Research On Personalized Information Recommendation Integrated Multi-modal Knowledge Graph And Attention Mechanism

With the rapid development of digital technology and the Internet,we live in an era of information explosion,surrounded by a large amount of information and choices every day.As an intelligent technology,the recommendation system can use users ’historical data and personal preferences to recommend goods,services or content related to their interests to users,so as to help users find what they want faster and more effectively.However,the recommendation system is also facing some problems,the most important of which is data sparseness.Data sparsity refers to the fact that the user’s historical behavior data is too small to accurately analyze and recommend the user ’s interest.In this case,the recommendation system may have recommendation errors or cannot provide satisfactory recommendations for users.Aiming at this problem,the current research trend is to introduce external information into the recommendation model,in which the knowledge graph is widely used.However,the existing knowledge graph recommendation methods mainly focus on using the path structure information in the knowledge graph or the semantic information in the knowledge graph,ignoring the rich high-order information inside the knowledge graph.Aiming at the problem of data sparseness in the recommendation system and the problem that the high-order information inside the knowledge graph has not been fully utilized,this thesis proposed a personalized information recommendation method that combined multi-modal knowledge graph and attention mechanism.The method mainly included two parts : the knowledge graph representation method that integrated multimodal information and the attention mechanism propagation model based on relational space.The specific research contents were as follows :(1)Addressing the problem of data sparsity faced by recommendation systems in practical applications,this thesis proposed a new knowledge graph representation method that integrated multi-modal information.Specifically,visual images(such as movie posters)corresponding to items were used as internal multimodal information in an entity-based form to expand knowledge graph recommendation.However,in the process of knowledge graph embedding,the text attributes of items were directly embedded into an ID,thus ignoring some semantic information contained in items,such as movie name information.Therefore,this thesis introduced the title text semantics of items as external multimodal features to make up for the semantic loss caused by knowledge graph embedding.At the same time,adding user feature attributes to external multimodal information can more accurately analyze users ’ interests.Through this method,the recommendation system can better deal with the problem of data sparsity.(2)Tackling the problem of insufficient mining of high-order information in knowledge graph recommendation,this thesis proposed an attention mechanism propagation fusion model based on relational space.Specifically,the model first obtained multimodal feature vectors from the internal multimodal knowledge graph layer as input,and learned the weight of each neighbor in the propagation process through the relationship-based perceptual attention mechanism,so that entities with closer relationships could be found in the relationship space.In the feature aggregation stage,this thesis used addition,connection and double-ended mixing to test the aggregation effect.Finally,the external multi-modal user and item feature vectors were added,and the final user and item feature vectors were fused to predict the recommendation.In order to test the effectiveness of the personalized information recommendation method combining multi-modal knowledge graph and attention mechanism proposed in this thesis,this thesis conducted a comparative experiment based on the MultiMovie Lens1 M dataset,and compared it with several current mainstream recommendation methods in terms of accuracy and recall rate.Experimental results showed that the proposed method was better than MKR,Ripple Net and other knowledge graph recommendation methods in terms of accuracy,recall and ACC.A comparative experiment was conducted on data sparsity.The experiment showed that the data sparsity problem of the recommendation system was further effectively alleviated while ensuring the accuracy of the recommendation.