审计 AI 辅导平台中日志衍生行为特征的来源：效度框架及其在错误后坚持中的应用

英文标题/原题

Auditing the Provenance of Log-Derived Behavioral Features in AI Tutoring Platforms: A Validity Framework and Its Application to Post-Error Persistence

作者

刘康

通讯作者

刘康

研究方向

学习分析；测量效度；AI 辅导平台

关键词

Learning analytics; Behavioral features; Interaction logs; Validity; AI tutoring platforms; Logging design

投稿/发表信息

发表期刊/目标

Education and Information Technologies

投稿/成果状态

投稿中

投稿日期

2026-10-23

发表日期

-

原地址/投稿系统地址

https://www.editorialmanager.com/eait/

期刊与检索情况

检索情况

SSCI、Scopus（未核实）

摘要/说明

Dashboards, learner profiles, and adaptive recommenders in AI tutoring platforms rely on behavioral features computed from interaction logs, such as persistence after an error, and present them as learner properties. Yet the platform that produced the log also decided which next action could occur. We present a feature-provenance audit: a four-step procedure and open-source toolkit. It classifies logged transitions by delivery regime, estimates within-learner associations by regime, and tests whether fixed-sequence label permutations reproduce each association. It also decomposes learner-level scores into regime exposure and residual behavior. In simulations with known ground truth, the audit recovered learner effects that conventional estimates missed when regime labels were correct, and it exposed its failure when labels were wrong or when the platform rerouted learners outside recorded regimes. Applied to 100,000 EdNet learners, the audit showed that 61% of post-error same-skill continuations occurred within item bundles answered and submitted together. Permutations reproduced the post-error association there. Repetition of the same question carried the only error-contingent signal, and about two thirds occurred in matched adaptive re-offers. Bundle exposure explained 75% of variance in a conventional persistence score, which did not improve retention prediction. The threshold-based warning had low sensitivity on held-out simulations. Counting bundles rather than responses as learning opportunities in a knowledge-tracing model changed learning-rate estimates while leaving predictive accuracy nearly unchanged. We derive reading rules for dashboard designers and a logging-design recommendation under which delivery regimes would be directly observable.

基金信息

无

论文备注

项目目录：E:\小作坊\01_论文\10_错误后伪坚持（P002）。EAIT 投稿包：05_投稿上传文件\3_期刊系统_EAIT\（title_page.pdf、manuscript.pdf、online_resource_1.pdf、cover_letter.pdf、manuscript_latex_source.zip、中英对照/）。数据与代码：OSF 匿名只读链接（view_only 5d3ft）。投稿历程：JLA 稿号 10065，2026-07-25 投、2026-08-22 初审拒稿（原题 When Logged Actions Are Partly Platform-Structured…）；IEEE TLT 稿号 TLT-2026-10-1160，2026-10-07 投、2026-10-12 以范围不符初审拒稿；2026-10-23 改投 EAIT。