Fig 1.
Overview of the GELT method.
Fig 2.
The graph embedding process in GELT.
When a student answers a question, GELT follows a series of steps. Firstly, it aggregates the skill features associated with the answered question (L1). Next, it updates the question features related to the skill (L2) and calculates the similarity for each skill node (L3). In addition, the model incorporates the difficulty information of the question as an attribute feature (L4). Lastly, the student’s knowledge states are updated through supervised learning using four loss functions.
Fig 3.
Framework of the attention mechanism.
Fig 4.
Computational graph for Lite-Transformer.
With low-dimensional binary queries and keys, Lite-Transformer replaces the majority of energy-intensive floating point multiplications in the conventional Transformer with simple additions. This substitution results in a significant reduction in computational complexity.
Table 1.
Datasets used for experiments.
Table 2.
The AUC results over three datasets.
Table 3.
The results of ablation experiments of three kinds of modules.
Fig 5.
Comparison of ACC and AUC among three methods on the ASSISTments2009 dataset.
Fig 6.
Comparison of training duration.
Fig 7.
The graph illustrates the correlation between skills, which is determined by the weights learned from associating the same questions.
By examining the graph, we can observe that it uncovers the implicit relationships between skill nodes, leading to a clear clustering effect.
Fig 8.
The heatmap representing the weighted values that indicate the relationships between different skills.