Data-Driven Analysis of Instructional Behaviors: Multi-Feature Fusion Clustering Based on Embedded Instructional Behavior Texts

As the demand for assessing teaching quality increases, elaborate analysis of instructional behaviors has become crucial for improving teaching skills. Traditional methods, based on observation and manual coding, are not only time-consuming and labor-intensive but also subject to bias. In response, we focus on a data-driven analysis of instructional behaviors, and propose a text clustering-based framework to visualize various instructional behaviors, aiming to reduce the labeling workload and discover unnoticed yet valuable instructional behaviors. The framework integrates three core components: (1) content masking to filter out knowledge unrelated to instructional behaviors, (2) an embedding model fine-tuned with pedagogical knowledge, and (3) multi-feature fusion clustering for behavior analysis. Moreover, we have built a novel instructional behavior dataset from online teaching videos to evaluate the clustering results. Experimental results demonstrate that the framework effectively clusters instructional behaviors, providing intuitive data visualizations that facilitate educational researchers in mining and formulating behavior labels more suited to current teaching practices.