Dual-Profile Modeling for Tobacco Sensory Evaluation via Graph Algorithms and Deep Learning

Sensory evaluation of tobacco products relies heavily on professional tasters, but results are often subjective and costly. Meanwhile, much consumer review data remains underutilized. To address these issues, this paper proposes a “dual-profile” framework integrating multi-dimensional attributes and textual data. For personnel management, a heterogeneous graph is constructed based on tasters’ behavior, skills, motivation, and personal attributes, and hierarchical clustering is performed using the Louvain algorithm. The approach achieves a modularity of 0.783 and reduces personnel misclassification by 15.8%. For product evaluation, a TobBERT tobacco-specific pre-trained model combined with a BiLSTM-Transformer architecture significantly outperforms baseline models in sentiment analysis (accuracy 89.6%) and entity recognition (F1 score 92.7%). Experimental results demonstrate that the system improves tasting task allocation efficiency and enhances objectivity in product quality evaluation, providing reliable technical support for intelligent management in the tobacco industry.