Student Performance Prediction in Intelligent Tutoring Systems: Methods, Challenges, and Future Perspectives

Abstract

The application of predictive models within Intelligent Tutoring Systems (ITS) has seen significant advancements, enhancing personalised learning and early identification of at-risk students. However, challenges related to model generalisability, interpretability, data quality, and fairness persist, limiting their widespread adoption across diverse educational environments. This review synthesizes findings from 58 studies published between 2019 and 2024, sourced from databases such as IEEE Xplore, Scopus, Science Direct, Google Scholar and SpringerLink. The literature selection was based on keywords like "student performance prediction" and "AI in education". Techniques, datasets, and evaluation metrics were categorized to identify thematic trends and highlight challenges and opportunities for ITS. The findings indicate that hybrid models (e.g., CNN-LSTM, RLCHI), ensemble learning, and integration of multimodal data significantly enhance predictive performance. Techniques like SHAP and LIME have improved interpretability, while models incorporating socio-demographic and behavioural data provide better insights into student learning patterns. Nonetheless, generalisability and fairness remain key challenges, requiring diverse datasets and fairness-aware modelling approaches. While ITS has made notable progress in personalised learning, further efforts are needed to enhance model generalisability, interpretability, and fairness. Addressing these challenges will be instrumental in developing inclusive, adaptive, and equitable learning environments, maximising the impact of ITS across diverse educational contexts.