Research On The Message Passing Mechanism Of Graph Neural Network

With the development of deep learning technology,Graph Neural Networks(GNNs)have attracted extensive attention as one of the most effective and prominent graph representation learning methods.At the same time,graph neural networks have been widely used in various graph analysis tasks due to its convincing performance and high interpretability,including recommendation,natural language processing,computer vision and other real-world tasks.Actually,various graph neural networks models have different design characteristics,but their convolutional operations on graph mainly depend on the carefully designed message passing mechanism.Message passing on graph mainly refers to the process of wich graph neural networks propagate and aggregate node features along the network topology and further learn the effective node representations.Despite the importance of the message passing mechanism,there are still few effective explorations on the flexibility,uniformity and applicability of the message passing process.Therefore,this paper aims to explore the message passing mechanism of graph neural networks,and further summarize the characteristics of the flexibility,uniformity and applicability.This paper then completes the summary and exploration of existing message passing mechanism,imporves the shortcomings and further designes new strategies for graph neural networks.In order to explore the flexibility of the message passing mechanism for GNNs,this paper first present an experimental investigation,and proves that the fusion capability of classical GCN in fusing node features and network topology id distant from optimal.Secondly,in order to improve the fusion flexibility,we propose an adaptive multi-channel graph neural networks model as AM-GCN,which can retain the advantages and remedy the weakness of current message passing mechanism.Meanwhile,considering that the input topology and node features in AM-GCN are still predefined and fixed,we introduce the edge weights learning strategies and further propose the label propagation guided multi-channel graph convolutional networks named as LPM-GCN.Our extensive experiments on benchmark datasets clearly show that our proposed models can effectively improve the flexibility of the message passing process of graph neural networks and improves the classification accuracy with a clear margin.In order to explore the uniformity of the message passing mechanism for GNNs,we first summarize and conclude the general definition for the message passing process.Secondly,we establish a surprising connection between different propagation mechanisms with an optimization problem,and propose a unified optimization framework to explain the message passing process of classical graph neural networks.Then,we improve the existing framework and propose a generalized unified optimization objective framework with a flexible feature fitting function and a generalized graph regularization term.We also analyze the general solutions of the unified framework,which provide a more convenient way for deriving corresponding propagation results.Based on the proposed two optimization framework,we further consider the graph neural network model designing problem,and we develop three novel objective functions considering adjustable graph kernels or high-order structural information during propagation respectively.Extensive experiments on benchmark datasets clearly show that the newly proposed GNNs not only outperform the state-of-the-art methods but also have good ability to alleviate oversmoothing,which further verify the feasibility for designing GNNs with the unified optimization framework and complete the exploration of the uniformity of the message passing mechanism for GNNs.In order to explore the applicability of the message passing mechanism for GNNs,we mainly focus on the personalized recommendation application in our paper.Actually,we take the user coldstart problem in social recommendation as an example to explore how to improve the message passing mechanism for GNNs.User cold-start problem is a long-standing challenge that restricts the performance of a recommendation system.Most traditional recommendation methods usually fail to learn precise preferences for new users due to the lack of user-item interactions,and the existing graph neural network methods seldom investigate the natural mechanism of GNNs,i.e.,which neighbors should be included and how to aggregate rich side information,for solving data sparsity problems in cold-start recommendation scenarios.Then in theis paper,we propose a novel receptive-feld adaptive Graph Neural Network for user cold-start recommendation,named READ.READ tailored the receptive-field not only consider the sparsity information,but also model the high-order social knowledge to fully estimate user preferences.Extensive experimental results on two real-world industrial datasets and one benchmark dataset clearly demonstrate that our proposed READ signifcantly outperforms the state-of-the-art models.