, accuracy, and convenience.
教室利用率和占用率的计算是高等教育机构预算规划和战略规划的一部分，尤其是在不动产是优质资产的情况下[108]，这也是教室监控和可视化分析的另一个关键应用。现代线下或线上学位的学生拥有许多技术驱动的优势，但同样也面临着过高的要求，这往往会导致辍学和教室利用率不足。虽然预测教室占用率/利用率是一个老问题[109]，但使用现代人工智能技术作为测量或提高教室利用效率的工具却是一个新课题。Sutjaritthamet 等人[110]在尊重学生隐私的前提下，使用校内传感器仪器监测课堂出勤率。他们在实验室实验中评估了几种测量方法，以确定在成本、准确性和便利性方面最佳的传感器技术。

AI also has an impact on technology enhanced learning (TEL) by providing several interesting applications in many subdomains. One such area is to aid the difficult task of understanding the various dimensions of TEL in schools. One reason for this difficulty is the limitation of monitoring classrooms for a longer period to analyze teachers' teaching methods and students' learning experiences. Howard et al. [111] explored the area of observing, analyzing, and visualizing TEL classrooms over time and used sensors to collect observation data over two months. This data are presented as insights to academic administrators and teachers for reflection and corrective action to enhance student learning.
人工智能也对技术强化学习（TEL）产生了影响，在许多子领域提供了一些有趣的应用。其中一个领域是帮助理解学校中技术强化学习的各个层面这一艰巨任务。造成这一困难的原因之一是，对课堂进行较长时间的监测以分析教师的教学方法和学生的学习经验受到了限制。Howard 等人[111]探索了长期观察、分析和可视化 TEL 课堂的领域，并使用传感器收集了两个月的观察数据。这些数据将作为见解呈现给教学管理人员和教师，供他们反思和采取纠正措施，以提高学生的学习效果。

Other applications of EDM and AI for the transformation of the traditional classroom include the analysis of student facial expressions to assess their level of engagement in the classroom. Soloviev et al. [112] proposed a system that analyzes (in real-time) the data feeds from video cameras that are installed in the classroom and applies AI and facial recognition technology to recognize student emotions to determine their level of enjoyment. Although Chua et al. [113] reviewed case studies and technologies developed to collect and analyze educational data. Several aspects of the learning environment, which is a combination of the physical and digital classroom setting, are studied. Moreover, different aspects of the learning process are assessed and analyzed to quantify teaching and learning processes, student assessments are also analyzed automatically. The authors introduce data pipelines that leverage data and information collected from both physical spaces as well as digital spaces.
EDM 和人工智能在改造传统课堂方面的其他应用包括分析学生的面部表情，以评估他们在课堂上的参与程度。Soloviev 等人[112]提出了一种系统，该系统（实时）分析安装在教室里的视频摄像头提供的数据，并应用人工智能和面部识别技术来识别学生的情绪，以确定他们的喜爱程度。Chua 等人[113]回顾了为收集和分析教育数据而开发的案例研究和技术。研究了学习环境的几个方面，即物理教室和数字教室环境的结合。此外，还对学习过程的不同方面进行了评估和分析，以量化教学和学习过程，并自动分析学生的评估结果。作者介绍了利用从物理空间和数字空间收集的数据和信息的数据管道。

The majority of the current efforts in classroom monitoring and visual analysis rely on state-of-the-art DL algorithms, such as CNNs, convolutional encoder/decoder networks, and hybrid models by combining CNNs and RNNs [114]. Moreover, the key tasks involved in the process generally include facial landmark extraction, face segmentation, head pose estimation, and facial expression recognition. As part of the efforts, several AI-based tools have been developed that can help in classrooms in several ways, such as security, marking attendance rolls, and emotional and movement monitoring for better classroom dynamics analysis. Table VI summarizes some of the existing classroom monitoring systems.
目前，课堂监控和视觉分析领域的大部分工作都依赖于最先进的数字逻辑算法，如 CNN、卷积编码器/解码器网络，以及结合 CNN 和 RNN 的混合模型 [114]。此外，这一过程中涉及的关键任务一般包括面部地标提取、面部分割、头部姿态估计和面部表情识别。作为努力的一部分，已经开发出几种基于人工智能的工具，可以在多个方面为教室提供帮助，如安全、考勤打分、情绪和动作监测，以便更好地分析教室动态。表 VI 总结了一些现有的教室监控系统。

