Heterogeneous Information Network Recommendation Algorithm Based On Graph Convolution

Recommender systems have proven to be valuable means for online users to cope with the information overload and have become one of the most powerful and popular tools in electronic commerce.Collaborative filtering based on matrix factorization is popular to fill in a missing rating matrix.In order to include these different types of extra data together,Heterogeneous information network(HIN)is an important research field.In recent years,graph convolutional neural networks have been widely used in network mining.Heterogeneous network graph convolution is mostly used for node classification tasks.In order to migrate it to recommendation tasks,an Attention based Interactive Mechanism(AIM)is proposed in this dissertation to make full use of the node’s multi-order neighbor information and explore the complex interaction relationship between users and items,thereby constructing a recommendation system model based on heterogeneous graph convolution.Extensive experimental results on three real-world heterogeneous graphs,Amazon,Yelp,DoubanMovie,show that the proposed model with heterogeneous graph convolution and AIM can outperform the classical matrix factorization model and many other new related models.And the exper ments show that when the model goes deeper,MAE will decrease but RMSE may increases,which means the samples with smaller error will be improved and samples with larger error are not so important.In order to verify the robustness of AIM,we tried some other graph convolution structures.We found that all structures have similar accuracy.Therefore,the combination of graph embedding models and AIM is a way to generalize graph classification to recommendation.