Learning states enhanced knowledge tracing: Simulating the diversity in real-world learning process

TL;DR

LSKT tackles two core gaps in ATT-DLKT: underutilization of real-world interaction diversity and neglect of evolving learning states. It introduces three IRT-inspired embedding granularities (LSKT-1PL/2PL/3PL), a Learning State Extraction module to capture multi-scale state changes, and a learning-state–enhanced knowledge state extractor with sparse attention and clustering to fuse learning and knowledge states for prediction. Across four real-world datasets, LSKT and particularly the LS KT-3PL variant outperform baselines, with ablations confirming the necessity of both learning-state modeling and knowledge-state extraction, and visualizations demonstrating clearer, more nuanced state representations. The approach yields more accurate predictions and interpretable, fine-grained representations of the learner’s knowledge and learning progression, suggesting meaningful gains for adaptive educational systems and real-time tutoring.

Abstract

The Knowledge Tracing (KT) task focuses on predicting a learner's future performance based on the historical interactions. The knowledge state plays a key role in learning process. However, considering that the knowledge state is influenced by various learning factors in the interaction process, such as the exercises similarities, responses reliability and the learner's learning state. Previous models still face two major limitations. First, due to the exercises differences caused by various complex reasons and the unreliability of responses caused by guessing behavior, it is hard to locate the historical interaction which is most relevant to the current answered exercise. Second, the learning state is also a key factor to influence the knowledge state, which is always ignored by previous methods. To address these issues, we propose a new method named Learning State Enhanced Knowledge Tracing (LSKT). Firstly, to simulate the potential differences in interactions, inspired by Item Response Theory~(IRT) paradigm, we designed three different embedding methods ranging from coarse-grained to fine-grained views and conduct comparative analysis on them. Secondly, we design a learning state extraction module to capture the changing learning state during the learning process of the learner. In turn, with the help of the extracted learning state, a more detailed knowledge state could be captured. Experimental results on four real-world datasets show that our LSKT method outperforms the current state-of-the-art methods.

Figures (7)

• Figure 1: Figure (a) illustrates the learning process of a learner. As shown in the figure, during the process of answering these six exercises, the learner is influenced by many factors. We need to observe the learner's performance in $e_{1}$ to $e_{5}$, infer their latent knowledge state and learning state, and consider both aspects to predict the learner's performance in exercise $e_{6}$. Figure (b) shows a real slice of the ASSIST12 dataset. It demonstrates that recent learning states have an impact on subsequent performance, even though different concepts are involved, which should be taken into account.
• Figure 2: (a) The overall architecture of our LSKT. (b) The learning state extraction module, where we extract and integrate multi-scale learning state information through multi-layer causal convolutions. (c) The learning state enhanced knowledge state extraction module
• Figure 3: Comparison of ablation experiment results on three datasets. Different colors are used to distinguish between different ablation models, and the specific experimental results are labeled at the top of each bar chart.
• Figure 4: The influence of different clustering quantities on the AUC values of three embedding models. The horizontal axis shows the change in k values from 1 to 10, while the vertical axis displays the percentage of AUC values. Three different colors distinguish the three embedding methods in the graph.
• Figure 5: Visualization of Knowledge State and Learning State. Figure (a) represents the distribution of knowledge state features extracted without the guidance of a learning state. Figure (b) illustrates the feature distribution corresponding to the learning state. Figure (c) depicts the distribution of knowledge state features extracted with the guidance of a learning state.