TABLE VI Summary of AI-Based Tools/Platforms for Classrooms' Monitoring and Visual Analysis
表 VI 基于人工智能的教室监控和可视化分析工具/平台一览表

A. Supervised Learning A.监督学习

The majority of the works on AI in education rely on supervised learning, as detailed in Section IV. Supervised learning aims at function approximation or curve fitting by finding a relation/function f:x→y using a training set {x,y}. Though the efficiency of supervised learning largely depends on the availability and quality of training data, it is a far more accurate learning strategy compared to its counterparts [121]. Supervised ML algorithms can further be divided into several categories at different hierarchies. A complete taxonomy of supervised learning techniques can be found in [121]. Some well-known techniques include RF, conditional RFs (CRFs), SVMs, decision trees, NNs, logistic and linear regressions, belief networks, naive Bayes, and Markov random fields and Markov models.
大多数教育领域的人工智能作品都依赖于监督学习，详见第四部分。监督学习的目的是通过使用训练集 {x,y} 找到关系/函数 f:x→y 来实现函数逼近或曲线拟合。虽然监督学习的效率在很大程度上取决于训练数据的可用性和质量，但与同类算法相比，监督学习是一种更为精确的学习策略 [121]。有监督的 ML 算法可进一步分为不同层次的多个类别。监督学习技术的完整分类法可参见 [121]。一些著名的技术包括 RF、条件 RF（CRF）、SVM、决策树、NN、逻辑回归和线性回归、信念网络、天真贝叶斯以及马尔可夫随机场和马尔可夫模型。

In education, supervised learning is mostly used in predictive analysis, such as grading [122], retention, and dropout prediction [123]. For instance, Majeed et al. [122] proposed several supervised learning techniques for students' grade prediction. In detail, around 2500 students' records were collected from a degree-awarding institution to train different supervised learning algorithms including naive Bayes and K-nearest neighbor classifiers. Similarly, in [124], several supervised learning algorithms including decision tree-based algorithms, naive Bayes, k-NN, linear models, and deep learning, are employed for the identification of students at risk using around 15 825 samples from the Budapest University of Technology and Economics.
在教育领域，监督学习主要用于预测分析，如评分[122]、留级和辍学预测[123]。例如，Majeed 等人[122] 提出了几种用于学生成绩预测的监督学习技术。具体而言，他们从一所学位授予机构收集了约 2500 份学生记录，用于训练不同的监督学习算法，包括天真贝叶斯和 K 近邻分类器。同样，在文献[124]中，利用来自布达佩斯技术经济大学的约 15 825 个样本，采用了多种监督学习算法，包括基于决策树的算法、天真贝叶斯、k-NN、线性模型和深度学习，来识别有风险的学生。

One of the main limitations of a supervised learning-based strategy in the education sector is the lack of quality datasets, as detailed in Section V. In order to overcome these limitations, a modified form of supervised learning, namely, semisupervised learning aiming to exploit partially labeled train sets for classification tasks, has been introduced. For instance, Livieris et al. [38] proposed a semisupervised learning-based framework for secondary school students' performances.
在教育领域，基于监督学习的策略的主要局限之一是缺乏高质量的数据集，详见第五部分。为了克服这些局限，人们引入了监督学习的一种改进形式，即半监督学习，旨在利用部分标记的训练集来完成分类任务。例如，Livieris 等人[38] 针对中学生的表现提出了一种基于半监督学习的框架。

