AdvKT: An Adversarial Multi-Step Training Framework for Knowledge Tracing

TL;DR

AdvKT addresses the mismatch between common single-step KT training and the multi-step inference used in real-world ITS simulations, which causes error accumulation and data sparsity problems. It introduces an adversarial multi-step framework with a generator that predicts multi-step responses and a discriminator that guides data augmentation to produce realistic, diverse sequences, thereby aligning training with inference. The method combines KT-specific embeddings, GRU-based state modeling with attention, and specialized BCE/advantage losses plus autoregressive objectives, along with richly augmented and synthetic data to improve robustness. Experiments on four public KT datasets show AdvKT achieving state-of-the-art results in multi-step settings and notable improvements in data-sparse scenarios, demonstrating practical potential for student simulators and intelligent tutoring systems.

Abstract

Knowledge Tracing (KT) monitors students' knowledge states and simulates their responses to question sequences. Existing KT models typically follow a single-step training paradigm, which leads to discrepancies with the multi-step inference process required in real-world simulations, resulting in significant error accumulation. This accumulation of error, coupled with the issue of data sparsity, can substantially degrade the performance of recommendation models in the intelligent tutoring systems. To address these challenges, we propose a novel Adversarial Multi-Step Training Framework for Knowledge Tracing (AdvKT), which, for the first time, focuses on the multi-step KT task. More specifically, AdvKT leverages adversarial learning paradigm involving a generator and a discriminator. The generator mimics high-reward responses, effectively reducing error accumulation across multiple steps, while the discriminator provides feedback to generate synthetic data. Additionally, we design specialized data augmentation techniques to enrich the training data with realistic variations, ensuring that the model generalizes well even in scenarios with sparse data. Experiments conducted on four real-world datasets demonstrate the superiority of AdvKT over existing KT models, showcasing its ability to address both error accumulation and data sparsity issues effectively.

Figures (7)

• Figure 1: (a) Illustration of the students simulator's application and the comparison between single-step and multi-step inference KT. (b) Comparison of single-step inference and multi-step inference (simulator) and the primary application of the simulator. Input of single-step inference: historical question and ground-truth response. Input of multi-step inference: historical question and predicted response. (c) Performance comparison of the single-step inference and multi-step inference with a single-step training approach. With increased sequence length, errors accumulate gradually under the multi-step inference setting.
• Figure 2: The overview of the proposed AdvKT framework.
• Figure 3: Comparison results of the influence of learning sequence length of the generator with BCE loss and AdvKT.
• Figure 4: Distribution of question-response pair visualized with two-dimensional t-SNE on ASSIST09. Green points are real data in dataset. Blue points correspond to data generated by positive data augmentation methods. Red points represent generative data.
• Figure 5: Ablation study on the hyper-parameter $\gamma$ on four datasets.