Teacher classroom teaching behavior indicators serve as a crucial foundation for guiding instructional evaluation. Existing indicator system suffers from limitations such as strong subjectivity and weak contextual generalization capabilities. Generalized category discovery (GCD) enables automatic data clustering to identify known categories and discover novel ones. Drawing inspiration from GCD mechanisms, this paper proposes a data-knowledge-driven framework for automatic category discovery of classroom teaching behavior indicators (DKD-TBICAD). The framework utilizes partially labeled data as constraint guidance and leverages extensive unlabeled data as pattern mining carriers to achieve automatic discovery and classification of teaching behavior categories. Specifically, the framework enhances spatiotemporal feature discriminability through supervised contrastive learning and spatiotemporal neighborhood aggregation contrastive learning. Additionally, we design a dynamic domain feature aggregation strategy to optimize the adaptability of feature learning, further enhancing the framework's capabilities in feature aggregation and novel class discovery. Experimental results on the proprietary TBU dataset and public UCF101 dataset demonstrate that the proposed method achieves 4% higher overall accuracy than baseline models. On UCF101, it surpasses baselines by 8.9% in old-class accuracy, while on the TBU dataset, it achieves 10% higher accuracy in new-class recognition. We believe this study provides valuable insights for indicator generation research driven by bidirectional integration of expert knowledge and data knowledge. [For the complete proceedings, see ED675583.]