Recommendation Algorithm Based On Graph Neural Network And Recurrent Neural Network

With the development of technology in the information age,and the amount of data shows a trend of massive increase,the problem of information overload has become increasingly important,and thus recommendation systems have emerged at the historic moment.Recommendation system utilizes the user’s historical data to dig out the user’s preferences and recommend some information that the user may want.The technology of recommendation systems is also developing,especially the deep learning technology is developing rapidly,from the Convolutional Neural Networks(CNN)for processing image data to Recurrent Neural Networks(RNN)which processes sequence information,and then to the current popular Graph Neural Networks(GNN).The development of these technologies promotes the development of recommendation systems.Since the user-item interactions can be regarded as a kind of graph structure data,graph neural networks can naturally be used in recommendation systems.In recent years,user-item recommendation algorithms based on graph neural networks have been widely used.However,most existing GNN-based recommendation methods only use the weighted sum of neighboring node’s feature embeddings when performing information aggregation.These methods assume that neighboring nodes are independent individuals,and ignore the potential correlation between neighbors,which leads to the failure to capture the "co-occurrence" signal.This thesis proposes a new recommendation algorithm(GCNNIRec)based on graph convolutional network and neighbor interaction network to capture possible co-occurrence signals between node neighborhoods.Specifically,GCNNIRec contains two types of modules,namely Linear-Aggregator module and Interaction-Aggregator module.The previous module linearly aggregates the features of adjacent nodes to obtain a representation of the target node.The latter uses the interactions between neighbors to capture the co-occurrence characteristics.In order to further utilize the high-order information of the graph and the sequence information of the interactive data,this thesis extends the GCNNIRec model(named GCNNIRec-ext).Specifically,the thesis extends the Linear-Aggregator module from one layer to multiple layers to learn high-order information,uses RNN to learn sequence features,and retains the Interaction-Aggregator module.The aggregation part is used to learn the "co-occurrence" feature,and then the three parts are combined to obtain the final embedding representation of the node,which is used for downstream tasks.The experiments on three real data sets such as Yelp,ML-100 K and Amazon have achieved excellent results.Compared with several classical methods in this thesis,the RMSE of the extended model has increased by 5.8%,4.6%,and 2.9% respectively.It confirms the better performance of GCNNIRec-ext and the performance improvement achieved by introducing Interaction-Aggregator module in GNN,and also confirms that recurrent neural network embedding part contributes a lot to the model performance enhancement.