面向多GPU的图神经网络训练加速

图神经网络由于其强大的表示能力和灵活性最近取得了广泛的关注. 随着图数据规模的增长和显存容量的限制, 基于传统的通用深度学习系统进行图神经网络训练已经难以满足要求, 无法充分发挥GPU设备的性能. 如何高效利用GPU硬件进行图神经网络的训练已经成为该领域重要的研究问题之一. 传统做法是基于稀疏矩阵乘法, 完成图神经网络中的计算过程, 当面对GPU显存容量限制时, 通过分布式矩阵乘法, 把计算任务分发到每个设备上, 这类方法的主要不足有: (1)稀疏矩阵乘法忽视了图数据本身的稀疏分布特性, 计算效率不高; (2)忽视了GPU本身的计算和访存特性, 无法充分利用GPU硬件. 为了提高训练效率, 现有一些研究通过图采样方法, 减少每轮迭代的计算带价和存储需求, 同时也可以支持灵活的分布式拓展, 但是由于采样随机性和方差, 它们往往会影响训练的模型精度. 为此, 提出了一套面向多GPU的高性能图神经网络训练框架, 为了保证模型精度, 基于全量图进行训练, 探索了不同的多GPU图神经网络切分方案, 研究了GPU上不同的图数据排布对图神经网络计算过程中GPU性能的影响, 并提出了稀疏块感知的GPU访存优化技术. 基于C++和CuDNN实现了该原型系统, 在4个不同的大规模GNN数据集上的实验表明: (1)通过图重排优化, 提高了GPU约40%的缓存命中率, 计算加速比可达2倍; (2)相比于现有系统DGL, 取得了5.8倍的整体加速比.     

一种面向图神经网络的图重构防御方法

近年来,图神经网络在图表示学习领域中取得了较好表现广泛应用于日常生活中,例如电子商务、社交媒体和生物学等.但是研究表明,图神经网络容易受到精心设计的对抗攻击迷惑,使其无法正常工作.因此,提高图神经网络的鲁棒性至关重要.已有研究提出了一些提高图神经网络鲁棒性的防御方法,然而如何在确保模型主任务性能的前提下降低对抗攻击的攻...     

大规模图神经网络系统综述

图神经网络(GNN)是一类基于深度学习的处理图域信息的方法,它通过将图广播操作和深度学习算法结合,可以让图的结构信息和顶点属性信息都参与到学习中,在顶点分类、图分类、链接预测等应用中表现出良好的效果和可解释性,已成为一种广泛应用的图分析方法.然而现有主流的深度学习框架(如TensorFlow、PyTorch等)没有为图...     

Pre-Train and Learn:Preserving Global Information for Graph Neural Networks

Graph neural networks(GNNs) have shown great power in learning on graphs.However,it is still a challenge for GNNs to model information faraway from the source node.The ability to preserve global information can enhance graph representation and hence improve classification precision.In the paper,we propose a new learning framework named G-GNN(Global information for GNN) to address the challenge.First,the global structure and global attribute features of each node are obtained via unsupervised pre-training,and those global features preserve the global information associated with the node.Then,using the pre-trained global features and the raw attributes of the graph,a set of parallel kernel GNNs is used to learn different aspects from these heterogeneous features.Any general GNN can be used as a kernal and easily obtain the ability of preserving global information,without having to alter their own algorithms.Extensive experiments have shown that state-of-the-art models,e.g.,GCN,GAT,Graphsage and APPNP,can achieve improvement with G-GNN on three standard evaluation datasets.Specially,we establish new benchmark precision records on Cora(84.31％) and Pubmed(80.95％) when learning on attributed graphs.     

Towards Defense Against Adversarial Attacks on Graph Neural Networks via Calibrated Co-Training

Graph neural networks (GNNs) have achieved significant success in graph representation learning. Nevertheless, the recent work indicates that current GNNs are vulnerable to adversarial perturbations, in particular structural perturbations. This, therefore, narrows the application of GNN models in real-world scenarios. Such vulnerability can be attributed to the model’s excessive reliance on incomplete data views (e.g., graph convolutional networks (GCNs) heavily rely on graph structures to make predictions). By integrating the information from multiple perspectives, this problem can be effectively addressed, and typical views of graphs include the node feature view and the graph structure view. In this paper, we proposeC²oG, which combines these two typical views to train sub-models and fuses their knowledge through co-training. Due to the orthogonality of the views, sub-models in the feature view tend to be robust against the perturbations targeted at sub-models in the structure view.C²oGallows sub-models to correct one another mutually and thus enhance the robustness of their ensembles. In our evaluations,C²oGsignificantly improves the robustness of graph models against adversarial attacks without sacrificing their performance on clean datasets.     

