【作者】 李东；

【导师】 付永强；

【作者基本信息】 哈尔滨工业大学 ， 应用统计（专业学位）， 2023， 硕士

【摘要】 近年来,图神经网络和半监督学习方法的结合,在医药、交通、社交网络等诸多领域有着广泛的应用,同时由于其能够弥补因标记样本不足而导致的模型性能下降问题的优势,能够帮助人们节省大量标注样本所要消耗的各种资源,受到学术界和工业界的广泛关注。然而,当前绝大多数提出的方法都存在着过度平滑和无法处理高异质率的图数据结构的问题。为了解决这两个问题,本文首先分析了邻接矩阵特征值与这两个问题的关系,以此为根据提出了一种创新的方法,即通过引入校准模块到传统的图神经网络中,使得模型在信息传递过程中能够自适应地校准邻接矩阵的特征值。接着,提出了一种新的损失函数来进一步提升模型性能。最后通过在广泛现实数据集上进行实验,我们发现与传统的图神经网络相比,该方法在同时解决过度平滑和异质性问题方面表现出色。此外,该方法还具有适应性,可以轻松应用于不同任务的其他图神经网络模型。这一研究为解决图神经网络过度平滑问题和异质性问题提供了一个全新的方法和统一的视角。更多还原

【Abstract】 In recent years,the combination of graph neural networks and semi-supervised learning methods has a wide range of applications in many fields such as medicine,transportation,and social networks,and has received wide attention from academia and industry due to its advantage of being able to compensate for the problem of model performance degradation caused by insufficient labeled samples,and can help people save various resources that would be consumed by a large number of labeled samples.However,most of the currently proposed methods suffer from the problems of over-smoothing and inability to handle graph data structures with high heterogeneity.To address these two problems,this paper first analyzes the relationship between the eigenvalues of the adjacency matrix and these two problems,as a basis for proposing an innovative approach that enables the model to adaptively calibrate the eigenvalues of the adjacency matrix during the information transfer process by introducing a calibration module into the traditional graph neural network,and proposes a new loss function to further improve the model performance.Through experiments on a wide range of realistic datasets,we find that the method performs well in solving both over-smoothing and heterogeneity problems compared to traditional graph neural networks.In addition,the method is adaptable and can be easily applied to other graph neural network models for different tasks.This study provides a novel approach and a unified perspective for solving the over-smoothing problem and the heterogeneity problem of graph neural networks.更多还原