Research On Recommendation Algorithm Based On User Reviews Semantics Information

With the rapid development of the fifth-generation mobile communication system 5G technology,the world has entered the era of big data,users can no longer effectively use the massive amount of Internet data information,"information overload" problem is becoming increasingly serious,the emergence of the recommendation system to a certain extent to alleviate the problem of information overload.The recommendation system effectively uses various behavioral information to model users and filter invalid information in a personalized way.Rating prediction has always been a core problem in the field of recommender systems,which predicts the scores of items not rated by users from their existing historical rating records.Although collaborative filtering methods based on matrix factorization are simple and powerful,they are often limited by the cold start problem caused by new user or new item entering the system.Therefore,introducing user review text to be modelled together with the rating matrix has become an important approach to improve the performance of recommendation systems.Reviews of items are generated by users,which contain the reasons for their high and low ratings and suggestions for items.Ratings and reviews are complementary resources,and joint training can build user models that better match user preferences and needs,and make more accurate personalized recommendations.Meanwhile,deep neural networks have powerful end-to-end feature extraction capabilities,overcoming the shortcomings of traditional bag-of-words model models that ignore contextual semantic relationships,better portraying user preferences and item features,and also providing a new direction for review-based item recommendation.To date,deep recommendation systems for review text have made many advances,and to a large extent alleviated the shortcomings of collaborative filtering related models based on matrix factorization,including CDL,DeepCoNN,Trans Net,D-Attn,NARRE,In-depth models such as Tar MF,CARP,CAML,etc.,based on natural language processing related technologies,fully feature extraction of review text,better modeling for users,and finally achieved good recommendation results.However,the following problems remain in the study of review-based recommendation systems:(1)It is difficult to infer the deep semantics of domain-specific sentiment words from the recommendation task alone,and there are no keywords that focus on the user's preferences and the attributes of the item;the relationship between the recommendation task and the sentiment analysis task is unclear and there are problems with the way information is shared between the two tasks.(2)DeepCoNN applies the target reviews directly to the item when predicting the user's rating of the target item.The construction of the test data is unreasonable because the user's reviews on the item cannot be known in advance;most existing models do not consider the focus of the user and the item in different aspects.To address the above issues,this thesis makes the following research:First,this thesis proposes a " Recommendation Models Incorporating Multi-task Learning and Attention Mechanisms",which introduces a sentiment analysis auxiliary task to help recommend the main task and clarifying the connection and distinction between the two tasks.The shared layer learns shared user preferences and item attribute representations from sentiment analysis and recommendation tasks through a local attention mechanism,so that the two types of tasks complement each other and work together.The private layer visualizes the different focus of the two tasks on different types of phrases through attention pooling,clarifying the difference between the recommendation task and the sentiment analysis task.Then,this thesis proposes a " Gated Convolution and Aspect-level Attention Mechanism for Rating Prediction ",which divides the modelling of user-item portraits into two networks:the target network and the source network.The gated convolution unit in the target network allows for a fuller representation of the sentiment of the target review features,while the aspect-based attention mechanism in the source network supervises the specific aspects of interest to the user,while the Trans network enables the review representation learned in the source network to converge infinitely with the target review,and finally the source network is used to predict the ratings.Experiments on the effectiveness of the model on five datasets from different Amazon domains and a Yelp restaurant dataset in this thesis show that the model outperforms existing methods overall,validating that the two review-based depth models proposed in this thesis can improve the accuracy of rating prediction and subsequently improve recommendation performance.