教学交互具有多模态的复杂数据表现。构建基于多模态课堂数据的教学交互特征模型，对于指向教学交互的学习分析研究具有重要意义。本研究首先基于扎根理论，以国家教育资源公共服务平台中的多模态课堂数据为样本开展了分析、编码与验证，构建了包含3个一级维度、9个二级维度、24个三级指标在内的教学交互特征理论模型；而后，基于机器学习方法，提出了多模态课堂数据的特征提取与融合算法，制订了表征个体和群体交互特征的高维向量结构；最后，研究构建了以多模态数据意义上的高维特征向量为输入，以理论模型中“交互秩序的稳定性”、“交互内容的开放性”和“交互行为的活跃性”三个指标为输出的支持向量机回归模型，验证了教学交互特征理论模型中部分指标的可计算、可拟合性（准确率大于0.75），提出了实现教学交互特征计算的自动化流程。本研究将为基于伴随式教育大数据的自动化课堂交互分析与评价提供方法和工具基础。

The performance of teaching interaction in data is multimodal and complex. Constructing instructional interaction feature model based on multimodal classroom data is of great significance for the researches of learning analysis associated with teaching interaction. On the one hand, based on Grounded Theory, this research collected an amount of data in the National Public Service Platform of Educational Resources as samples to carry out analysis, coding and verification. As a result, a theoretical model of instructional interaction feature including 3 first-class dimensions, 9 second-class dimensions and 24 third-class indicators was established. On the other hand, based on Machine Learning, this research proposed a feature extraction and fusion method based on multimodal data, and formulated high-dimensional vector structures representing individual and group interaction features. Finally, this research constructed a SVR model with the high-dimensional multimodal feature vector as the input and three interaction feature indicators (Stability of the interaction order, Openness of the interaction content, Activity of the interaction behavior) in the theoretical model as the output, verified the computability and predictability of certain indicators (accuracy > 0.75), and put forward an automatic process of instructional interaction feature calculation. This research will provide a method and tool basis for automated classroom interaction analysis and evaluation based on real-time education big data.