Deep Learning–Based Prediction of Public Opinion Dynamics in a Power Bank Product Safety Incident

Product safety incidents on social media tend to rapidly evolve into complex, multidimensional public opinion dissemination processes, characterized by fine-grained semantics, strong inter-label correlations, and dynamic evolution. Existing methods are limited in their ability to capture fine-grained semantic features and model label structural relationships, thereby constraining the accuracy and interpretability of public opinion evolution analysis. To address these challenges, this study takes the Weibo data from a power bank safety incident involving Romoss as the research object, constructs a manually annotated multidimensional semantic dataset, and proposes a deep learning model, HPLR-Net, which integrates hierarchical semantic modeling with label relationship constraints to achieve fine-grained semantic recognition. Building upon this, an OFF semantic monitoring mechanism is introduced, and a CPORI risk index is developed to transform semantic recognition outputs into daily aggregated indicators for dynamic characterization and trend prediction of public opinion evolution. Experimental results demonstrate that HPLR-Net outperforms multiple baseline models in primary label semantic recognition tasks, showing notable advantages in handling difficult labels and maintaining label structural consistency. Furthermore, the CPORI risk index effectively distinguishes different stages of public opinion development and reveals the coupling dynamics of multidimensional semantics throughout the evolution process. The findings indicate that the proposed approach enables fine-grained identification and dynamic prediction of public opinion in product safety incidents, balancing model performance and interpretability, and providing effective technical support for regulatory risk warning and enterprise crisis response.