Research On Recommendation With Heterogeneous Graph Neural Network

With the spread of information infrastructures,Internet users are daily facing increasingly enormous information.As the key technology to alleviate information overload problem,recommendation systems are receiving attention from more and more researchers,and have been widely used in various scenarios,such as e-commerce,social networks and micro-video platforms.The aim of recommendation systems is to predict the next interested item based on the observed user behavior data.As resourceful auxiliary features,users' social relations and items' domain knowledge have been widely applied by many research works,to refine the embedding learning for users and items and improve the effectiveness of recommender systems.But there are few recommendation methods introduce both categories of the information into the recommendation model so far.To this end,this thesis proposes a Heterogeneous Graph Memory Network,which is a recommendation model based on heterogeneous graph neural network and have the ability to introduce users' social information and items' domain knowledge information into the recommendation model.It includes the following three well designed modules:(1)Heterogeneous graph node feature encoding module.This module firstly extract nodes' features by multiple layers of graph neural network consisting of memory-augmented relation heterogeneity encoder and message aggregator.Then,we obtain the embedding vectors of nodes by aggregating features extracted from each order.(2)User embedding recalibration module.This module injects the influence of the users' social ties into the item representation learning by adding a recalibration term,which brings performance improvements.(3)Recommendation module.This module makes preference predictions by users' recalibrated embedding vectors and items' embedding vectors,and the final recommendation results are the items with the highest preference.In the experimental part,we carry out extensive contrast experiments with state-of-the-art algorithms on three large-scale real-world datasets.The experimental results demonstrate the rationality and superiority of the designed model.