Research On Product Recommendation Method Integrating Multiple Categories Text Information And User Preferences

Social and economic development has greatly improved people’s lives,and the people’s vision of material and cultural life has been continuously enhanced.At the same time,human beings have entered the era of big data.How to select the information that users are interested in from the massive data and recommend it to users has become the focus of both businesses and users.Therefore,the personalized recommendation system came into being.User reviews contain rich emotional information,which is less considered by traditional recommendation methods.Although some studies have introduced text processing methods,they only process a single text such as user comments,and cannot perform fine-grained sentiment classification of comments according to user preferences.These factors limit the accuracy of sentiment classification.By comprehensively utilizing multiple texts and user preferences in sentiment classification,more accurate sentiment polarity can be obtained.In addition,the existing recommendation methods do not make full use of the temporal features of user preferences.Temporal features are an important part of user preferences,and using them can make user preference modeling more accurate.At the same time,users’ hobbies and interests are constantly changing.This change is not only self-generated by users with time and experience,but also influenced by other users.Existing methods often only focus on modeling user preferences.However,it lacks the use of social relationships.Taking these factors into consideration can improve the accuracy of product recommendation.Aiming at the above problems,this thesis conducts product recommendation research,and proposes a product recommendation method integrating multiple categories text information and user preferences.The method consists of two parts: the first part is an aspect-level emotion classification method integrating multiple text information and attention mechanism.The function of this method is to classify the sentiment of user comments to extract the sentiment polarity of user opinions in the comment text;the second part is a product recommendation method based on multidimensional features,which makes use of the obtained sentiment polarity,and Combining the time factor,a more accurate user preference vector is obtained,and a more accurate recommendation result is obtained by comprehensively using the user preference and social relationship.The two parts are described below.（1）An aspect-level emotion classification method integrating multiple text information and attention mechanism.Aiming at the insufficient mining of user sentiment information in comment texts by existing sentiment classification methods,this thesis proposes a new sentiment classification method by comprehensively using multiple texts,attention mechanisms and user preferences.For a product,its introduction text is the most intuitive description of the product.Traditional sentiment classification methods often only deal with user comments,but this thesis comprehensively uses two different texts: product introduction and user reviews.First,slice the two kinds of texts to obtain multiple sub-sequences,and use the attention mechanism to make the two kinds of texts interact to obtain the textual representation that integrates the product introduction and user comments;Process to extract the hidden features of the text information more fully;finally,use the corresponding aspect processing modules to train different aspects involved in the text information,get the most interesting aspects of the user according to the user’s preference,and input the text feature vector into the The aspect processing module performs aspect-level sentiment polarity calculation,and finally obtains a more accurate sentiment classification result.（2）A product recommendation method based on multi-dimensional features.Aiming at the inaccurate modeling of user preferences and the lack of utilization of social relations in existing recommendation methods,this thesis proposes a new product recommendation method by comprehensively using emotional polarity,time factors and social relations.This thesis firstly obtains the user’s sentiment polarity for the product according to the user’s comments,and obtains the user’s preference for the product’s time attribute according to the user’s historical behavior;then,the two kinds of information are fused by adding two gating units to the GRU model.Then,the user interest evolution network is used to capture the change process of user interest,and then the product score based on user preference can be obtained according to the user interest vector;at the same time,the social relationship network is constructed according to the user’s comment behavior,and the product score based on the social relationship can be obtained by extracting the social relationship.By combining the two scores using an adaptive time function,the final product score can be obtained,which in turn can generate more accurate product recommendation results.Finally,the proposed method is experimentally verified,and the accuracy,precision,recall and F value are used as evaluation indicators.The experimental results show that the product recommendation method proposed in this thesis,which integrates multi-textual information and user preferences,performs better in various indicators than other comparison method.