In a product optimization design strategy, it is very important to carry out product attention mining and prediction of user satisfaction change law based on user preference evaluation information. Most of the existing research adopts a scale-based semi-structured data analysis model, which ignores the multi-dimensional nonlinear decision-making attributes of the evaluation process, especially the coupling problem between differential kansei images. In this regard, a heterogeneous emotional theme clustering simulation process is developed for unstructured data, and the web crawler is first used to obtain the user's online comment text. Then, with the help of Word2vec, the text is numerically encoded, and the sentiment binary classification judgment is completed by the sentiment analysis model, and the positive and negative sentiment datasets are established. Then, the BTM is used to carry out heterogeneous emotional theme clustering. Finally, the quantitative results of multi-dimensional evaluation indicators are output. The results show that the constructed simulation process can accurately judge dichotomous (the reliability test is greater than 0.85), and the topic clustering results are in line with the product optimization strategy.