Aiming at the shortcomings of low-value density and inconsistency in online reviews of complex experiential products,a deep mining model of user preferences is proposed to support enterprises to improve existing products based on product reviews accurately. First，establish a bidirectional long and short-term memory neural network（BiLSTMNN） to fine-grained mining the user emotional polarity implicit in the rough comments. Secondly，in order to mine user preferences from inconsistent comments，a partial regression model is used to mine users’ perceptions of different attributes Linear preference. Finally，according to the trained partial regression model，the Kano model is applied to discover the non-linear preferences of users with various attributes. Taking the data of Shanghai Disneyland company as an example，the user preference deep mining model designed in this paper can mine the user’s non-linear preferences implied in complex experiential product reviews with high accuracy and then put forward suggestions for product improvement.