The Study And Application Of Recommendation Algorithm Based On User Reviews Modeling

With the explosive growth of online information,the recommendation system has become an effective way to deal with information overload.The recommendation system personalizes the filtering of valid information by inferring user interest from the user's historical behavior.Classical collaborative filtering methods are simple and effective,but are often limited by sparse scoring data and cold-start problems.Therefore,the introduction of additional information such as user reviews becomes an important method to improve the performance of the recommendation system.In recent years,user reviews have played an increasingly important role in the recommendation system,effectively alleviating the limitations of collaborative filtering methods.Online platforms such as e-commerce encourage users to write reviews and extract a variety of information for user-object image modeling to improve recommendation performance.With the rapid development of deep learning in the fields of natural language processing and recommendation systems,user review texts have been more effectively utilized,and depth models such as D-ATT,DeepCoNN,and NARRE have achieved good results in userobject image modeling..However,there are several problems in these deep learning methods:(1)only use the characteristics of the item to represent the user's interest,but do not consider the user's feelings about the characteristics of the item;lack of ability to adapt to the variable length phrase,can not accurately and completely extract the item features and emotional phrases.(2)Due to the sparse data,it is difficult to infer the fine-grained,domain-related emotional word meaning from the recommendation task;the emotion classification task is used to improve the recommendation performance,but the relationship between the recommendation task and the emotion classification task is ambiguous.(3)When filtering a phrase,only the local context(word combination)of the word is considered,but the influence of the global context(the overall nature of the item)is ignored;only the comment text is scanned once,ignoring the effect of multiple rounds of iterative refinement.In response to the above questions,this article expands the following three tasks in turn:First of all,this paper proposes "Review Modeling at Multi-level for Recommendation".The model divides the user-object image extraction into four levels:(1)"phrase extraction layer",flexible extraction of variable-length phrases;(2)"phrase association layer","emotional phrase" and "item characteristics" The phrase is associated with the user's interest in the particular aspect of the item.(3)"Review layer",each review is independently modeled,each review represents a certain aspect of the user's relative concentration(4)"interaction layer",the user and candidate items are matched,and the predicted score is calculated.We conducted comprehensive experiments on multiple datasets from different areas of Amazon,and the results showed that the mean squared error model overall exceeded the most advanced methods,including MF,D-ATT,DeepCoNN,and NARRE.Then,this paper proposes the "Sentiment Analysis Oriented Multi-task Learning for Recommendation".The model introduces additional supervision information by means of "sentiment classification auxiliary task" to reduce the learning difficulty of "recommended task".By sharing the representation of the phrase between the two tasks,the recommendation task can better understand the semantics of the emotional phrase.In addition,the model selects taskrelated phrases from shared representations through attention mechanisms,clearly reflecting the relationship between the two types of tasks.Finally,this paper proposes "Iterative Refined Product Profile Modeling for Recommendation".The model uses multiple rounds of attention to iteratively refine the image of the item.The first round only determines the importance of the phrase from the local context,constructing the "rough portrait" of the item as the global context(the overall nature of the item).In the second round,the “rough portrait” is re-screened to generate a “fine portrait” to eliminate some noise and enrich the details of the portrait.