Research On The Predictive Model Of Product Recommendation List Based On Interactive Learning Of Review Content

In the era of information explosion,recommender systems are widely studied and applied to discover user preferences.Reviews often reflect user semantic information and play an important role in modeling user preferences.However,most of the existing comment-based recommendation methods still have some problems,for example,when using user reviews to construct user characteristics,they do not notice the difference in the importance of different words for modeling user characteristics,and most recommendation models lack the influence of time factors on the recommendation model,which reduces the recommendation performance.In order to solve these problems,two novel models are proposed in this thesis,and the specific research content is as follows:In order to distinguish the importance of each comment vocabulary to the characteristics of users or items,construct the interaction between users and item features more accurately,and improve the accuracy of user-item rating prediction,this thesis proposes a scoring prediction model(FLTRS)based on the interaction between users and items.First,the model uses an interaction-based approach to model scoring prediction questions as text matching questions.In the embedding layer,multiple different embedding vectors are learned for each feature,and the cosine similarity formula of the modified cosine similarity calculation method is used to calculate the similarity between the word vectors of user reviews and comments received by the project,and learn the user’s long-term preference features.Secondly,the user preference matrix is reweighted using a compressed stimulus network,learning the different weights of each element in the preference matrix,and feeding the reweighted matrix into the CNN framework to obtain the output.Sequence recommendation is based on the traditional recommendation method,taking into account the time factor of user interaction with the item,and recommending the project that the user will interact with.The traditional recommendation model does not consider the change trend of multiple contextual information in the user behavior sequence,and fails to accurately construct the dynamic preference characteristics of users.To solve this problem,a sequence recommendation model(SRMFS)based on scoring information is proposed.Firstly,in order to prevent information leakage,a bidirectional model that can achieve the expected purpose is trained,the Cloze target is used for sequence recommendation,and the two-way deep self-attention mechanism is proposed to simulate the user’s behavior sequence,and the context information contained in the adjacent terms of the random mask item in the user behavior sequence is comprehensively processed to predict any mask item in the sequence.In the prediction stage,the mask target item is put at the end of the sequence,and after training,the probability of user interaction with the target item at the next moment is obtained,and then the prediction score of the target user is scored after integrating the long-term preference characteristics of the user learned in Chapter 3,and the recommended value of the item is obtained,and the recommendation list is formed in descending order according to the recommended value.In this thesis,a large number of experiments on FLTRS model and SRMFS model are carried out on two publicly available datasets,Amazon and Yelp,and the results show that the proposed model has great advantages over the latest model.