Fine-grained Sentiment Analysis Based Recommendation System

In recent years,with the rapid development of e-commerce,more and more attention has been paid to recommendation algorithms.A large number of users,items,reviews and other related information emerged on the online platform,which also provides an important reference for recommendation algorithms.In this paper,a recommendation system combining the information of users and items is proposed,starting with the interaction between users and items and the construction of user and item profiles through text reviews.For recommendation system,it is necessary to construct user and item profile,to express and learn the relationship between user and item properly.The recommendation system based on neural network can capture the interaction between user and item by using stitching and activation functions after vector representation of user and item for the recommendation score prediction.Experiments show that,compared with traditional recommendation algorithms,the recommendation system based on neural network can better represent and learn the relationship between user and item.In addition to the purchasing and purchased information,the shopping platform usually provides another important reference-reviews written by users.Reviews are usually written by users spontaneously after purchasing items.They often reflect users' preferences and items' characteristics.By using the effective information in the reviews,we can construct the user and item profiles on the one hand,and on the other hand,we can get the reasons for whether the system is recommended or not,which can increase the credibility of the system and users' experience.For the recommendation system,the important information in the review usually includes two aspects: one is the opinion words in the text review,the other is the sentiment polarity of the text review.We can use sentiment analysis technology to solve the problem.There is a correlation between the two tasks.Usually the sentiment polarity of the review depends on the opinion words in the sentence.Therefore,we use a hierarchical network model to jointly learn these two tasks.At the bottom level,we use attention weight to capture more important opinion words for sentiment semantic expression in sentences,and at the top level,we learn the representation of sentiment semantic information.In the experiment,the proposed hierarchical network can effectively extract opinion words from reviews and distinguish the sentiment polarity of sentences.Inspired by the above work,we use review to complement the interaction between users and items.Few studies have considered the importance of both review level and word level.We propose an explainable recommendation algorithm based on hierarchical attention mechanism.Firstly,the interaction information between user and item is properly represented,and then the hierarchical network is used to process the reviews of user and item respectively,as a supplement to the user and item profiles.Among them,attention weights on reviews are used to distinguish the importance of different words in reviews,and attention weights between reviews are used to distinguish the importance of different reviews.Experiments on four real data sets from Amazon show that our model has improved in recommendation rating prediction compared with several state-of-the-art methods.In addition,we extract several review examples from the test data and visualize the attention for word level and review level which verifies the effect of the model for selecting more important words and reviews.