Personalized Recommendation Based On Multimodal Heterogeneous Data

The rapid development of Internet technology has profoundly affected and changed people's way of life.With the development of e-commerce websites(such as Taobao,Jingdong,Amazon,etc.),social networking platforms(such as Sina,tiktok,WeChat,etc.)and streaming media platforms(such as NetEase cloud music,Tencent video,and voice buffet),Internet service platform has gradually become the main channel for public consumption,and has promoted the rapid development of Internet economy.How to help users rapidly and accurately find their favorite items from a large number of items is the core problem to improve service quality,enhance user viscosity and platform revenue.Although researchers have conducted many studies on Personalized Recommendation and obtained significant achievements,the performance of traditional collaborative filtering based recommendation is still limited by the data sparsity problem.Multimodal heterogeneous data contains huge valuable information concernd with users or items.Mining valuable information from multimodal heterogeneous data is the key to alleviate the data sparsity problem and improve the quality of recommendation.In this thesis,the methods which can integrate multimodal heterogeneous data with user-item complex interactions for user preference modeling are proposed,so as to improve the performance of recommendation system.In this theis,the main research objectivity is to explore multi-modal heterogeneous data for modeling user preferences.From the view of recommendation,the multimodal heterogeneous data could be divided into three groups,including 1)user-item interactions,2)textual information,and 3)visual information.On the basis of these information,the method of exploring valuable information for user preference modeling is further studied.In summary,I focuses on the research of personalized recommendation towards multimodal heterogeneous data,and verified its effectiveness on several real-world datasets.The contributions of this thesis are listed as follows:(1)User-Item Interactions Based RecommendationIn this theis,the contribution of high-order information on user preference modeling is deeply studied.The nosiy information leveraged by high-order information and over smoothing problems caused by high-order graph convolution are analysised.For exploring the valuable information in high-order information and improving the utilization of high-order information,this theis proposed an Interest-aware Message Passing strategy Graph Convolution Network recommendation model.The model groups users and their interaction items into different subgraphs according to user interest,and conduct high-order graph convolution on the subgraphs.As the proposed model can explore higher-order information,it can effectively alleviate the problem of data sparsity.(2)The Fusion of User-Item interactions and Textual Information for RecommendationThis theis deeply analyzes the influence of item attributes on user preference modeling and the methods to alleviate the problem of data sparsity.For exploring the contribution of attribute information to user interaction and optimizing the representation learning of users and items,this theis proposed An Attributeaware Attentive GCN Model.This model integrates the item attribute into user preference modeling,and introduces the attention mechanism which can filter the information from different adjacent nodes.(3)The Fusion of User-Item Interactions and Multimodal Information for RecommendationA novel Multimodal Attentive Metric Learning model is presented to model user diverse preferences on different aspects of items.This is achieved by a proposed attention neural network,which exploits the multimodal features to estimate the user's specifc attention on each aspect of the target item.Moreover,the model is developed based on a metric-based learning approach,which avoids the inherent limitation of matrix factorization based methods and thus can capture fine-grained user preference and alleviate the data sparsity problem.