Research On Personalized Semantic Recommendation Algorithm For Online Reviews

With the rapid development of science and technology in today's society,artificial intelligence,cloud computing and other technologies have gradually matured.Countless network services have made the scale of data and information volume show an exponential growth.In order to process these huge data information,the recommended system Application is indispensable.The recommendation system needs to record the explicit or implicit behavior of the user's historical interactive behavior,discover the user's preference characteristics,and then make different recommendations to different users according to product attributes.This paper makes two classifications of traditional online review recommendation algorithms,document-based modeling and review-based modeling.The recommendation algorithm based on review modeling has the problem of high computing power consumption and user preferences will change over time.Improvements are proposed.The algorithm based on document modeling has low prediction accuracy and low interaction between user preferences and item attributes.To improve,in view of the above problems,the research work of this paper is as follows:First of all,in view of the different attribute weights that the same user cares about for different products,this article uses the Word2Vec model to characterize the input review content,and then extracts the aspect features of each review based on the review modeling method,which adds local-The attention mechanism assists in aspect-level feature extraction,uses the attention mechanism to calculate the weight of each aspect,then performs feature aggregation,and uses optimized LFM for result prediction.After adjusting the parameters of the algorithm to reach the optimal MSE and MAE,the experimental comparison is carried out.The experimental comparison selects the aspect recommendation algorithm based on document modeling and the two algorithms MPCN and TNET based on review modeling.Three sets of 10,000 public test sets are used respectively.Using 80%,10%,and 10%ratios as training data,verification data and test data for testing,the results show that the algorithm proposed by the above theory is very effective in predicting results.Secondly,for the cold start problem and the problem of low interaction between user preferences and item attributes based on document modeling,this paper uses the Glove model to characterize the input comment content,and then uses two parallel CNNs to extract features of user preferences and product attributes respectively.Finally,feature fusion is performed and PMF is used to optimize the prediction results.This algorithm is different from the ordinary bag-of-words model because the Glove model retains the word order and does not need to set the length of the data set.Two parallel CNNs solve the interaction between user preferences and item attributes.Finally,PMF was chosen to solve the problem of matrix sparseness,and then the parameters of the algorithm were adjusted to reach the optimal values of MSE and MAE,and then experimental comparisons were carried out.For comparison experiments,MEUMF,DeepCoNN,and D-ATT based on document modeling were selected for experimental comparison.The experimental data and data division method are the same as above,and the results show that the algorithm proposed by the above theory has a very good effect on the prediction results.In summary,this article proposes and implements different improvement methods for the current two types of modeling recommendation algorithms,and analyzes the final results.Based on review modeling,it reduces computing power consumption and improves the predicted value of the results.Based on document The modeled recommendation algorithm alleviates the cold start problem,increases the interaction between user preferences and item attributes,and the integrated Glove model avoids the drawbacks of the preset result set size,and also improves the result prediction problem.