Incomplete multi-modal clustering (IMmC) is challenging due to the unexpected missing of some modalities in data. A key to this problem is to explore complementarity information among different samples with incomplete information of unpaired data. Despite preliminary progress, existing methods suffer from (1) relying heavily on paired data, and (2) difficulty in mining complementarity on data with high missing rates. To address the problems, we propose a novel method, Integrated Heterogeneous Graph ATtention (IHGAT) network, for IMmC. To fully exploit the complementarity among different samples and modalities, we first construct a set of integrated heterogeneous graphs based on the similarity graph learned from unified latent representations and the modality-specific availability graphs formed by the existing relations of different samples. Thereafter, the attention mechanism is applied to the constructed integrated heterogeneous graph to aggregate the embedded content of heterogeneous neighbors for each node. In this way, the representations of missing modalities can be learned based on the complementarity information of other samples and their other modalities. Finally, the consistency of probability distribution is embedded into the network for clustering. Consequently, the proposed method can form a complete latent space where incomplete information can be supplemented by other related samples via the learned intrinsic structure. Extensive experiments on eight public datasets show that the proposed IHGAT outperforms existing methods under various settings and is typically more robust in cases of high missing rates.

由于数据中某些模态的意外缺失，不完全多模态聚类 (IMmC) 具有挑战性。该问题的关键是利用未配对数据的不完整信息来探索不同样本之间的互补信息。尽管取得了初步进展，现有方法仍面临以下问题：（1）严重依赖配对数据，（2）难以挖掘高缺失率数据的互补性。为了解决这些问题，我们提出了一种用于 IMmC 的新方法，即集成异构图注意力（IHGAT）网络。为了充分利用不同样本和模态之间的互补性，我们首先基于从统一潜在表示中学习到的相似性图和由不同样本的现有关系形成的模态特定可用性图构建一组集成异构图。此后，将注意力机制应用于构建的集成异构图，以聚合每个节点的异构邻居的嵌入内容。这样，可以根据其他样本及其其他模态的互补信息来学习缺失模态的表示。最后将概率分布的一致性嵌入到网络中进行聚类。因此，所提出的方法可以形成一个完整的潜在空间，其中不完整的信息可以通过学习的内在结构由其他相关样本来补充。对八个公共数据集的广泛实验表明，所提出的 IHGAT 在各种设置下都优于现有方法，并且在高缺失率的情况下通常更稳健。