| In the information-rich Internet age,recommendation systems are vital to alleviate the problem of information overload.As more and more people turn to the Internet to learn and work,tens of thousands of links to resources surround users,making it difficult to access relevant information.As one of the most important information services on the Internet today,it has become a high-profile research topic in the field of machine learning.Recommender systems not only provide convenience to users,but also help commercial organisations to fund relevant research and innovation,improve product quality and functionality,and better understand the consumer market.Variations of graph neural networks(GNNs)such as graph convolutional networks(GCNs),graph attention networks(GATs)and graph recurrent networks(GRNs)have shown groundbreaking performance in many deep learning tasks.By propagating information over graphs,higher-order connections can be effectively modelled.By treating user-project interactions as bipartite graphs and using GCNs to model high-level synergistic signals,researchers can capture the complex synergistic relationships between users,projects and other entities in the graph.By incorporating higher-order connections,graph convolutional networks can discover finer and more meaningful representations of graph structure,thereby improving recommendation effectiveness.However,the main challenges facing GCN-based recommendation models remain how to obtain more effective representations from user-item interactions,and how to address issues such as training difficulties,over-smoothing and the generation of redundant information.Therefore,the main research of this paper is as follows:A propagation layer weighting algorithm mechanism is proposed which is designed for the combination of propagation layers in GCN-based recommendation models.The default graph convolutional propagation layer uses an average weighting scheme that normalises and sums its own nodes as well as the node embeddings of other layers,and does not purposefully distinguish the importance of different propagation layers.As a result,this leads to a model that does not learn the embedding representations of users and items well.In-depth research and comparative experiments on the combined weights of each propagation layer in graph convolutional networks revealed that larger weights should be assigned to more differentiated propagation layers,such as higher-order propagation layers.In addition,a fast graph convolutional network is proposed for problems such as increasingly large datasets and long time model training.In this work,a simplified strategy for generating node representations is presented.It is found experimentally that using only the last layer of the embedding representation as the final representation can give better results than models that integrate multilayer graph embeddings.In convolutional operations,the stitching or stacking of multi-layer graph embeddings can add redundant information and noise to the final node representation.In contrast,a ’last layer’ strategy can improve the performance of the model while making it lighter and easier to scale to large datasets.Through extensive comparison experiments with different frontier recommendation models and datasets,this study achieves good results and further validates the effectiveness of the improvement method proposed in this paper. |
