Modeling and analysis of teacher-student interaction in electronic commerce courses from a multidimensional interaction perspective

Teaching activities in e-commerce courses are increasingly supported by digital platforms and multimodal data records. Teacher–student interaction behavior is therefore characterized by high-dimensional, heterogeneous, and dynamically changing evidence, which provides a data basis for fine-grained classroom interaction analysis. This study proposes a multidimensional framework for modeling teacher–student interaction behavior by integrating temporal analysis and graph-based relationship representation across four heterogeneous data sources: video, audio, operation logs, and discussion texts. It uses timestamp alignment, window segmentation, and semantic coding to construct a four-dimensional feature matrix comprising visual, auditory, operational, and semantic dimensions. The model was evaluated on a dataset containing 120 class sessions from five e-commerce courses. The proposed model achieved an accuracy of 91.3%, a Macro-F1 score of 90.1%, an AUC of 95.2%, a sequence consistency score of 88.9%, and an anomaly detection recall of 86.7%. Compared with the single-mode CNN-LSTM baseline, the accuracy increased from 87.4 to 91.3%, corresponding to an improvement of 3.9 percentage points. Compared with the Graph Transformer baseline, the proposed model improved accuracy by 0.6 percentage points and anomaly detection recall by 0.9 percentage points. In real-world teaching scenarios, the model achieved an average system response latency of 41.2 ± 3.5 ms, an interaction chain complete recognition rate of 91.1 ± 2.2%, an anomaly segment localization accuracy of 88.1 ± 2.4%, a teacher feedback trigger effectiveness of 86.2 ± 2.5%, and a classroom conversation coverage of 93.7 ± 1.8%. These results indicate that multimodal fusion, teacher–student bipartite interaction graph constraints, and temporal dependency modeling jointly improve interaction recognition and abnormal segment detection. However, the relatively weaker performance in promotional campaign planning shows that the current 32-window sequence and teacher–student bipartite graph are less effective in open-ended discussions involving delayed responses, parallel peer interactions, and rapid topic transitions. Cross-course transfer therefore remains to be validated across different task structures, semesters, and institutions.