Research On Cross-domain Deep Recommendation Algorithm Based On Reviews

With the rapid development of smart devices such as the Internet and computers,and the increasingly mature technologies such as cloud computing and big data,the scale of information data has grown exponentially.In order to alleviate the urgent situation that users are difficult to find valuable information in the face of numerous information,the personalized recommendation system was born and has been widely used in today’s business competition.To achieve personalized recommendation,the most important thing is the interaction between users and projects.As a key interaction behavior,comments contain a large number of user and project characteristics,providing a new way to solve the problem of data sparsity and cold start in the recommendation system.Therefore,this thesis uses comment information as the basis for recommendation,uses cross domain recommendation technology to improve the accuracy of recommendation in cold start and data sparse situations,and proposes two comment based recommendation algorithms to work in different recommendation scenarios.The main work is as follows:(1)Whether the feature extraction of review information is accurate and sufficient will directly affect the performance of review-based recommendation algorithms.In order to obtain more accurate review feature information,starting from the feature extraction process of review information,this thesis proposes a review-based Deep Feature Integrate Recommendation(DFIR)algorithm.First of all,a network structure of feature stacking is proposed,which uses a jump connection method similar to the residual structure to stack shallow features and deep features to solve the problem of feature loss in the feature extraction process of comments,Then the context feature information of comments is introduced to distinguish the different meanings of the same words in the context,and feature fusion is used to combine the two and introduce the attention mechanism,Finally,the extracted user features and project features are fused and predicted in the fusion layer.Through the experimental verification,we have achieved good recommendation results in a single field.(2)In order to further solve the problem of accurate recommendation under cold start,this thesis proposes a deep aspect cross domain recommendation(DACDR)algorithm incorporating aspect level.By introducing cross domain recommendation technology,an auxiliary domain is built on the basis of DFIR to realize knowledge migration between different domains.In order to solve the effectiveness of user preference migration between different domains,an Aspect level feature modeling method is introduced,In addition,a cross domain Aspect attention mechanism is used to learn the Aspect correlation between different domains,so that users’ preferences can be effectively transferred between different domains to solve the recommendation problem in the cold start environment.Figure 32 Table 6...