图神经网络在复杂图挖掘上的研究进展

图神经网络对非欧式空间数据建立了深度学习框架,相比传统网络表示学习模型,它对图结构能够实施更加深层的信息聚合操作.近年来,图神经网络完成了向复杂图结构的迁移,诞生了一系列基于复杂图的图神经网络模型.然而,现有综述文章缺乏对复杂图神经网络全面、系统的归纳和总结工作.将复杂图分为异质图、动态图和超图3种类型.将异质图神经网络按照信息聚合方式划分为关系类型感知和元路径感知两大类,在此基础上,分别介绍普通异质图和知识图谱.将动态图神经网络按照处理时序信息的方式划分成基于循环神经网络、基于自编码器以及时空图神经网络三大类.将超图神经网络按照是否将超图展开成成对图划分为展开型和非展开型两大类,进一步按照展开方式将展开型划分成星形展开、团式展开和线形展开3种类型.详细阐述了每种算法的核心思想,比较了不同算法间的优缺点,系统列举了各类复杂图神经网络的关键算法、(交叉)应用领域和常用数据集,并对未来可能的研究方向进行了展望.     

基于复杂结构信息的图神经网络序列推荐算法

图结构因其在序列推荐场景中的自然适应性而备受关注,而现有的基于图神经网络的会话序列推荐算法虽然能够利用图结构信息达到较好的推荐效果,但是没有考虑用户在会话序列中的重复点击行为和项目之间的复杂转换,且未很好地利用图中复杂的结构信息,导致推荐的效果受到一定程度的限制。提出有向与无向信息同注意力相融合的图神经网络序列推荐算法,并基于推荐算法给出项目隐含向量建模算法,结合会话序列图中的有向结构信息与无向结构信息,通过考虑用户的重复点击行为和引入注意力机制建立会话中点击项目的复杂转换模型。图节点在特征传播的过程中平衡邻居节点信息与自身信息的比例,以更准确地预测推荐过程中生成的会话向量。在Diginetica、Yoochoose 1/64、Yoochoose 1/4 3个数据集上的实验结果表明,与SR-GNN、TAGNN算法相比,该算法精度最高提升4.34%,能够更好地预测用户在会话中的下一次点击精度。     

图神经网络前沿进展与应用

图结构数据是现实生活中广泛存在的一类数据形式.宏观上的互联网、知识图谱、社交网络数据,微观上的蛋白质、化合物分子等都可以用图结构来建模和表示.由于图结构数据的复杂性和异质性,对图结构数据的分析和处理一直是研究界的难点和重点.图神经网络(Graph Neural Network,GNN)是近年来出现的一种利用深度学习直接对图结构数据进行学习的框架,其优异的性能引起了学者高度的关注和深入的探索.通过在图中的节点和边上制定一定的策略,GNN将图结构数据转化为规范而标准的表示,并输入到多种不同的神经网络中进行训练,在节点分类、边信息传播和图聚类等任务上取得优良的效果.与其他图学习算法相比较,GNN能够学习到图结构数据中的节点以及边的内在规律和更加深层次的语义特征.由于具有对图结构数据强大的非线性拟合能力,因此在不同领域的图相关问题上,GNN都表现出更高的准确率和更好的鲁棒性.本文在现有GNN研究的基础上,首先概述了GNN的出现历程,并介绍了相关概念和定义.之后本文着重讨论和对比了GNN中的各种算法框架,包括核心思想、任务划分、学习方式、优缺点、适用范围、实现成本等.此外,本文对GNN算法在多个不同领域下的应用场景进行了详细的阐述,将GNN与其他图学习算法的优缺点作了联系和比较.针对存在的一些问题和挑战,本文勾画了GNN的未来方向和发展趋势,最后对全文进行了全面而细致的总结.     

