Research On Dual-target Cross-domain Recommendation Algorithm Based On Graph Neural Network

With the rapid development of information technology and the era of big data,major Internet vendors have built their own data ecosystem and are involved in the recommendation scenario of multiple domains at the same time,which will inevitably lead to the phenomenon of cross-platform or cross-domain user information,in order to enable the information of different domains to share and complement each other,to achieve the bidirectional transfer of information between different domains so as to achieve the purpose of improving the accuracy of different domain recommendations at the same time.Dual-Target Cross-Domain Recommendation(DTCDR)has emerged.However,the existing DTCDR models have the following problems: First,some graph information is lost when learning the node embedding representation of user-item interaction graphs with graph structure types,resulting in the inability to obtain better node higher-order embedding representation and thus affecting the recommendation performance.Second,the traditional DTCDR model focuses on modelling the historical user-item interaction data in the respective domains,ignoring the influence of the domain characteristics of the nodes on the node embeddings,and second,ignoring the reality that different neighbour nodes in the user-item interaction graphs have different degrees of importance to the central node.Addressing the above issues and challenges facing DTCDR research,the main work is summarised as follows:(1)A Dual-Target Cross-Domain Recommendation Model with Graph Convolutional Neural Network(GCN＿DTCDR)is proposed to address the problem of learning higher-order embedding representations of users/items nodes from existing DTCDR models.The model consists of three main parts: firstly,the users-items interaction graph is constructed based on the historical interaction information between users and items;then,a graph convolutional neural network is introduced on this model,and the end-to-end graph embedding learning method relies on the iteration of convolutional layers to simulate the process of local propagation of node information,and the node embedding representation is updated by aggregating the information of multi-order neighbours nodes,so that the node embedding contains both graph structure information and Finally,it uses the idea of transfer learning to transfer user preferences from one domain to another,thus realizing dual-target cross-domain recommendation.The model is compared on three real datasets of Douban(Book,Movie,Music),and the experimental results show that the model has better rating prediction ability compared with the baseline model,and effectively alleviates the problem of sparse node data.(2)To further alleviate the problem of sparse node data,the above GCN＿DTCDR model does not consider the influence of node domain feature information and the importance of neighboring nodes on the recommendation performance,a Dual-Target Cross-Domain Recommendation Algorithm model for Fusion Graph Attention Node Classification Task is proposed,called DTCDR＿GANCT,(A Dual-Target CrossDomain Recommendation Algorithm for Fusion Graph Attention Node Classification Task).Unlike the GCN＿DTCDR model,which is modelled in the respective domains,this model innovatively constructs a multi-domain users-items interaction graph by historical interaction behaviours of user and item nodes on multiple domains,and introduces a node classification task on graph neural network to predict the domain to which nodes belong,so that node embedding eventually carries domain feature information and enriches the embedding representation of nodes;using graph attention network to assign different weights to neighbourhood nodes and identify more important neighbourhood nodes,which is more realistic.The model was subjected to comparison experiments,ablation experiments and parameter sensitivity experiments on three standard sub-datasets of Douban.The results of the comparison experiments show that the model significantly outperforms other baseline models on the Top-N recommendation task;the designed ablation experiments demonstrate the effectiveness of two major factors,the graph attention mechanism and node label classification,in improving the performance of the model.