Abstract
Knowledge tracing (KT) is the problem of predicting students' future performance based on their historical interactions with intelligent tutoring systems. Recently, many works present lots of special methods for applying deep neural networks to KT from different perspectives like model architecture, adversarial augmentation and etc., which make the overall algorithm and system become more and more complex. Furthermore, due to the lack of standardized evaluation protocol, there is no widely agreed KT baselines and published experimental comparisons become inconsistent and self-contradictory. Therefore, in this paper, we provide a strong but simple baseline method to deal with the KT task named simpleKT. Inspired by the Rasch model in psychometrics, we explicitly model question-specific variations to capture the individual differences among questions covering the same set of knowledge components. Furthermore, instead of using sophisticated representations to capture student forgetting behaviors, we use the ordinary dot-product attention function to extract the time-aware information embedded in the student learning interactions. Extensive experiments show that such a simple baseline is able to always rank top 3 in terms of AUC scores and achieve 57 wins, 3 ties and 16 loss against 12 DLKT baseline methods on 7 public datasets of different domains.
Introduction
Knowledge Tracing (KT) is a fundamental task in the field of intelligent tutoring systems, aimed at modeling a student's evolving knowledge state to predict their future performance. Despite the proliferation of complex Deep Learning-based Knowledge Tracing (DLKT) models—incorporating graph neural networks, adversarial training, and intricate forgetting mechanisms—the field has struggled with inconsistent evaluation protocols and self-contradictory results.
The authors argue that the current trajectory toward complexity lacks a robust, standardized baseline. Reported AUC scores for established models like DKT vary significantly across literature, making it difficult to ascertain true progress. To address this, we propose simpleKT, a streamlined yet powerful baseline that achieves state-of-the-art results through two primary refinements: question-specific modeling and standard attention mechanisms.
Methodology
The simpleKT model focuses on capturing individual differences among questions and leveraging standard attention to model temporal dynamics.
Question-Specific Modeling
Most KT models aggregate questions by their "Knowledge Components" (KCs)—the underlying skills or concepts required. However, questions within the same KC often vary in difficulty. Inspired by the Rasch model from psychometrics, simpleKT represents a question's embedding as the sum of its average KC embedding and a unique, question-specific variation vector. This allows the model to differentiate between an easy and a hard question covering the same concept.
KC-Level Transformation
To handle datasets where questions may map to multiple KCs, simpleKT employs a data transformation technique. This process expands question-level interactions into KC-level interactions, ensuring the model can learn from every relevant concept involved in a task.
Dot-Product Attention
Rather than employing complex, hand-crafted forgetting functions, simpleKT utilizes the standard dot-product attention mechanism found in Transformers. This approach allows the model to naturally learn the relevance of past interactions to the current prediction task, effectively capturing student learning and forgetting patterns without unnecessary architectural complexity.
Datasets and Evaluation
The model was rigorously evaluated on seven public datasets using the pyKT benchmark to ensure fair and reproducible comparisons against 12 baseline models. The datasets were categorized into two groups:
- D1 (Question + KC info): ASSISTments2009, Algebra2005, Bridge2006, and NIPS34.
- D2 (Question OR KC info): Statics2011, ASSISTments2015, and POJ.
Key Results
- Performance: simpleKT achieved 57 wins, 3 ties, and 16 losses against 12 deep learning baselines. It consistently ranked in the top 3 for AUC scores across all datasets.
- Ablation Study: Removing the question-specific difficulty component (the "NoDiff" variant) resulted in a performance drop of over 2% on D1 datasets, confirming that modeling individual question variations is critical for success.
- Robustness: The model's simplicity translates to high reproducibility and competitive performance even against models with significantly more parameters.
Discussion and Conclusion
While simpleKT is highly effective, its reliance on question-specific modeling means its advantages are most pronounced in datasets containing distinct information for both questions and KCs. Future work could explore more complex interactions between student ability and question difficulty, or integrate auxiliary features like response time and textual content.
In summary, simpleKT proves that simplicity is a powerful tool in Knowledge Tracing. It offers a robust baseline that the community can use to reliably measure the contributions of more complex architectural innovations.
Summary updated