基于图神经网络的自闭症辅助诊断研究综述

自闭症谱系障碍是一种复杂的神经系统发展障碍疾病,截至目前其病因尚不明确。图神经网络作为非欧几里得空间深度学习的重要分支,在处理图结构数据的相关任务中取得优异表现,为医学领域的成像和非成像模式的集成提供了可能,因此利用图神经网络进行自闭症等脑部疾病神经成像诊断逐渐成为研究热点。阐述传统机器学习方法在自闭症疾病预测中应用,介绍图神经网络的基本分类,按照图中节点与边关系的建模方法,从基于人群图和基于个体图两个角度对图神经网络在自闭症辅助诊断中的应用进行梳理和分析,并归纳现有诊断方法的优劣势。根据目前基于图神经网络的自闭症神经成像诊断的研究现状,总结了脑神经科学领域辅助诊断技术面临的主要挑战和未来研究方向,对于自闭症等脑部疾病辅助诊断的进一步研究具有指导意义和参考价值。     

基于改进遗传算法和图神经网络的股市波动预测方法

针对支持向量机（SVM）、长短期记忆（LSTM）网络等智能算法在股市波动预测过程中股票评价特征选择困难及时序关系维度特征缺失的问题,为能够准确预测股票波动、有效防范金融市场风险,提出了一种基于改进遗传算法（IGA）和图神经网络（GNN）的股市波动预测方法——IGA-GNN。首先,利用相邻交易日间的时序关系构建股市交易指标图数据;其次,通过评价指标特性优化交叉、变异概率来改进遗传算法（GA）,从而实现节点特征选择;然后,建立图数据的边与节点特征的权重矩阵;最后,运用GNN进行图数据节点的聚合与分类,实现了股市波动预测。在实验阶段,所研究的股票总评价指标数为130个,其中IGA在GNN方法下提取的有效评价指标87个,使指标数量降低了33.08%。应用所提IGA在智能算法中进行特征提取,得到的算法与未进行特征提取的智能算法相比,预测准确率整体提升了7.38个百分点;而与应用传统GA进行智能算法的特征提取相比,应用所提IGA进行智能算法的特征提取的总训练时间缩短了17.97%。其中,IGA-GNN方法的预测准确率最高,相较未进行特征提取的GNN方法的预测准确率整体提高了19.62个百分点;而该方法与用传统GA进行特征提取的GNN方法相比,训练时间平均缩短了15.97%。实验结果表明,所提方法可对股票特征进行有效提取,预测效果较好。     

图知识蒸馏综述：算法分类与应用分析

图数据, 如引文网络, 社交网络和交通网络, 广泛地存在现实生活中. 图神经网络凭借强大的表现力受到广泛关注, 在各种各样的图分析应用中表现卓越. 然而, 图神经网络的卓越性能得益于标签数据和复杂的网络模型, 而标签数据获取困难且计算资源代价高昂. 为了解决数据标签的稀疏性和模型计算的高复杂性问题, 知识蒸馏被引入到图神经网络中. 知识蒸馏是一种利用性能更好的大模型(教师模型)的软标签监督信息来训练构建的小模型(学生模型), 以期达到更好的性能和精度. 因此, 如何面向图数据应用知识蒸馏技术成为重大研究挑战, 但目前尚缺乏对于图知识蒸馏研究的综述. 旨在对面向图的知识蒸馏进行全面综述, 首次系统地梳理现有工作, 弥补该领域缺乏综述的空白. 具体而言, 首先介绍图和知识蒸馏背景知识; 然后, 全面梳理3类图知识蒸馏方法, 面向深度神经网络的图知识蒸馏、面向图神经网络的图知识蒸馏和基于图知识的模型自蒸馏方法, 并对每类方法进一步划分为基于输出层、基于中间层和基于构造图知识方法; 随后, 分析比较各类图知识蒸馏算法的设计思路, 结合实验结果总结各类算法的优缺点; 此外, 还列举图知识蒸馏在计算机视觉、自然语言处理、推荐系统等领域的应用; 最后对图知识蒸馏的发展进行总结和展望. 还将整理的图知识蒸馏相关文献公开在GitHub平台上, 具体参见: https://github.com/liujing1023/Graph-based-Knowledge-Distillation.     

