Sim-GAIL with SAINT for Knowledge Completion in Intelligent Tutoring Systems

¹ Department of Computers Techniques Engineering, College of Technical Engineering, The Islamic University, Najaf, Iraq
² Department of Information Science and Engineering, Dayananda Sagar Academy of Technology and Management, Bengaluru, India
³ Department of Electronics and Communication Engineering, Nitte Meenakshi Institute of Technology, Nitte (Deemed to be University), Bengaluru, India
⁴ Department of Business Informatics and Software Engineering, University of Technology, Mauritius
⁵ Department of Computer Science and Engineering, BMS College of Engineering, Bengaluru, India

^* Corresponding author: keerthi.kumar@nmit.ac.in

Abstract

In recent times, learner profiles and knowledge graphs in educational platforms have gaps due to missing interactions or unseen concepts; knowledge completion helps fill missing parts and understand learners. Conventional evaluation methods depend poorly on targeted questions, which is insignificant in uncovering a student’s true understanding, even though they demand long hours of testing. The continuous integration of Artificial Intelligence (AI) into digital education involves an advanced Intelligent Tutoring Systems (ITS), online course delivery, and learning management platforms. Traditional simulation methods fail to produce high-quality and diverse data, limiting their usefulness in training Intelligent Tutoring Systems (ITS). While ITS initiates learning paths, it faces challenges, such as cold-start issues, time demands, manual effort, and high costs. The proposed Sim-GAIL (Generative Adversarial Imitation Learning) integrates SAINT (Self-Attentive Neural Knowledge Tracing) to combine simulated student trajectory generation with attentive control prediction. The preprocessing step extracts raw interactions into trajectory elements based on actions and temporal patterns, enabling policy learning within a Markov Decision Process (MDP). Sim-GAIL generates realistic learning sequences to address cold-start limitations, whereas SAINT uses self-attention to trace evolving knowledge states for an accurate performance prediction. Experimental results show that the integrated Sim-GAIL with SAINT method performs better than the existing Item Response Theory (IRT) with the XGBOOST model for both the EdNet and ASSISTment17 datasets, achieving 98.42% and 98.71 for the Area Under the Curve (AUC) metric.