Collaborative Filtering Recommendation Model Based On Graph Convolutional Neural Network

With the boom of the Big Data era,"information overload" is a key factor that affects the ability of users to access valuable information quickly and efficiently on the Internet.The existence of recommendation systems is one of the effective ways to solve this problem and has become a hot topic of research in both academic and commercial circles and has been successfully applied in many fields.However,with the increasing size and diversity of data,traditional collaborative filtering recommendation models are underperforming and deep learning techniques have been introduced to capture the characteristics of users and items at a more granular level.At the same time,the successful application of deep learning techniques in various fields such as image,text and video has provided theoretical and practical support for the many research results obtained by Graph Convolutional Neural network(GCN)in many areas.The task of recommendation technology can be seen as a kind of link prediction problem on graphs.Compared to traditional recommendation algorithms,graph convolutional networks have more powerful data representation capabilities and can better tap into the implicit representations in graph-structured data,thus effectively improving the accuracy of recommendation tasks.The core goal of a recommendation system is to infer user preferences in order to proactively provide users with valuable and targeted messages,and collaborative filtering is a classic and widely used approach in recommendation models.However,GCN-based recommendation methods still suffer from the following problems:(1)GCN recursively aggregates messages from different sequential neighbourhoods,but has difficulty distinguishing mixed messages from different nodes,making it difficult to extract key information and eliminate useless information;(2)height nodes have a greater influence on representation learning,further extending the influence of observed edges with neighbourhood aggregation schemes;(3)susceptible to noisy interactions effects as well as problems such as oversmoothing.These limitations impose significant constraints on the recommendation model.Based on the above elaboration,we improve and optimise the graph convolutional network based on the graph convolutional network collaborative filtering recommendation algorithm with node hierarchical aggregation and self-supervised learning assisted tasks to enhance node representation learning.We propose a simple collaborative filtering recommendation model based on an improved GCN.Unlike current GCN-based approaches,we first use the simple GCN model to aggregate neighbourhood messages from neighbourhoods of different orders by separating them from neighbourhoods of different orders,and then aggregates them in a hierarchical manner for collaborative filtering,discarding the non-linear activation of GCN and without introducing additional model parameters.Secondly,self-supervised learning on user-item graphs is also explored.The idea is to complement the learning of the classical recommendation supervision task with an auxiliary self-supervised task,and then the Dropout idea is migrated to the model by designing three data enhancement operators : Node Dropout,Edge Dropout and Randow Walk to generate node sub-views to enhance node representation learning.This is in addition to mitigating the effects of oversmoothing,and is also good at preventing overfitting and improving model performance.Finally,through analysis of the model and experiments on three different recommendation algorithm datasets,the analysis demonstrates the effectiveness of the proposed model in improving recommendation accuracy and robustness to interaction noise.