Research On Collaborative Filtering Based On Graph Neural Network

Nowadays,Collaborative Filtering still plays an important role in the field of recommender system.It is a recommendation algorithm with a wide range of applications and strong interpretability.From the earlier methods based on matrix factorization to the later methods based on deep learning,they optimize the model with interaction information instead of encoding them into the embeddings,which makes it difficult for the embeddings to capture the collaborative information.To address this problem,Graph Neural Network is introduced into collaborative filtering tasks and is widely used.Firstly,in recommender systems,existing collaborative filtering algorithms based on graph convolutional networks usually assume that neighbor nodes only have positive impacts on the target node,and do not consider the possible negative impacts of neighbors on collaborative filtering.To solve this problem,this paper proposes the Light Graph Adaptive Convolution Network Model.It retains only the most important components of graph convolutional networks-neighborhood aggregation and layer combination.It captures the positive and negative information of the target node through low-frequency and high-frequency filters.It learns user and item embeddings by propagating their positive and negative information on user-item interaction graphs through adaptive attention-based methods.At the same time,it uses the self-attention mechanism to combine the embeddings learned at each layer as the final embedding.This model is not only easy to implement,but also has certain interpretability.Experiments show that the model outperforms strong baseline models on standard recall and normalized discounted cumulative gain.Secondly,this paper proposes the Graph Adaptive Convolutional Network based on Receptivity and Similarity of Neighborhood(RSN-GACN).It is an improvement on the Light Graph Adaptive Convolution Network Model.Since the node itself is only used to uniquely identify itself in the collaborative filtering algorithm based on the graph neural network,most of the existing graph neural network models use nonlinear activation and feature transformation to learn the feature vector of the node.This approach results in the model not only capturing less information but also being slow to train.To solve this problem,this paper proposes a neighborhood-aware method,which captures the latent information of nodes and grasps the structural information of graphs through pre-training of nodes.In addition,the sparsity in the dataset has a huge impact on the performance of the model.To solve this problem,this paper proposes a self-attention layer combination module base on neighborhood relationship to capture the correlation between neighborhood information of different layers as auxiliary information for model training.To alleviate the negative impact on model performance due to the data sparsity problem,this paper defines the distance between users and items and their second-order neighbors as the second-order similarity loss,and combine it with the adaptive margin-based Bayesian Personalized ranking loss as a new loss function for model training.Experiments show that the loss function can effectively improve the performance of the model compared with the Bayesian personalized ranking loss function.Finally,this paper proposes a modular framework of collaborative filtering algorithm based on graph neural network.On the basis of unifying the above two models,this framework further generalizes the existing collaborative filtering algorithm model based on graph neural network.Since most of the existing collaborative filtering models based on graph neural networks are single-structure models for specific scenarios,the best-performing models cannot be obtained quickly in the face of new scenarios.Therefore,by analyzing the existing collaborative filtering recommendation model based on graph neural network,this paper studies and analyzes the in+uence of different design dimensions on the model effect.The search space is reduced by a controlled random search method.Guided by the empirical research results,this paper obtains a high-quality model component selection space,which provides a modular framework with excellent performance for model design in new scenarios.