Abstract: With the rapid development of smart connected products, the vast volume of user-generated online reviews has emerged as a vital source for understanding user preferences and tracking market demand dynamics. To uncover the evolution of user opinions from high-dimensional, heterogeneous, and temporally structured reviews, this study develops a multimodal analytical model capable of dynamically capturing user needs and opinion evolution, thereby informing product iteration and enhancement. Specifically, the study first integrates textual and visual features of online reviews to construct the initial nodes of the review network using spectral clustering. Subsequently, a dynamic DeGroot-LSTM model is introduced by embedding the temporal memory structure of long short-term memory (LSTM) networks and a weight-learning mechanism into the DeGroot model, enabling adaptive updates of influence relationships among nodes and capturing the nonlinear dynamics of opinion propagation. By analyzing node opinions and weight matrices, the model enables prediction of user demand and opinion evolution trends across different product features. A case study on the Echo Dot smart speaker demonstrates that the proposed model achieves average MSE, MAE, and R² values of 0.0032, 0.0276, and 0.8295, respectively, across all features, significantly outperforming the static DeGroot model and the linear Ridge regression model. The findings not only provide data-driven decision support for the optimization of smart connected products but also offer a novel methodological framework for studying opinion evolution from a multimodal online review perspective.