随着信息技术的发展与学习信息化的普及，在线学习迅速兴起并成为一种重要的学习方式。已有的研究表明，在线学习场景中学习者更容易出现不良的投入度，并因此产生较差的学习效果。为了改善在线学习场景下学习者的投入度，教学实践中出现了要求学习者与工具持续进行交互才能完成学习的交互式学习系统。该系统的日志数据中记录了学习者交互发生的具体时间，可以从中提取出反映学习者投入度的反应时数据进行分析。本研究获取了某公司交互式学习系统的日志数据，基于该数据对学习者的投入度状态进行分析，并对不同的投入度状态与学习者特征之间的关系进行探索。

本研究使用对数正态反应时模型（Log-Normal Respond Time Model）对学习者在单个内容页上的反应时数据进行建模，提取模型残差作为学习者投入度状态的特征，并对该特征构成的序列使用堆栈自编码器（Stacked Autoencoder）进行降维。根据降维后的学习者投入度状态特征在二维空间上的分布特点，本研究将密度聚类（Density-Based Spatial Clustering）与k均值聚类（K-Means Clustering）进行组合，对学习者进行了聚类分析。最后使用随机森林模型（Random Forest）特征重要性分析方法完成不同聚类类别的投入度状态定义。上述的建模与分析的结果表明，在本研究中，学习者存在四种投入度状态，分别为“专注”、“良好”、“一般”、“异常”。后续对学习者的投入度状态与其个人特征进一步分析发现，具有位于一线城市、月收入低于8000元、非学生三项特征的学习者更可能在本研究选取的课程上出现好的投入度。

本研究通过反应时数据对在线交互学习情境中学习者的投入度状态进行分析，并进一步分析了学习者特征与投入度的关系，希望能为学习投入度相关的理论研究提供案例参考。相比已有投入度评测研究所使用的问卷调研等方法，本研究所使用的分析方法具有成本低、信度高、时效快的特点，可应用与在线交互学习情境下的学情实时分析，具有一定的实践价值。

With the development of information technology and the popularization of learning informatization, online learning has risen rapidly and has become an important learning method. Existing studies have shown that learners in online learning scenarios are more likely to have poor engagement, resulting in poorer learning outcomes. In order to improve learners' engagement in online learning scenarios, interactive learning systems have emerged in teaching practice that require learners to continuously interact with tools to complete learning. The log data of the system records the specific time of the learner's interaction, from which the reaction-time data reflecting the learner's engagement can be extracted for analysis. In this study, log data of a company's interactive learning system was obtained, and based on the data, the state of learners' engagement was analyzed, and the relationship between different engagement states and learner characteristics was explored.

This study uses a log-normal reaction time model to model learners' reaction time data on a single content page, extracts model residuals as a feature of learner engagement status, and uses stack autoencoding for the sequence of features. dimensionality reduction. According to the distribution characteristics of learner engagement state features in two-dimensional space after dimensionality reduction, this study combines density clustering and k-means clustering to conduct cluster analysis on learners. Finally, the feature importance analysis method of the random forest model is used to complete the definition of the engagement state of different cluster categories. The results of the above modeling and analysis show that in this study, learners have four states of engagement, namely "focused", "good", "average", and "abnormal". Subsequent analysis of learners’ engagement status and their personal characteristics found that learners with three characteristics, located in first-tier cities, with a monthly income of less than 8,000 yuan, and non-students, were more likely to show good engagement in the courses selected in this study.

In this study, the state of learners' engagement in online interactive learning situations was analyzed through reaction time data, and the relationship between learner characteristics and engagement was further analyzed, hoping to provide case references for theoretical research on learning engagement. Compared with the questionnaire survey and other methods used in the existing engagement evaluation research, the analysis method used in this research has the characteristics of low cost, high reliability and fast timeliness. It has certain practical value.