Multivariate time series prediction of complex systems based on graph neural networks with location embedding graph structure learning

Graph convolutional neural networks (GNNs) have an excellent expression ability for complex systems. However, the smoothing hypothesis based GNNs have certain limitations for complex process industrial systems with high dynamics and noisy environment. In addition, it is difficult to obtain an accurate information about the interconnections of sensor networks in manufacturing systems, which brings challenges to the application of GNNs. This paper introduces a graph convolution filter with a serial alternating structure of low-pass filter and high-pass filter to alleviate the problem of node feature loss. Furthermore, we propose a simple and effective method to learn graph structure information during training. This method combines the advantages of graph structure learning based on metric method and direct optimization method. Finally, a spatiotemporal parallel feature extraction framework for multivariate time series prediction is constructed. Experiments are carried out on real industrial datasets, and the results demonstrate the effectiveness of the model.     

图匹配技术研究

图(Graph)在众多的科学领域和工程领域(如模式识别和计算机视觉)中具有广泛的应用 ,其具备 强大的信息表达能力。当图被用来表示物体结构时,衡量物体的相似程度将会被转化成计算两个图的相似度,这就是图匹配(Graph Matching)。近几十年来,对图匹配相关技术和算法的研究已经成为了研究领域内的一个重要课题,尤其是随着大数据时代的来临,图作为数据之间关系的一种表示形式,将会受到越来越多的关注。文中对图匹配技术的发展现状进行了综述,详细介绍了该技术的理论基础,梳理了解决图匹配问题的几种主流思路。最后,结合图匹配技术的一种具体应用对几种算法的性能进行了对比分析。     

A graph-based framework for sub-pixel image segmentation

Many image segmentation methods utilize graph structures for representing images, where the flexibility and generality of the abstract structure is beneficial. By using a fuzzy object representation, i.e., allowing partial belongingness of elements to image objects, the unavoidable loss of information when representing continuous structures by finite sets is significantly reduced, enabling feature estimates with sub-pixel precision.This work presents a framework for object representation based on fuzzy segmented graphs. Interpreting the edges as one-dimensional paths between the vertices of a graph, we extend the notion of a graph cut to that of a located cut, i.e., a cut with sub-edge precision. We describe a method for computing a located cut from a fuzzy segmentation of graph vertices. Further, the notion of vertex coverage segmentation is proposed as a graph theoretic equivalent to pixel coverage segmentations and a method for computing such a segmentation from a located cut is given. Utilizing the proposed framework, we demonstrate improved precision of area measurements of synthetic two-dimensional objects. We emphasize that although the experiments presented here are performed on two-dimensional images, the proposed framework is defined for general graphs and thus applicable to images of any dimension.     

KENN: 线性结构熵的图核神经网络

图神经网络(graph neural network, GNN)是一种利用深度学习直接对图结构数据进行表征的框架,近年来受到人们越来越多的关注.然而传统的基于消息传递聚合的图神经网络(messaging passing GNN, MP-GNN)忽略了不同节点的平滑速度,无差别地聚合了邻居信息,易造成过平滑现象.为此,研究并提出一种线性结构熵的图核神经网络分类方法,即KENN.它首先利用图核方法对节点子图进行结构编码,判断子图之间的同构性,进而利用同构系数来定义不同邻居间的平滑系数.其次基于低复杂度的线性结构熵提取图的结构信息,加深和丰富图数据的结构表达能力.通过将线性结构熵、图核和图神经网络三者进行深度融合提出了图核神经网络分类方法.它不仅可以解决生物分子数据节点特征的稀疏问题,也可以解决社交网络数据以节点度作为特征所产生的信息冗余问题,同时还使得图神经网络能够自适应调整对图结构特征的表征能力,使其超越MP-GNN的上界(WL测试).最后,在7个公开的图分类数据集上实验验证了所提出模型的性能优于其他的基准模型.     

Graphix a graph theory sub language

GRAPHIX is a PASCAL based graph theory sub-language. It provides a set of graph theory algorithms, and allows the user to program directly in terms of graph structures theory algorithms, and allows the user to program directly in terms of graph structures with no knowledge of the internal representation of the graph. A preprocessor has been written to convert GRAPHIX/PASCAL programs into pure PASCAL. The main problems of a sub-language and preprocessing have been overcome in the design of the language and the preprocessor. An example of a GRAPHIX/PASCAL program is provided.     

