Improving the Efficiency of Knowledge Tracing Algorithms: LinSAKT
Izhan Ali
23/02/2026
Knowledge sharing through teaching is an essential aspect of human intelligence. With the advent of online teaching platforms, teachers have the ability to track student progress and personalize teaching to tailor their learning experience. Predicting a student’s performance by tracking their learning progress and interactions with the study material is called the Knowledge tracing (KT) problem. The KT problem in general has found usage in intelligent tutoring systems (ITS), curriculum learning and other computer-aided learning tools. Due to the availability of massive datasets, KT problem is effectively solved using various deep learning algorithms. Each of these algorithms have exploited the deep KT (DKT) problem through memory-augmented neural networks, attention mechanisms, graph-based networks and feature engineering techniques. These techniques drastically improve the accuracy of these algorithms making them more suited for various teaching related tasks. In this research, we focus on one of the well-known DKT algorithms called SAKT (attention-based framework) and study its computational efficiency and accuracy. Attention based frameworks (SAKT) have higher accuracy, incur higher memory footprint and greater runtime. To take advantage of their accuracy but lower the runtime and memory usage, we propose a linear attention model for the KT task called LinSAKT. This model preserves the contextual information captured by the attention mechanism through Linear attention and significantly reduces the runtime and memory footprint. This model is a relevant addition to solving the KT problem while enhancing efficiency. Three datasets were used to study efficiency compared to the baseline SAKT. LinSAKT reduced training time and memory usage. Averaged across three ASSIST datasets and all tested hyperparameter configurations (36 paired comparisons total), LinSAKT reduces training time per epoch by 14.1% and peak memory per epoch by 30.6% (unweighted mean of per-configuration percent reductions). LinSAKT enables KT in low-resource environments supporting equitable access to AI-driven personalized learning tools in underserved regions.