The comments of users before purchasing goods or services not only reflect their experience and feelings, but also imply their preferences and usage demands. Based on this, this article first uses the qualitative research paradigm to sort out six common categories of user needs and construct a comprehensive demand structure model; Further extract subjective emotions, satisfaction feedback, and types of needs involved in the comments, and introduce quantile regression method to analyze their intrinsic correlation with actual user behavior. Secondly, in response to the shortcomings of existing recommendation methods in capturing comment information, a Review aware Graph Recommendation System (RGRS) that integrates comment perception was designed. This model first uses a deep language representation model to semantically compress and vectorize the demand features reflected in comments, and then integrates them into the interaction network between users and products through graph neural mechanisms to form an accurate individual oriented recommendation strategy. The model was experimentally validated on a real public dataset, and the results demonstrated that, compared with mainstream algorithms such as LGCN, NGCF, and BPR, the proposed method achieved a 2–4% improvement in key indicators like Precision and Recall by incorporating user preference semantics from comments, highlighting its superior ability to model nuanced user demands and generate more personalized and accurate recommendations. This study not only expands the application dimensions of user demand mining in comment scenarios, but also provides new methodological support and theoretical basis for the construction of intelligent recommendation systems.