An Implementation-Friendly Model-Agnostic Approach for Process Data Analysis
提出一种通用框架，从计算机化问题解决测试的日志文件中提取内在特征并汇总为概率曲线，用于识别关键动作和揭示反应模式，首次将GRU和Transformer等深度学习模型应用于过程数据，经PISA真实数据验证具有鲁棒性和模型无关性。
Computer-based problem-solving assessments capture detailed human–computer interactive processes through log files, providing rich information for understanding response behaviors and analyzing complex problem-solving (CPS) abilities. However, existing analytical approaches often rely on item-dependent features or task-specific models, limiting their generalizability. This paper proposes a comprehensive analytical framework that extracts intrinsic features from process data and summarizes them into probability curves, which reflect students’ progression toward correct answers. These curves are successfully applied to downstream tasks including identifying key actions and uncovering response patterns. Within this framework, we also, for the first time, apply modern deep learning models like GRU and Transformer to process data, exploring the potential of neural networks in this domain. The framework is validated by using real data from multiple PISA items. It is demonstrated that the proposed probability curves are robust and model-agnostic statistics. This approach provides a generalized and scalable solution for analyzing diverse CPS items, offering improved insights into students’ problem-solving strategies and capabilities. A user-friendly code repository is open for public use.