Research On Recommendation Model Based On Enhanced Comments

With the rapid development of Internet technology in recent years,we have taken steps to the era of big data.Useful business information can be obtained according to users' past behavior records,and users' preferences for certain items can be predicted through analysis,so that users can quickly find items that they are interested in,so recommendation technologies urgently need to be studied.Based on the traditional collaborative filtering recommendation system only uses the information in the user's item rating matrix to make recommendations after analyzing users or items.Although it is simple and convenient,there are many problems.Although many scholars put forward a lot of ways to try to solve it,the effect is still not ideal so far.This paper studies the two most serious defects of the traditional collaborative filtering recommendation system and tries to improve them,which are excessively sparse data and cold start of the recommendation system respectively.Too sparse data refers to the number of items in a user item rating matrix is excessively sparse,and the cold start means that the recommendation system is often faced with new users and new items,as well as the cold start of the system,but the traditional recommendation system algorithm is difficult to deal with,cause the result deviation,even difficult to use or need more long time to get the recommendation result,in recent years,the score predicts recommendation algorithm based on the comment text is getting more and more researchers' attention.The comment text covers a large number of users' usage information and personal preferences,etc.In-depth mining of this information can grasp users' interests,hobbies,article characteristics more accurately,and make accurate recommendations to users under the support of this information.This project adopts a text-based approach to analyze the comment and forecast problems.The following are the main tasks completed:(1)A recommendation model based on enhanced comment graph is proposed.This model is divided into 4 stages,namely,item embedding,deep processing knowledge,construction features and recommendation prediction.The model needs to embed the item identification in the text comment to form the enhanced comment text;Then,the Word2 vec model is used to analyze and extract the semantic features of enhanced comment text.The important keywords are extracted based on the graph model,and the semantic relations among the comment keywords are established through the knowledge graph,so as to expand the users' interests and preferences.The feature of network nodes extracted from the knowledge map is fused with the semantic features of comments,which is used to represent the feature vector of items and users.After the completion of feature construction,we enter the stage of recommendation and prediction,and use Wide&Deep model to predict the final recommendation results.A two-level attention recommendation algorithm model based on enhanced comment crossover is proposed.This model is composed of two network models,which fuse the user's hidden feature vector and the item's hidden feature vector respectively.In modeling,we study the two aspects of lexical feature and comment feature.Firstly,the article representation feature is embedded into the text comment to form the enhanced comment text.Then,the enhanced comment vector and connective vector are input into the network model,and the word2 vec algorithm model is used to generate the enhancement.In addition,the attention layer is added to the convolution layer to enhance the effect of key features on the model,and the interpretability of the recommended model is enhanced;at the top level,the driver decomposition machine model is used to extract high-order hidden features,and then the recommendation results are predicted.By analyzing the experimental data,compared with the benchmark recommendation algorithm,the root mean square error of the proposed method is smaller,and the prediction accuracy is higher.