B. Unsupervised Learning B.无监督学习

Unsupervised learning, which aims to discover or extract patterns of regularities and irregularities in a set of observations, has also been widely exploited in educational data analysis. Unsupervised algorithms process and discover hidden patterns in input samples without needing any training samples, and thus, are easy to implement and deploy in an application. Unsupervised ML algorithms can be mainly divided into two categories, namely, 1) clustering and 2) dimensionality reduction techniques, which are further divided into subgroups. A detailed taxonomy of unsupervised ML algorithms has been provided in [121]. Clustering algorithms aim to divide a collection of samples into clusters or segments while dimensionality reduction algorithms are used to extract a small set of relevant features for building a reliable model.
无监督学习旨在发现或提取一组观测数据中的规律性和不规则性模式，在教育数据分析中也得到了广泛应用。无监督算法无需任何训练样本即可处理和发现输入样本中隐藏的模式，因此易于在应用中实施和部署。无监督 ML 算法主要可分为两类，即 1) 聚类和 2) 降维技术，这两类算法又进一步分为若干子类。文献[121]对无监督 ML 算法进行了详细分类。聚类算法旨在将样本集合划分为群组或片段，而降维算法则用于提取一小部分相关特征，以建立可靠的模型。

In literature, unsupervised learning—especially clustering algorithms— has been mostly used in EDM to extract useful information for a diverse set of applications from raw data [125]. Some of the applications of EDM in which clustering has been proven very effective include students' performance prediction [126], students' profiling and modeling [127], recommendation systems [128], enrollment management [129], constructing course contents [130], and analyzing students' behavior [131]. Similarly, dimensionality reduction algorithms, such as principal component analysis (PCA) and linear discriminant analysis (LDA), have also been employed in educational data analysis. For instance, Borges et al. [132] employed PCA for students' performance prediction and data analysis.
在文献中，无监督学习--特别是聚类算法--被广泛应用于 EDM，从原始数据中提取有用信息，用于各种应用[125]。聚类已被证明非常有效的 EDM 应用包括学生成绩预测[126]、学生分析和建模[127]、推荐系统[128]、招生管理[129]、课程内容构建[130]和学生行为分析[131]。同样，主成分分析（PCA）和线性判别分析（LDA）等降维算法也被用于教育数据分析。例如，Borges 等人[132] 将 PCA 用于学生成绩预测和数据分析。

C. Reinforcement Learning
C.强化学习

In the beginning, RL was mostly restricted to robotics and game theory, however, more recently it has been deployed in other application domains as well [121]. A significant portion of literature, especially the work presented in top venues, as detailed in Section IV, is based on RL. RL provides a set of recommended actions to maximize reward in a particular situation/application. RL differs from supervised learning in several ways. For instance, supervised learning algorithms are trained on class labels to predict a class while RL algorithms are trained on a reward signal and predict/recommend an action to solve a particular problem. Moreover, RL performs a task in a sequential way where the input depends on the previous decision.
最初，RL 主要局限于机器人学和博弈论，但最近它也被应用于其他应用领域 [121]。很大一部分文献，尤其是第四节详述的顶级会议上介绍的工作，都是基于 RL 的。RL 提供了一组推荐行动，以在特定情况/应用中获得最大回报。RL 在几个方面不同于监督学习。例如，监督学习算法是根据类别标签进行训练以预测类别，而 RL 算法是根据奖励信号进行训练，并预测/推荐解决特定问题的行动。此外，RL 以连续的方式执行任务，输入取决于之前的决定。

Similar to supervised and unsupervised learning, RL algorithms can be divided into different categories. RL algorithms can mainly be categorized as Markovian or evolutionary. A complete taxonomy of RL can be found in [121].
与有监督和无监督学习类似，RL 算法也可分为不同类别。RL 算法主要分为马尔可夫算法和进化算法。关于 RL 的完整分类可参见 [121]。

In education, RL has been mainly used for generating feedback for students on time series data [133], modeling students' learning style [134], personalized learning [135], adaptive tutorial modeling [136], and improving students' problem-solving capabilities [137].
在教育领域，RL 主要用于为学生生成时间序列数据反馈 [133]、学生学习风格建模 [134]、个性化学习 [135]、自适应辅导建模 [136]，以及提高学生解决问题的能力 [137]。

D. Generative AI D.生成式人工智能

Generative AI is a cutting-edge field within AI that focuses on creating models capable of generating new content, such as text, images, and even music. Unlike traditional ML techniques like supervised learning, unsupervised learning, and RL, which primarily focus on pattern recognition and optimization, generative AI models aim to simulate human creativity and produce original outputs in various forms including text, audio, images, and videos. In recent years, generative AI has gained immense popularity, thanks to the emergence of pretrained large language models (LLMs) trained on vast amounts of Internet data using large-scale parallel computing with new transformer-based techniques [138]. These pretrained models, often referred to as foundation models [139], are designed to be versatile and applicable to a wide range of tasks that were not explicitly specified during their training. They can be utilized with prompts to enable zero-shot learning, where they provide direct answers, or one-shot/multishot learning, where prompts include examples. Furthermore, these models can be fine-tuned for more specific applications later on. Prominent examples of foundation models include BERT, T5, and the GPT family, including GPT3 and GPT4. The widespread adoption of generative AI has been fueled by the popularity of platforms like ChatGPT, which leverages a conversational interface built on top of the GPT3.5 model, capturing the public's attention and imagination. These foundation models have revolutionized entire fields such as NLP by eliminating the need for developing bespoke models. Instead, general-purpose models are built once and can be utilized in various contexts for a multitude of tasks.
生成式人工智能是人工智能中的一个前沿领域，其重点是创建能够生成文本、图像甚至音乐等新内容的模型。与监督学习、无监督学习和 RL 等主要关注模式识别和优化的传统 ML 技术不同，生成式人工智能模型旨在模拟人类的创造力，并生成文本、音频、图像和视频等各种形式的原创输出。近年来，生成式人工智能大受欢迎，这要归功于利用大规模并行计算和基于变压器的新技术在海量互联网数据上训练的预训练大型语言模型（LLMs）的出现[138]。这些预训练模型通常被称为基础模型[139]，其设计目的是使其具有多功能性，并适用于在其训练过程中未明确指定的各种任务。这些模型可与提示一起使用，以实现零点学习（直接提供答案）或单点/多点学习（提示包括示例）。此外，这些模型以后还可以针对更具体的应用进行微调。基础模型的突出例子包括 BERT、T5 和 GPT 系列，包括 GPT3 和 GPT4。ChatGPT 等平台的流行推动了生成式人工智能的广泛应用，该平台利用建立在 GPT3.5 模型之上的对话界面，吸引了公众的注意力和想象力。这些基础模型消除了开发定制模型的需要，从而彻底改变了 NLP 等整个领域。取而代之的是，通用模型只需构建一次，就能在各种情况下用于多种任务。

This advancement holds great promise in the field of education, particularly since the introduction of ChatGPT on top of GPT3.5 and GPT4, which has captured the interest of the education community since the release of ChatGPT near the end of the year 2022. Educators have recognized the transformative potential of this technology in enabling personalized education through a conversational interface. Recent studies and research have delved into the implications of generative AI and LLMs in education, highlighting both the promise and the potential risks associated with these technologies [140], [141], [142]. The generative AI models are used for several tasks in the education sector, such as sentiment analysis and feedback [143], interactive chat-based teacher training [144], and student assessments [145] etc.
这一进步在教育领域大有可为，尤其是在 GPT3.5 和 GPT4 的基础上引入 ChatGPT 后，自 ChatGPT 于 2022 年年底发布以来，已引起教育界的浓厚兴趣。教育工作者已经认识到这项技术在通过对话界面实现个性化教育方面的变革潜力。最近的调查和研究深入探讨了生成式人工智能和LLMs在教育领域的影响，强调了这些技术的前景和潜在风险[140], [141], [142]。生成式人工智能模型被用于教育领域的多项任务，如情感分析和反馈 [143]、基于聊天的交互式教师培训 [144] 和学生评估 [145] 等。

A. Authors-Based Productivity Analysis

Author productivity is one of the common methods to evaluate significant entities. By consulting the work of top authors in a domain, the directions of a research domain can be easily determined. Fig. 1 shows the most published authors in the field of AI in education. We observe that authors from USA-based organizations are significantly contributing to the field of AI in education. Neil Heffernan from Worcester Polytechnic Institution, Kenneth Koedinger from Carnegie Mellon University, and Ryan S.J.D. Baker from the University of Pennsylvania are collectively ranked as the top three authors in the field of EDM. Most of the subsequent authors on the list also belong to USA-based institutions.

Fig. 1.

Authors with the highest publication count during 2014–2022. USA-based authors appear to be prominent in this list.

