An Explainable <scp>AI</scp> Framework for Game‐Based Assessment: From Model Tuning to Human Insights

ABSTRACT Background Games are one of the most popular activities that transcend cultures and ages. Game‐based assessment (GBA) integrates game elements into the assessment of abilities, skills, or knowledge and has already been applied in education. However, the complex behavioural sequence data of GBA poses a challenge to the explainability of artificial intelligence (AI)‐based models. Objectives The objective of this research is to construct a deep learning framework for game‐based educational assessment and transform model decisions into human‐understandable educational insights through explainable AI technology, ultimately achieving precise prediction and intervention support for learners' learning states. Methods This paper proposes a Mamba‐based explainable AI framework named MambaGBA for game‐based education assessment. MambaGBA employs a Mamba State Space Model as backbone and integrates a cognitive science‐inspired Episodic Memory Module to capture key behavioural patterns, aiming to predict learners' performance in GBA. Furthermore, an eXplainable Artificial Intelligence (XAI) analysis reveals MambaGBA model's decision‐making logic. Results and Conclusions Experimental results demonstrate that MambaGBA outperforms the baseline models in predicting student performance, including LightGBM, LSTM (Long Short‐Term Memory), and transformer. XAI helps to distill the complex knowledge learned by models into insights and tools that human experts can understand. This study also develops a lightweight detector for identifying learners' struggling states based on MambaGBAs' interpretable insights. This study not only provides a high‐performance and highly interpretable GBA framework but also offers a new theoretical perspective and practical evidence on how to apply XAI technology more meaningfully in education.