【KDD 2020】 HGCN: A Heterogeneous Graph Convolutional Network-Based Deep Learning Model Toward Collec

背景挑战

Collective classification [9, 25] aiming to collectively categorize the inter-connected entities in networked data, plays an important
role in the various in-depth analysis tasks [3, 23, 31], such as genre identification in a movieLens network, social circle learning and software delay prediction, etc.

集合分类[9,25]旨在对网络数据中相互连接的实体进行集中分类，具有重要的意义,参与各种深度分析任务[3,23,31]，如电影网络中的体裁识别、社交圈学习和软件延迟预测等。

In the meantime, heterogeneous information networks (HINs) have received increasing attention in the past decade due to its capability to represent rich and complex semantics, from which the latent relations can be inferred to improve the performance of collective classification.

同时，异构信息网络（HINs）以其丰富而复杂的语义表示能力，从中可以推断出潜在的关系，从而提高集合分类的性能，近年来受到了越来越多的关注。

However, owing to the heterogeneity of HINs, collective classification is confronting several severe challenges, e.g., complex relational hierarchy, potential incompatible semantics and node-context relational semantics. Thus, the collective classification is still a challenging topic in HINs at present.

然而，由于HINs的异构性，集体分类面临着复杂的关系层次、潜在的不相容语义和节点-上下文关系语义等严峻的挑战。因此，在目前的HINs中，集合分类仍然是一个具有挑战性的课题。

Previously, most researchers focus on conventional relational models [16, 17, 20, 21], wherein the relational features depend heavily on the experts’ empiricism by design. Thereby, it is inevitable to introduce noise/partial relational features, furthermore, these relations are usually too shallow to learn the complex/deep relational features.

以前，大多数研究者关注传统的关系模型[16,17,20,21]，其中关系特征在很大程度上依赖于专家的经验论设计。因此，不可避免地引入噪声/部分关系特征，而且这些关系通常太浅，无法学习复杂/深层关系特征。

However, many real-world applications involve complex dependencies among the entities, such as colleague relationships
in bibliographic networks, and so forth.

然而，许多实际应用程序涉及实体之间复杂的依赖关系，例如同事关系在书目网络中，等等。

Take the bibliographic network in Fig. 1 as an example. Suppose author A is the co-author of both author B and author C. If we only consider the co-author relation, author B and author C should be close to author A in the metric space. But if we consider the latent relations, such as colleague relation (i.e., two authors sharing similar co-authors) between author A and author C, author A and author C should be closer to each other and assigned the same label.

As a result, an effective model for relational data should be able to capture the dependencies among different entities for collective classification.

因此，一个有效的关系数据模型应该能够捕获不同实体之间的依赖关系，以便进行集体分类。

问题现状

Taking the shallow dependencies into account, many graph convolution related works [4, 12, 30] have been recently proposed for deep learning applications. They in essence extended the deep neural networks to deal with arbitrary graph-structured data and offer automatic end-to-end feature learning for entities.

考虑到浅层依赖性，最近有许多与图卷积相关的工作被提出用于深度学习应用。它们本质上扩展了深层神经网络来处理任意图结构的数据，并为实体提供了端到端的自动特征学习。

1) Unfortunately, the weakness lies in that these works only tackle homogeneous networks and learn the simple relational features. They have not yet been extended into the applications of extracting deep relational features in HINs.

不幸的是，这些工作的缺点在于这些工作只处理同质网络和学习简单的关系特性。它们还没有被扩展到在hin中提取深层关系特征的应用中。

2) To solve the above weakness, a few scholars proposed deep relational feature learning schemes [23, 32, 33] to performd collective classification in HINs. Nevertheless, these works learned the deep relational features on the top of multiple homogeneous networks constructed via a set of meta-paths [28], which often require users to specify the meta-paths or leverage supervision to make the meta-path selection. However, a set of meta-paths specified or selected in this way often only reflects certain aspects of the HIN. As a result, they are not always capable of transcribing the rich and complex correlations among entities in HINs comprehensively.

