Enhancing Predictive Fairness in Deep Knowledge Tracing with Sequence Inverted Data Augmentation

Ensuring fairness in predictive models is a critical challenge, particularly in sensitive domains such as education. This paper addresses the issue of predictive bias in Deep Knowledge Tracing (DKT), a widely used model in Intelligent Tutoring Systems (ITS) for modeling student learning and predicting future performance. While DKT and its successors have advanced the state-of-the-art in personalized learning, they remain vulnerable to biases that emerge due to distribution shifts in student data, often occurring at the end of academic cycles or semesters. In response to this challenge, we introduce a novel sequence-inverted data augmentation method that significantly enhances both predictive accuracy and fairness in DKT. Our approach generates synthetic student sequences representing diverse performance extremes, thereby bolstering the model’s robustness against distribution shifts and reducing predictive bias across gender groups.Through extensive experiments on a real-world, large-scale student dataset, we show that our method significantly outperforms the previously state-of-the-art Balanced-3 technique, which, despite its prominence in bias mitigation for educational contexts, proves ineffective in sequential prediction tasks like DKT. Unlike sampling-based methods, our proposed method achieves superior predictive fairness without sacrificing accuracy, maintaining stable performance even in the presence of severe distribution shifts.Our work is the first to introduce a bias mitigation approach tailored for KT models and to thoroughly evaluate the issue of predictive fairness using real student data. Previous research often addresses predictive fairness in a broad and generalized manner, overlooking the various factors that contribute to bias. In contrast, this paper delves into a specific and realistic factor influencing predictive fairness — distribution shifts in test data. This targeted approach discussed in this paper not only refines our understanding of predictive fairness in the context of KT but also establishes a new perspective on analyzing and mitigating bias. This targeted approach paves the way for more equitable AI solutions in education and offers valuable insights for other domains involving sequential prediction, such as time series analysis.