International Journal of Artificial Intelligence and Education Technology

The academic success of students who are nearing academic failure should be Identifying students who are at risk of academic failure or course withdrawal at an early stage of their enrolment remains one of the most pressing challenges in higher and distance education. The research assesses the performance of seven machine learning classifiers which include Logistic Regression Decision Tree Random Forest Gradient Boosting Decision Tree (GBDT) AdaBoost Naive Bayes and Multilayer Perceptron for predicting student risk at an early stage based on a behavioural and demographic dataset derived from the Open University Learning Analytics Dataset (OULAD). The dataset contains 7895 student records which represent a single module and show eight demographic factors together with eight Virtual Learning Environment (VLE) usage patterns. All classifiers were evaluated through five-fold stratified cross-validation. The GBDT model achieved the best results with an AUC-ROC value of 0.782 (} 0.003) and an accuracy rate of 0.708 (} 0.005) which produced an F1 score of 0.729 (} 0.006) and a recall rate of 0.769 (} 0.006). The analysis of feature importance showed that late sub-mission count (I = 0.304) and total VLE clicks (I = 0.150) together with first assessment score (I = 0.135) serve as the three most valuable predictive indicators because they help identify student engagement patterns which become evident through VLE traces that educational institutions collect from students during their first module. Educational institutions can utilize learning management system data to implement effective combi-nation methods which enable them to execute necessary teaching methods even though they do not need to gather additional expense data. The article presents design elements which both create early warning systems and manage the ethical use of predictive analytics within educational systems.

AI-based quiz and assessment tools are widely studied for supporting adaptive learning, yet existing work is distributed across different tasks (e.g., question generation, automatic evaluation, feedback, and conversational assessment) and often uses inconsistent datasets and metrics, making comparisons difficult. This paper reports a Systematic Literature Review (SLR) conducted under PRISMA 2020 to summarize approaches and evaluation practices for AI-based quiz and assessment systems. Searches were performed in IEEE Xplore, ACM Digital Library, and Google Scholar using keyword combinations related to automated question generation, assessment, evaluation, and large language models. The search returned Nidentified=57 records; after duplicate removal, Ndedup=55 records remained for screening. Following title/abstract screening and full-text eligibility assessment, Nincluded=9 studies were included for qualitative synthesis and structured data extraction. The reviewed studies show strong attention to transformer/LLM-based question generation, automatic scoring and evaluation frameworks, and formative feedback generation for learning. However, recurring limitations include reliability of automated judging, lack of standardized benchmarks, domain transferissues, and risks impacting fairness and academic integrity. We conclude with practical recommendations for stronger evalua-tion design (e.g., shared benchmarks, transparent rubrics, and human-in-the-loop validation) to improve trust and real-world adoption.

Artificial intelligence is increasingly used to monitor learning processes in higher education; however, many analytics pipelines still terminate at prediction and provide limited support for instructional action. The research establishes an explainable artificial intelligence framework which utilizes digital learning environment behavioral data and contextual information to create customized instructional support solutions. The analysis uses xAPIEdu-Data dataset which contains 480 records to build engagement index and create support profiles and predict multiclass performance through rule based action allocation. The study tests three classification models using stratified cross validation. The study selects Random Forest as the most effective system because it delivers superior results across all tests. The selected model demonstrates 0.8021 accuracy and 0.8204 macro precision and 0.8010 macro recall and 0.8084 macro F1 score and 0.9140 macro area under the curve on the hold-out sample. The analysis shows that student absence and composite engagement index andgender and student guardian relationship and support profile and digital resource access arethe most important factors that determine student performance. The final decision layer manages student assignment to instructional support plans which contain attendance-first intervention and adaptive engagement support and family-engagement reinforcement and structured progression coaching and challenge-and-extend pathways. The study develops an analytical framework which connects explainable artificial intelligence to digital higher education instructional decision support systems.