Sunday 06 April 2025
Artificial intelligence has long been touted as a panacea for personalized education, promising to tailor learning experiences to individual students’ needs and abilities. But what happens when those algorithms are biased? Researchers have long known that AI systems can perpetuate and even amplify existing social inequalities, but a new study sheds light on how this plays out in the context of intelligent tutoring systems.
The authors of the paper propose a novel approach called Disentangled Knowledge Tracing (DisKT), designed to alleviate cognitive bias in these systems. Cognitive bias refers to the tendency for AI models to favor certain types of students or responses over others, often based on flawed assumptions about human learning and behavior. In the context of intelligent tutoring systems, this can lead to inaccurate predictions and suboptimal learning experiences.
To understand how DisKT works, it’s helpful to consider the basic architecture of these systems. Intelligent tutoring systems use a combination of machine learning algorithms and item response theory (IRT) to model student knowledge states over time. IRT is a statistical method that estimates an individual’s ability based on their responses to a set of questions. In this context, the goal is to predict which questions a student will answer correctly or incorrectly.
DisKT builds upon existing work in this area by introducing a novel causal graph that models the relationships between student knowledge states, question difficulty, and response outcomes. This allows the system to disentangle the complex interactions between these variables and better understand how they impact learning outcomes.
The authors demonstrate the effectiveness of DisKT using 11 publicly available datasets from various educational domains. They compare their approach to 16 state-of-the-art models, showing that DisKT outperforms them in terms of evaluation accuracy while also reducing cognitive bias.
One key insight from this study is the importance of considering the contradictory psychology of mistaking and guessing. In traditional IRT approaches, a student’s incorrect response can be attributed solely to their lack of knowledge or ability. However, DisKT recognizes that students may also make mistakes due to factors such as overconfidence or distraction. By accounting for these additional variables, the system is better able to distinguish between true knowledge gaps and temporary lapses in attention.
The implications of this work are significant, particularly in the context of personalized education. By developing AI systems that are more equitable and transparent, we can help ensure that all students have access to high-quality learning experiences, regardless of their background or abilities.
Cite this article: “Disentangling Cognitive Bias in Knowledge Tracing: A Novel IRT-variant Approach”, The Science Archive, 2025.
Artificial Intelligence, Personalized Education, Cognitive Bias, Intelligent Tutoring Systems, Disentangled Knowledge Tracing, Item Response Theory, Machine Learning Algorithms, Educational Domains, Evaluation Accuracy, Contradictory Psychology
