Deep Learning Recommendation Algorithm Based On Reviews

With the rapid development of the Internet,the problem of information overload affects people's ability to quickly and accurately find the information they need,so the recommendation system has become an increasingly important helper in the process of browsing information.It is of great significance to study how to improve the accuracy of recommendation to optimize user experience.The sparsity of the rating data limits the recommendation algorithm based on ratings to further improve accuracy,and the rich and diversified information of users and items contained in reviews makes the recommendation algorithm based on reviews become the main development trend in the recommendation field in recent years.Improving the ability to mine user characteristics and item attribute characteristics in the review text,and obtaining more comprehensive user and item related information are the core keys to improving the accuracy of recommendation.However,the recommendation algorithm based on reviews is still in an imperfect research stage,and there are still some problems,such as insufficient feature extraction ability,limited attention to user preference changes,and insufficient comprehensiveness of text representation.To solve these problems,this paper proposes a deep learning recommendation algorithm based on reviews and item descriptions and a deep learning recommendation algorithm based on multi-dimensional sentence embedding.Compared with the current mainstream algorithm,they have significant improvements in feature extraction,corpus diversification,text representation,and attention to changes in users' interests,which can greatly improve the accuracy of recommendation results.The main research contents and innovations of this paper are as follows:(1)Aiming at the shortcomings of the current recommendation algorithms based on reviews,such as insufficient feature extraction ability,limited attention to the change of user preferences,and single corpus,a deep learning recommendation algorithm based on reviews and item descriptions was proposed.Based on the reviews,the algorithm supplemented the item description documents,which are helpful to learn more comprehensive item attribute characteristics and alleviate the cold start problem of the item.The algorithm uses the BERT model with strong feature extraction ability to process text data,uses global information to predict words,so as to improve the feature extraction ability.After that,the algorithm uses LSTM method to learn the changes of user interests over time,accurately grasp the user's current preferences,and further improve the recommendation accuracy.In the model training stage,the reviews data from 1 to 5 points were randomly sampled according to the proportion of five equal points to ensure that the data amount of each score value was equal,so as to reduce over-fitting and improve the robustness of the model.Experimental results show that compared with the recommendation algorithm based on reviews proposed recently,the recommendation accuracy of the proposed algorithm is improved by 5.1%.(2)At present,the mainstream recommendation algorithms mostly use a deep learning method to mine text features,so as to obtain word embedding vector or sentence embedding vector,but few works consider using multiple methods to capture more information to optimize the recommendation effect.To solve the problem of incomplete text representation,this paper proposes a deep learning recommendation algorithm based on multi-dimensional sentence embedding.This algorithm uses three advanced word embedding methods,ELMo,GPT and BERT,to get three different groups of word embedding,calculates the average value,maximum value and minimum value of the three groups of word embedding respectively,and connects the three groups of word embedding to get multi-dimensional sentence embedding.Compared with the sentence embedding obtained by the single method,this multi-dimensional sentence embedding includes word embedding of multiple models in multiple spatial dimensions,which effectively improves the comprehensiveness of text representation and helps to better depict the features of users and items.Finally,the review level attention mechanism and AFM were used to learn the importance of different reviews and the user-item feature interaction,respectively,to get the predicted score.The experimental results show that for the same reviews,the proposed recommendation algorithm based on multi-dimensional sentence embedding can extract more user and item information,and the recommendation accuracy is improved by 3.8% compared with the current mainstream algorithm.