模式识别实验室知识库

图结构是一类重要的数据类型，能够用于建模对象间复杂的交互关系。现实生活中的许多应用都可以被建模为图结构，如社交网络、化学分子结构等。在对图数据进行分析时，如何恰当地表示图数据中蕴含的信息是一个核心的问题。随着深度学习的发展，图神经网络逐渐成为图数据分析的一个标准方法。然而，大多数现有的图神经网络模型都建立在半监督或监督学习的基础上，这需要大量的有标签训练数据辅助模型的训练，而现实中，对数据进行标记往往是一项费时费力的工作。因此，如何利用丰富的无标签数据来训练图网络模型值得深入探索并且具有重要的意义。本论文围绕自监督图表达学习进行展开，针对不同类型的图数据提出不同的自监督训练策略。

论文的主要工作和创新点归纳如下：

（1）面向聚类感知的图自监督学习模型

图数据中的社群将图中具有相似功能的节点分组，其表征了图的内在语义和连接信息，因此可以作为模型训练数据的来源。基于此，本文提出一个面向聚类感知的图自监督学习框架，通过预测节点的社群标签来训练模型。之后，为了避免模型学到聚类分配的平凡解，本文提出将图学习的过程建模为一个最优传输问题，并提出一个均衡聚类策略来对社群分配结果进行规约。除此之外，为了缓解图结构中固有的连接噪声，本文提出在图表示学习的同时优化图的拓扑结构，从而学到更紧凑的隐空间，进而提高聚类算法的效果。本文在4个真实数据集上对节点分类和节点聚类两个基准任务进行了广泛的实验。实验结果证明了所提出方法的有效性。

（2）基于自适应增强的图对比学习模型

近来，对比学习在自监督图表达学习方面已经取得了不错的效果。然而，大多数现有的图对比方法都存在信息损失的问题。此外，已有工作大多利用简单的启发式图数据增强方法，忽视了数据增强对于对比学习的重要价值，在训练过程中无法保留原始数据中重要的信息。为了解决上述问题，本文提出了基于自适应数据增强的图对比学习模型。本文首先提出在节点级别上的对比学习目标函数，通过最大化数据增强视图中节点表达向量的相似度来训练模型。这一方法避免了对要求单射性质的读出函数的依赖。在此基础上，本文还提出了自适应的图数据增强方法。该方法结合了拓扑和语义两方面的先验信息。基于网络科学中的中心性度量，该方法能够使得模型学到图数据中结构上的重要连接关系和节点特征层面上关键的语义信息。本文在5个真实数据集上进行的实验表明，该方法优于目前现有的无监督训练模型。值得注意的是，在部分数据集上本文方法甚至超过了几个有监督的模型。

（3）基于结构增强的异构图对比学习模型

许多实际生活中的问题都可以被建模为带有不同类型节点和边的异构图。然而，已有的图对比学习方法却不能很好地应用在异构图上。针对异构图的特点，本文提出了基于结构增强的异构图对比学习模型。首先，为了捕捉不同元路径中的丰富信息，本文提出了一个基于多视图的对比聚合损失函数，通过优化节点在不同元路径视图下的表达一致性对每个中的视图的信息进行适应性编码。此外，本文提出显式地利用节点的结构信息，通过结构表达向量来表征节点的局部结构模式，并据此提出使用结构增强的负样本挖掘策略，通过对真实、难分负样本进行加权来增强对比学习的表达能力。本文针对3个真实场景中的异构数据集进行了深入实验，实验结果证明了所提方法的有效性。

Graphs are ubiquitous in modeling the complex interaction of objects in real-world applications, ranging from social networks to chemical molecules. Therefore, representing the knowledge underneath graph-structured data has a key issue in addressing graph-based problems. With the development of deep learning, graph neural networks have become the de facto model for analyzing graph-structured data. However, most existing graph neural network models focus on semi-supervised or supervised learning, which requires access to a large number of labeled training data. In reality, it is often time-consuming and labor-intensive to collect manual annotations for training datasets. Therefore, how to leverage the abundant unlabeled data for training the graph networks has attracted a lot of attention. This thesis develops a series of models in self-supervised learning of graph neural networks. The major contribution of this thesis can be summarized as follows.

(1) Cluster-aware graph self-supervised learning

Considering that clusters group nodes sharing similar functionalities in a graph, they convey intrinsic semantic and connectivity information of graphs and thus can be used as training signals. In light of this, we propose a graph self-supervised learning framework with cluster-awareness that trains the model by predicting the community labels of the nodes. Then, to avoid downgraded solutions of cluster assignments, we model graph representation learning as an optimal transport problem and propose an equipartition strategy to regularize the cluster assignments. After that, in order to alleviate the inherent noise in the graph structure, we propose to optimize the graph structure along with graph representation learning, which results in a more compact latent space for representations. We have conducted extensive experiments on node classification and node clustering are conducted using four real datasets. The results demonstrate the effectiveness of the proposed method.

(2) Graph contrastive learning with adaptive graph augmentation

Recently, contrastive learning has emerged as a promising method for self-supervised graph representation learning. Although promising performance has been achieved, most existing contrastive schemes for graph data suffer from information loss. Also, they leverage simple heuristic graph augmentation schemes, which fail to preserve important information during training. To address the aforementioned problem, we propose a graph contrastive learning with adaptive graph augmentation. Specifically, we propose a novel node-level contrastive objective that trains the model by maximizing the similarity node-level representations in augmented views and eschews the need of overly restricted injective readout functions. In addition, we propose an adaptive augmentation module that incorporates various graph priors from topological and semantic aspects. Based on the centrality measure in network science, important connectivities at the structure level and key semantic information at the node feature level are highlighted during training. Experiments conducted on five real datasets show that our proposed method outperforms existing unsupervised models and sometimes even surpasses several of supervised counterparts.

(3) Structure-enhanced heterogeneous graph contrastive learning

Many practical application problems can be modeled as heterogeneous graphs that involve various types of nodes and connections. To enable self-supervised learning on such heterogeneous graphs, we propose a contrastive learning model with structure-enhanced hard negative mining. Firstly, to capture the rich information in different metapaths, we propose a multi-view contrastive aggregation objective that can ensure global consistency among metapath-induced views and adaptively encode information from each view. In addition, we advocate the explicit use of structure embedding, which enriches the model with local structural patterns and propose a structure-enhanced negative mining strategy, so as to better mine true and hard negatives for contrastive objectives. Empirical studies on three real real-world heterogeneous datasets verify the effectiveness of the proposed method.