In the real world, various kinds of data are able to be represented as heterogeneous graph structures. Heterogeneous graphs with multi-typed nodes and edges contain rich messages of heterogeneity and complex semantic information. Recently, diverse heterogeneous graph neural networks (HGNNs) have emerged to solve a range of tasks in this advanced area, such as node classification, knowledge graphs, etc. Heterogeneous graph embedding is a crucial step in HGNNs. It aims to embed rich information from heterogeneous graphs into low-dimensional eigenspaces to improve the performance of downstream tasks. Yet existing methods only project high-dimensional node features into the same low-dimensional space and subsequently aggregate those heterogeneous features directly. This approach ignores the balance between the informative dimensions and the redundant dimensions in the hidden layers. Further, after the dimensionality has been reduced, all kinds of nodes features are projected into the same eigenspace but in a mixed up fashion. One final problem with HGNNs is that their experimental results are always unstable and not reproducible. To solve these issues, we design a general framework named Robust Feature Reinforcement (RFR) for HGNNs to optimize embedding performance. RFR consists of three mechanisms: separate mapping, co-segregating and population-based bandits. The separate mapping mechanism improves the ability to preserve the most informative dimensions when projecting high-dimensional vectors into a low-dimensional eigenspace. The co-segregating mechanism minimizes the contrastive loss to ensure there is a distinction between the features extracted from different types of nodes in the latent feature layers. The population-based bandits mechanism further assures the stability of the experimental results with classification tasks. Supported by rigorous experimentation on three datasets, we assessed the performance of the designed framework and can verify that our models outperform the current state-of-the-arts.

在现实世界中，各种数据都可以表示为异构图结构。具有多类型节点和边的异构图包含丰富的异构消息和复杂的语义信息。最近，出现了多种异构图神经网络 (HGNN) 来解决这一高级领域的一系列任务，例如节点分类、知识图谱等。异构图嵌入是 HGNN 中的关键步骤。它旨在将来自异构图的丰富信息嵌入到低维特征空间中，以提高下游任务的性能。然而，现有方法仅将高维节点特征投影到相同的低维空间，然后直接聚合这些异构特征。这种方法忽略了隐藏层中信息维度和冗余维度之间的平衡。此外，在降维之后，各种节点特征被投影到同一个特征空间中，但以混合的方式。HGNN 的最后一个问题是它们的实验结果总是不稳定且不可重现。为了解决这些问题，我们为 HGNN 设计了一个名为 Robust Feature Reinforcement (RFR) 的通用框架，以优化嵌入性能。RFR 由三种机制组成：单独映射、共同隔离和基于种群的 bandits。当将高维向量投影到低维本征空间时，分离映射机制提高了保留最多信息维度的能力。共分离机制最大限度地减少了对比损失，以确保从潜在特征层中不同类型的节点提取的特征之间存在区别。基于种群的bandits机制进一步保证了分类任务实验结果的稳定性。在对三个数据集进行严格实验的支持下，我们评估了所设计框架的性能，并可以验证我们的模型是否优于当前的最新技术水平。