A cross-domain knowledge tracing model based on graph optimal transport

Knowledge tracing (KT) is an essential technology in intelligent education that predicts students’ performance by analyzing their learning behavior data. Despite significant advances in KT, most existing research focuses on improving performance within a single domain (i.e., a specific academic discipline or school), with limited attention to Cross-Domain Knowledge Tracing (CDKT). CDKT aims to efficiently transfer the capabilities of a KT model from a source domain to a target domain. The main challenge in CDKT is aligning the knowledge state distributions between the two domains to compensate for the lack of interaction data in the target domain. To address this challenge, we propose an Auto-encoder Embedding and Graph Optimal Transport based CDKT model (AEGOT-CDKT). In the initial pre-training phase, we use an auto-encoder to learn the embedding representations of exercises and knowledge concepts in both the source and target domains. Next, we design a cross-domain alignment module that aligns the embedding representations of the two domains using the Graph Optimal Transport (GOT) strategy. This approach effectively supplements students’ knowledge state representations acquired from both domains and guides the learning process in the target domain. Additionally, intra-domain alignment is conducted through a graph decoding objective, bringing node features with similar knowledge concepts or exercises closer together. Finally, we train an LSTM model using the aligned source domain embeddings and use the trained LSTM to predict student responses in the target domain. Transfer experiments in different educational contexts demonstrate that our approach successfully accomplishes CDKT task and validates the effectiveness of aligning knowledge state distributions across different domains. Copyright © 2024 The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature.