DASKT: A Dynamic Affect Simulation Method for Knowledge Tracing

TL;DR

DASKT tackles the challenge that affective states influence knowledge tracing by deriving four affect-related factors from non-affect-oriented behavior and modeling their dynamic evolution with interval-based clustering and a dynamic affect trajectory graph. The framework fuses these affect signals with KT through a sequence model that uses affect-enhanced representations to predict future performance more accurately. Empirical results on ASSIST2012 and ASSISTchall show that DASKT consistently outperforms strong baselines across RMSE, ACC, AUC, and $r^2$, and ablation studies confirm the necessity of its MAF, ICA, and DASE components. By improving both the interpretability and accuracy of KT, DASKT offers a scalable path toward affect-aware adaptive learning and personalized interventions.

Abstract

Knowledge Tracing (KT) predicts future performance by modeling students' historical interactions, and understanding students' affective states can enhance the effectiveness of KT, thereby improving the quality of education. Although traditional KT values students' cognition and learning behaviors, efficient evaluation of students' affective states and their application in KT still require further exploration due to the non-affect-oriented nature of the data and budget constraints. To address this issue, we propose a computation-driven approach, Dynamic Affect Simulation Knowledge Tracing (DASKT), to explore the impact of various student affective states (such as frustration, concentration, boredom, and confusion) on their knowledge states. In this model, we first extract affective factors from students' non-affect-oriented behavioral data, then use clustering and spatiotemporal sequence modeling to accurately simulate students' dynamic affect changes when dealing with different problems. Subsequently, {\color{blue}we incorporate affect with time-series analysis to improve the model's ability to infer knowledge states over time and space.} Extensive experimental results on two public real-world educational datasets show that DASKT can achieve more reasonable knowledge states under the effect of students' affective states. Moreover, DASKT outperforms the most advanced KT methods in predicting student performance. Our research highlights a promising avenue for future KT studies, focusing on achieving high interpretability and accuracy.

Figures (5)

• Figure 1: A toy example illustrating the mastery level of three knowledge concepts in the same exercise sequence for two students who have identical cognitive abilities at the initial stage. Student $s1$ experiences affect of frustration and boredom throughout the learning process, while Student $s2$ consistently maintains a state of concentration.
• Figure 2: The main structure of our DASKT model is presented. The diagram illustrates the method of obtaining affective factors and the process of calculating affect at intervals. Based on this, at time step $t$, the inputs include affect $affect_t$, problem embedding $P_t$, knowledge concept of the problem $KC_t$, student's response $r_t$, and knowledge states $h_t$. In addition, we can also predict the student's performance at time step $t+1$, denoted as $y_{t+1}$.
• Figure 3: An illustrative example of interval-based affect calculation is shown where a student's sequence of 14 attempts is split into three interval segments, correlating to time intervals for answering 5 problems each. If the student ceases interaction, remaining segments are zero-filled. Attempt counts vary according to the number of problems addressed.
• Figure 4: This is a comparative diagram that illustrates the differences in a student's knowledge states when completing a sequence of nine exercises, with and without affect involvement. The knowledge states is updated every three exercises, with numerals representing the value of the knowledge states. A higher value indicates a better predicted mastery of the corresponding KC. The top half of the diagram showcases the knowledge states in the absence of affect, while the bottom half depicts the knowledge states when affect are factored in.
• Figure 5: This is an example demonstrated on the ASSISTchall dataset, revealing the impact of dynamic affective states on the changes in knowledge states and the performance of the DASKT model. We interpret future predictions by comparing the dynamic affective states in the learning sequences of two students. Accuracy of the pred describes the predictive accuracy of the DASKT model for the exercise problem at time t. Solid circle symbols indicate that the prediction matches the student's actual answer, while hollow circle symbols show that the prediction is inconsistent with the student's actual response.