Recommendation Algorithm Based On Heterogeneous Graph Attention Network And Recurrent Neural Network

With the increasing amount of information in the internet,users require accurate and personalized recommendations while surfing the internet.How to cultivate latent patterns and preference in historical records of users to make accurate recommendation become a priority in the field of recommendation system.Streaming platform such as You Tube and Online shopping platform such as Alibaba utilized varieties of recommendation algorithms to improve user experience.Collaborative Filtering is one of the classic recommendation algorithms,it is based upon an assumption,which assumes that users who purchase the same item have similar preference.Therefore,it recommends the items bought by other users who made similar purchases.However,it is not capable of reflecting changes of preference in a long term and suffered from sparseness of data as well.In many real-world scenarios,data is stored as graphs consist of multiple kinds of edges,such as item-item relations,user-user relations,user-item relations(purchase,click,comment).Besides,many edges contain timestamps.Based on the Heterogeneous Graph Attention Network(HGAN)and Recurrent Neural Network(RNN),this thesis propose a novel approach namely Heterogeneous Graph Attention Network with Time for Recommendation,i.e.HGANTRec,which can learn information from both graph structure and sequential data to make personalized prediction,thus,further improve user experience.This thesis conducts systematical evaluation in three challenging subsets of Amazon data set,all of them have several million edges,and result of tests further confirmed both the effectiveness and efficiency of the HGANTRec model in practice.