Extracting valuable information from online reviews and providing users with personalized product ranking is a new direction for e-commerce operations in the digital and intelligent era. Existing research results in biases in sentiment analysis of textual reviews, which may affect the accuracy of ranking results. Additionally, current recommendation models rarely consider the impact of user preference uncertainty on ranking outcomes. To fill these gaps, this study proposes a ranking model that combines deep learning with Bayesian preference disaggregation methods. First, to reduce sentiment classification bias for unlabeled reviews, a sentiment classification method based on a pre-trained large language model is proposed. Specifically, the BERTopic model is employed to extract product evaluation criteria and keywords that users are interested in from online reviews. The LoRA (low-rank adaptation) framework is utilized to fine tune the Chinese-RoBERTa sentiment classification pre-trained model based on labeled data. Monte Carlo simulation is used to convert criteria preferences into Bayesian prior information, constructing different types of likelihood functions to model both certain and uncertain product preference information. By combining the Bayesian rule and preference disaggregation method, the probability distribution for product ranking can be calculated, ultimately providing personalized recommendations to users. A case study of medical smartwatches recommendations on JD.com is conducted to verify the effectiveness of the proposed model.