图神经网络回归的人脸超分辨率重建

人脸超分辨率（super-resolution,SR）即将输入模糊的低分辨率（low-resolution,LR）人脸图像通过一系列算法处理得到较为清晰的高分辨率（high-resolution,HR）人脸图像的过程．相比自然图像,不同人脸图像的相同位置通常具有相似的结构．本文针对人脸图像的局部结构一致性特点,提出一种新的基于图结构的人脸超分辨率神经网络回归方法．将输入低分辨率图像表示为图结构,进而为图结构中每一个结点的局部表示训练一个浅层神经网络进行超分辨率回归．相比基于规则矩形网格的方法,图结构在描述一个像素的局部信息时,不仅考虑到图像坐标的相关性,同时关注了纹理的相似性,能更好表达图像局部特征．训练过程中,利用已收敛的相邻结点的神经网络参数初始化当前结点的神经网络参数,不仅加快神经网络的收敛速度,而且提高了预测精度．与包括深度卷积神经网络在内的基于学习的超分辨率最新算法比较实验表明,本文提出的算法取得了更高的准确率．本文提出的图神经网络（Graph Neural Networks,GNN）并不局限于解决人脸超分辨率问题,它还可以用于处理其它具有不规则拓扑结构的数据,解决不同的问题．     

Satellite image classification using granular neural networks

The increased synergy between neural networks (NN) and fuzzy sets has led to the introduction of granular neural networks (GNNs) that operate on granules of information, rather than information itself. The fact that processing is done on a conceptual rather than on a numerical level, combined with the representation of granules using linguistic terms, results in increased interpretability. This is the actual benefit, and not increased accuracy, gained by GNNs. The constraints used to implement the GNN are such that accuracy degradation should not be surprising. Having said that, it is well known that simple structured NNs tend to be less prone to over‐fitting the training data set, maintaining the ability to generalize and more accurately classify previously unseen data. Standard NNs are frequently found to be accurate but difficult to explain, hence they are often associated with the black box syndrome. Because in GNNs the operation is carried out at a conceptual level, the components have unambiguous meaning, revealing how classification decisions are formed. In this paper, the interpretability of GNNs is exploited using a satellite image classification problem. We examine how land use classification using both spectral and non‐spectral information is expressed in GNN terms. One further contribution of this paper is the use of specific symbolization of the network components to easily establish causality relationships.     

Improving vector space embedding of graphs through feature selection algorithms

Graph based pattern representation offers a versatile alternative to vectorial data structures. Therefore, a growing interest in graphs can be observed in various fields. However, a serious limitation in the use of graphs is the lack of elementary mathematical operations in the graph domain, actually required in many pattern recognition algorithms. In order to overcome this limitation, the present paper proposes an embedding of a given graph population in a vector space Rⁿ. The key idea of this embedding approach is to interpret the distances of a graph g to a number of prototype graphs as numerical features of g. In previous works, the prototypes were selected beforehand with heuristic selection algorithms. In the present paper we take a more fundamental approach and regard the problem of prototype selection as a feature selection or dimensionality reduction problem, for which many methods are available. With several experiments we show the feasibility of graph embedding based on prototypes obtained from such feature selection algorithms and demonstrate their potential to outperform previous approaches.     

Towards data augmentation in graph neural network: An overview and evaluation

Many studies on Graph Data Augmentation (GDA) approaches have emerged. The techniques have rapidly improved performance for various graph neural network (GNN) models, increasing the current state-of-the-art accuracy by absolute values of 4.20%, 5.50%, and 4.40% on Cora, Citeseer, and PubMed, respectively. The success is attributed to two integral properties of relational approaches: topology-level and feature-level augmentation. This work provides an overview of some GDA algorithms which are reasonably categorized based on these integral properties. Next, we engage the three most widely used GNN backbones (GCN, GAT, and GraphSAGE) as plug-and-play methods for conducting experiments. We conclude by evaluating the algorithm’s effectiveness to demonstrate significant differences among various GDA techniques based on accuracy and time complexity with additional datasets different from those used in the original works. While discussing practical and theoretical motivations, considerations, and strategies for GDA, this work comprehensively investigates the challenges and future direction by pinpointing several open conceivable issues that may require further study based on far-reaching literature interpretation and empirical outcomes.