Show All

We also observe the trends of coauthorship in AI in education in Fig. 2. The average number of authors during 2014–2022 remains more or less constant at three authors per paper but the spread of authorship has increased in recent years. Some papers have experienced a higher number of coauthorship as well, e.g., in 2022, the maximum number of authors of a paper was 14 authors.

Fig. 2.

Distribution of authorship during 2014–2022 in our dataset. Although a median number of authors remained more or less constant throughout the mentioned time period but spread of co-authorship increased in recent years.

Show All

B. Institution and Country-Based Productivity Analysis

This subsection deals with the varying research trends of AI in education in different institutions and countries. Fig. 3 shows the most publishing institutions in the field of AI in education. Almost all of the top institutions are from the USA, which shows the significant research contributions in this domain by the USA.

Fig. 3.

Institutions with the highest number of publications in our dataset during 2014–2022. Almost all of the top publishing institutions are from USA.

Show All

Fig. 4 shows the rank of a contributing country in the field of AI in education using a global heat map. The United States is in the highest position in the field of AI in education in terms of publication count. Other top countries include China, Canada, India, and the United Kingdom in AI in education.

Fig. 4.

Rank of countries based on their publication count in our dataset during 2014–2022. USA emerges as top contributor followed by China, Canada, India, and Australia/U.K.

Show All

C. Knowledge Flow

First, we extract the references from all papers and create a citation graph, as we are curious to understand how venues in AI in education cite each other. Fig. 5 is a Sankey diagram that shows the fraction of papers that AI in education papers reference (left), as well as the other papers that in turn cite the papers in our dataset (right).

Fig. 5.

Distribution references and citations in AI in education venues during 2014–2022. The left input shows the conferences that are referenced by our dataset; the right output shows which papers cite publications in our dataset. Major sources of references and citations in our dataset are from journals.

Show All

Interesting patterns emerge from this analysis. Most noteworthy is the bias for citing papers from the same venue. For example, 31% of the references in papers for AI in education are from other papers previously published in the AI in education conferences. In contrast, a far more diverse set of conferences and journals cites articles from venues considered in our dataset. 51% of the papers in our dataset are cited by journals, rather than conferences. Major citers of papers in our dataset include Computer and Education and LNCS (which subsumes many proceedings) besides the venues where the paper is published.

All that said, it is clear that several other publication venues feature heavily in the bibliographies of papers of our dataset, and these are dominated by journals rather than conferences.

D. Relationships Between Applications and Techniques

It is also important to provide readers with an overview of AI techniques employed in different applications of AI in education. To this aim, in Fig. 6 we provide the statistics of four main categories of AI techniques using data-driven AI methodologies in terms of the number of publications in different applications in top venues. As can be observed, in most of the applications, supervised or semisupervised techniques of learning have been employed suggesting the availability of the annotated data in the majority of the applications. Unsupervised learning techniques have also been widely employed in some of the applications, such as e-learning, student evaluation, ITS, and personalized learning. Similarly, RL has also been employed in several works on ITS, student evaluation and retention, and dropout prediction. We also observe the growing number of works on ITS, recommender systems, and personalized learning are using generative AI.

Fig. 6.

Relationship between applications and AI techniques.

Show All

E. Relationship Between Applications and Venues

Fig. 7 provides the statistics of some interesting applications of AI in education in terms of the number of papers published on each in the leading venues. The most popular application of AI in education is in developing ITSs followed by its use for evaluation and personalized learning.

Fig. 7.

Statistics of AI papers at top venues in terms of applications.

Show All

F. Relationship Between Techniques and Venues

Fig. 8 shows the statistics of the papers published based on the four types of data-driven AI methods, namely: 1) supervised learning, 2) unsupervised learning, 3) RL, and 4) generative AI. As can be seen in the figure, algorithms from each of the four categories have been deployed in educational data analysis works presented in the aforementioned venues in our dataset. The most common technique type by far is supervised learning followed by unsupervised learning and RL, and then generative AI.

Fig. 8.

Statistics of AI papers at different venues and techniques.

Show All

A. Insights

Some key insights from literature on the use of AI in education discussed in the article are summarized as follows.

B. Future Research Directions and Open Issues

In this section, we provide some potential directions for future research in the domain.