为了解决上述不足，一些学者提出了深层关系特征学习方案[23,32,33]来进行HIN中的集体分类。尽管如此，这些著作学习了通过一组元路径构建的多个同质网络之上的深层关系特征[28]，这些路径通常要求用户指定元路径或利用监督来选择元路径。然而，以这种方式指定或选择的一组元路径通常只反映HIN的某些方面。因此，它们并不总是能够全面地转录HINs中实体之间丰富而复杂的关联.

3) Besides, these works discard all  intermediate nodes along the meta-path by only considering two end nodes, which makes much fine-grained information (e.g., the fine-grained semantics of different types of nodes) lost, resulting in the over-smoothed relational features and indistinguishable clusters. In addition, although  the results of HIN embedding methods [5, 6, 8, 26] can be used to classify the entities with a supervised classifier, they may generate the similar representations for vertices from different classes due to the sparsity and noisiness of real-world HINs.

这些方法只考虑两个终端节点，从而丢弃了元路径上的所有中间节点，使得大量的细粒度信息（如不同类型节点的细粒度语义）丢失，导致关系特征过于平滑，聚类不易区分。此外，虽然HIN嵌入方法的结果[5,6,8,26]可以用于利用监督分类器对实体进行分类，由于现实世界HINs的稀疏性和噪声性，可以对不同类别的顶点产生相似的表示。

核心问题

1）The incompatibility among latent relations. As shown in Fig. 2, a HIN usually involves a multi-relational hierarchy, where the complex relation, like the colleague relation, can be inferred from the bottom-up learning using lower-layer relations. In other words, with the complexity of HINs increasing, the number and depth of latent relations typically increase to a great extent. However, due to the potential incompatible semantics among these relations, not all relations can benefit the collective classification. For example, the co-author relation can provide finer-grained information while the co-participant relation may introduce much noise. This is actually not surprising because a conference usually contains cross-research areas, which indicates the attendees may have somewhat different labels; By contrast, the co-authors of one paper are more possible
to focus on the same research topic. Such incompatibility among relations would degenerate the performance of collective classification. Hereby, how to correctively model the incompatibility among sophisticated hierarchical relations in HINs is a challenge for collective classification.

1） 潜在关系之间的不相容性。如图2所示，HIN通常涉及多关系层次，其中复杂关系，如同事关系，可以使用较低层关系从自底向上的学习中推断出来。换言之，随着hin复杂度的增加，潜在关系的数量和深度通常会大大增加。然而，由于这些关系之间潜在的语义不相容，并不是所有的关系都有利于集体分类。例如，合作作者关系可以提供更细粒度的信息，而共同参与者关系可能会引入很多噪音。这其实并不奇怪，因为一个会议通常包含交叉研究领域，这表明与会者可能有一些不同的标签；相比之下，一篇论文的合著者更有可能专注于同一个研究课题。这种关系之间的不相容性会降低集体分类的性能。因此，如何正确地建模HINs中)复杂层次关系之间的不相容性是一个挑战。

2) The node-context relational semantics. Due to the heterogeneity, a HIN often carry immensely rich and complex relational
semantics, such as the different effects of co-participant relations in collective classification stemming from the finer-grained
professional domains in a conference. In other words, compared with a cross-research conference, the authors attending a highlyprofessional conference may be more possible to have the same label, namely the conferences (nodes) have differential academic contexts. Hereby, how to effectively preserve such node-context relational semantics is another challenge for collective classification.

节点上下文关系语义。由于异构性，一个HIN常常携带非常丰富和复杂的关系语义方面，比如共同参与者关系在集体分类中的不同效果来源于细粒度会议中的专业领域。换言之，与交叉研究会议相比，参加高专业会议的作者更可能具有相同的标签，即会议（节点）具有不同的学术背景。因此，如何有效地保存这种节点上下文关系语义是集合分类的另一个挑战。

方法

Fig. 4 exhibits the architecture for the collective classification execution in HINs. For a target-type node, the model works in three steps: i) employ a heterogeneous graph convolution to aggregate the features of its K-order neighbors in each adjacency matrix, which is constructed by the one-type edges; ii) in each graph inception layer, collect all the features for the given node from different adjacency matrices, subsequently integrate the features as the input for the next layer; iii) select appropriate features for the given node from different-order and different-type neighbors by minimizing the cross-entropy between the predicted label and real label.