Spiking Neural Network Based on Bidirectional Variational Anomaly Detection for Knowledge Tracing

Jinru Hu, Mingkun Chen, Yige Zhu, Jianrui Chen

Abstract: Knowledge Tracing is a core technology in educational data mining and intelligent education systems. It aims to predict future personal ability by analyzing the interaction history between learners and learning systems, and dynamically modeling the evolution process of their knowledge states. Despite significant progress in knowledge tracing research, there are still abnormal interactions at the data level, such as guessing, slipping and cheating, etc. These anomalies interfere with the model’s estimation of the real knowledge state, and reduce the accuracy of the prediction. To address these issues, we propose a spiking neural network based on bidirectional variational anomaly detection to enhance the accuracy of the model by interactive denoising (i.e., denosing abnormal data). Specifically, we design a bidirectional variational detector to the answer sequence of learners. The underlying distribution of normal interactions is learned by an encoder-decoder mechanism, and the reconstruction error is calculated to identify abnormal behaviors that deviate from the sequence distribution. Finally, the spiking neural network is introduced for temporal decision-making, and its event-driven characteristics are utilized to perform pulse coding on interactions after denoising. Extensive experiments on three real datasets show that the proposed method can dynamically denoise and improve the robustness of the knowledge tracing model, which is suitable for sparse and noisy data in real education scenarios.

External IDs:doi:10.1007/978-981-95-3643-6_10