科学探究是科学教育最重要的理念和目标之一，也是科学素养的重要组成部分及培 养方式。随着计算机与信息技术的发展，计算机交互测验成为科学探究评价的重要形式。 学生完成计算机交互探究试题的过程中会产生大量行为过程数据，为分析学生的科学探 究行为过程、策略等以往较难考察的方面提供新的研究思路。

本研究基于我国 320 名小学四年级学生作答一道计算机交互探究试题产生的过程数 据，对学生完成交互探究任务过程中的典型行为模式、行为特点进行考察，并进一步分 析行为模式、行为特点等过程特征对学生交互探究任务表现的预测作用。具体而言，本 研究关注的问题包括：(1)学生在交互探究任务中表现出哪些典型的行为模式(具有显著先 后顺序关系的两个行为的组合)及行为特点？(2)不同性别、不同表现学生的典型行为模式 与行为特点是否存在差异？(3)哪些探究行为模式及行为特点可以有效预测学生的交互探 究任务表现？

为了回答上述问题，本研究综合运用行为序列分析、可视化分析、有监督的机器学 习分类预测模型等教育数据挖掘方法分析学生的作答过程数据。首先，从任务交互行为 层面(即学生在任务系统中的交互操作行为)、实验设计行为层面(即学生改变实验变量水 平并运行实验条件的操作行为)分别提取相应的过程数据，并把过程数据重新编码成按照 时间排列的行为序列。研究一采用滞后序列分析、可视化分析的方法考察学生任务交互 行为层面的典型行为模式，并比较不同性别、不同表现学生的行为模式差异。其中，不 同表现学生的划分采用任务表现排名前、后 27%为临界值把学生分成三组，并分别命名 为高、中、低表现组。研究二采用滞后序列分析、可视化分析的方法考察学生实验设计 行为层面的典型行为模式，并比较不同性别、不同表现学生的行为模式差异。此外，研 究二还以学生的实验设计行为模式为依据提取控制变量策略的行为指标。研究三以学生 的典型行为模式、控制变量策略行为指标作为过程特征，以学生交互探究表现的二分变 量(前 27%编码为 1，后 27%编码为 0)为预测变量，采用随机森林模型预测学生的交互探 究表现，并考察过程特征的相对重要性。

研究一、二的结果表明：(1)学生在任务交互行为层面和实验设计行为层面都表现出 典型的行为模式。例如研究一中，学生在任务交互行为层面的典型行为模式包括改变变 量并运行实验条件、作答题目等行为模式；而研究二中，学生在实验设计行为层面的典 型行为模式包含一次只改变一个变量水平的行为模式(即控制变量策略)。(2)男、女生在两个行为层面上的典型行为模式差异相对较小，说明不同性别的学生在本研究采用的交互 探究任务上表现出相似的探究行为模式。(3)不同表现组的学生在两个行为层面上的典型 行为模式差异均较大，尤其是高、低表现组。例如，研究一中低表现组学生有进入任务 页面后直接答题的典型行为模式，而高表现组学生这一行为模式出现较少；研究二中高 表现组学生表现出系统的控制变量策略行为模式，而低表现组学生控制变量策略行为模 式较为零散，缺乏系统性。不同表现学生行为模式的差异可能反映出两组学生在认知(例 如控制变量策略的使用)和元认知(例如动机水平)等方面的差异。

Scientific inquiry has been one of the most important concepts and goals of the science education, as well as an important aspect of scientific literacy development. With the advance of computer and information technology, interactive computer assessment has become an important way of evaluating scientific inquiry. In the process of completing an interactive computer inquiry assessment, students will generate large amounts of behavioral process data, providing new research perspectives for analyzing students' scientific inquiry performance, processes, strategies, and other aspects that previously have been difficult to assess.

Based on the process data generated by 320 fourth-grade primary school students completing an interactive computer inquiry task, this study analyzed the typical behavior patterns and behavioral characteristics of the students in completing the inquiry task, and further examined the predictive effects of these behavior patterns and behavioral characteristics on the inquiry task performance. Specifically, this study is concerned with three questions: (1) What are the typical behavioral patterns and behavioral characteristics that students exhibit in the interactive inquiry task? (2) Are there differences in the typical behavioral patterns and characteristics of students of different genders and performance groups? (3) What inquiry behavior patterns and behavioral characteristics can effectively predict student performance in the interactive inquiry task?

To answer the above questions, this study integrates educational data mining techniques such as behavioral sequence analysis, visualization, and supervised machine learning model to analyze the process data generated during the interactive computer inquiry task. First, the process data were extracted from the overall interactive behaviors and the experimental design behaviors respectively, and the corresponding process data were then recoded into chronologically-ordered behavioral sequences. Study 1 used lag sequential analysis and visualization to examine typical behavior patterns for task interactive behaviors, and to further compare differences in behavior patterns among students of different genders and performance groups. Students were divided into three groups according to threshold of top 27% named as high performance group, bottom 27% named as low performance group, and the remaining named as medium performance group. Study 2 used lag sequential analysis and visualization to examine typical behavioral patterns for experimental design behaviors. Similarly, the behavioral patterns of students in different gender and performance groups were compared. Study 3 used the typical behavior pattern and the behavior indicators of control of variables strategy as process features, the dichotomous variables of students' interactive inquiry performance (top 27% coded as 1 and bottom 27% coded as 0) as predictor variable, and then used a random forest model to predict students' interactive inquiry performance, and also examined the relative importance of process characteristics.

The results of study 1 and 2 indicate that: (1) students exhibit typical behavior patterns for both the task interactive behaviors and the experiment design behaviors. For example, in study 1, students' typical behavioral patterns for the task interactive behaviors included behavioral patterns of changing variables and applying the data, and behavioral patterns of answering questions; whereas in study 2, students' typical behavioral patterns for the experimental design included varied only one variable at a time (i.e., control of variables strategy). (2) Only negligible differences were detected in the typical behavioral patterns of male students and female students, indicating that in general, the inquiry behavioral processes of students of different genders were similar for the interactive inquiry tasks used in this study; (3) more salient differences were found in the inquiry behavioral patterns of students from different performance groups, especially the high and low performance groups. For example, in study 1, students in the low performance group exhibited a typical behavior pattern in answering questions directly after entering the task, while students in the high performance group rarely exhibited this behavior pattern; in study 2, students in high performance group exhibited systematic use of control of variables strategy, while students in the low performance group exhibited far less systematic use of control of variables strategy. Differences in behavioral patterns for students from different performance groups may reflect differences in their cognition (e.g., control variable strategies) and metacognition (e.g., motivational levels).

The results of study 3 showed that the random forest model had an accuracy rate of 94.23%, an accuracy rate of 94.83%, a recall rate of 94.23% and an F1 score of 94.21 % on the test data. The process features overall had good predictive effects on students' performance in the interactive inquiry task. Control of variables strategy, experimental implementation behavior patterns, and initial interactive behavior patterns contributed the most to the random forest model.

This study is an exploratory study using process data to analyze scientific inquiry processes, inquiry strategies, and to predict student performance on interactive inquiry tasks using process features. The overall results of the three studies demonstrate the feasibility of using process  data to analyze the characteristics of students' interactive scientific inquiry processes and of using process data combined with machine learning models to predict scientific inquiry performance. The findings of this study have some insights for research in the field of educational process data analysis and scientific inquiry. In addition, the results of this study can provide guiding evidence for teaching. Teachers can develop targeted inquiry learning materials for students based on the typical inquiry behavior characteristics to help them learn more efficiently.