一句话：发布首个区分记忆衰减与干扰的受控软件工程知识追踪数据集，并揭示DKT在干扰下性能下降。
现有知识追踪（KT）数据集无法区分两种遗忘机制：记忆衰减（知识随时间消退）与干扰（新学习破坏旧知识），阻碍了遗忘感知模型与教学评估。本文提出 FORGET-SE，首个受控遗忘数据集，来自 186 名大学生、10,873 次交互，覆盖 10 个软件工程知识点，每个知识点平均 5.85 次交互；通过策略性排序的每周测验评估记忆衰减与干扰效应。实验发现，最先进的深度知识追踪（DKT）模型在干扰下性能显著下降，为 DKT 改进提供新方向。最后讨论了该数据集超越传统 KT 问题的研究路径。
Forgetting is a fundamental aspect of human learning, yet existing Knowledge Tracing (KT) datasets fail to distinguish between two well-established forgetting mechanisms: memory decay (knowledge fades over time) and interference (new learning disrupts old knowledge). This limitation hinders researchers from developing accurate forgetting-aware models to trace students’ knowledge states and educators from understanding how teaching methodologies affect knowledge retention. This paper introduces FORGET-SE, the first controlled forgetting dataset collected from 186 university students (10,873 interactions) studying 10 software engineering knowledge components (KCs) (avg of 5.85 interactions per KC) through strategically sequenced weekly quizzes designed to assess memory decay and interference effects. State-of-the-art Deep Knowledge Tracing (DKT) models show significant performance degradation under interference, which opens new directions for advancing DKT. Finally, we outline research avenues that FORGET-SE can address beyond the traditional KT problem.
@inproceedings{paperbot11966,
title = {FORGET-SE: A Controlled Dataset to Investigate Memory Decay and Interference in Software Engineering Knowledge Tracing},
author = {Alyssa Sha and Bernardo Pereira Nunes and Yige Chen},
booktitle = {ICSE},
year = {2026},
doi = {10.1145/3786580.3786952}
}