Multi-view subspace clustering has been widely applied in practical applications. It fuses complementary information across multiple views and treats all samples of a view as a set of bases of a generalized subspace. Meanwhile, it assumes that an instance has all features corresponding to all views. However, each view may lose some features due to the malfunction, which results in an incomplete multi-view dataset. This paper presents an incomplete multi-view subspace clustering with adaptive instance-sample mapping and deep feature fusion algorithm (IMDF). Owing to the good performance of attention mechanism, we fuse deep features adaptively extracted by convolutional neural networks in a weighted way to integrate abundant information from distinctive views, reduce redundancy between features, and generalize a robust and compact representation before self-representation stage. We jointly process feature training and clustering for a specific task to weaken the sensitivity of model to pre-extracted features. At the same time, we propose a modified weighted view-specific instance-sample mapping strategy to solve the inconsistent dimensions problem induced by different numbers of samples of distinct views in the process of learning a common representation in a unified latent subspace. Experimental results demonstrate that our method outperforms five state-of-the-art methods on various real-world datasets including images and documents.

多视图子空间聚类已经在实际应用中得到了广泛的应用。它融合了跨多个视图的补充信息，并将视图的所有样本视为广义子空间的一组基础。同时，假设实例具有与所有视图相对应的所有功能。但是，由于故障，每个视图可能会丢失某些功能，从而导致不完整的多视图数据集。本文提出了具有自适应实例样本映射和深度特征融合算法（IMDF）的不完整多视图子空间聚类。由于注意力机制的良好性能，我们以加权方式融合了卷积神经网络自适应提取的深层特征，以整合来自独特视图的大量信息，减少特征之间的冗余，并在自我表示阶段之前概括出鲁棒而紧凑的表示形式。我们针对特定任务联合处理特征训练和聚类，以减弱模型对预先提取的特征的敏感性。同时，我们提出了一种改进的加权特定于视图的实例样本映射策略，以解决在学习统一的潜在子空间中的公共表示的过程中，不同视图的样本数量不同而导致的尺寸不一致问题。实验结果表明，我们的方法在包括图像和文档在内的各种实际数据集上均优于五种最新方法。我们提出了一种改进的加权特定于视图的实例样本映射策略，以解决在学习统一的潜在子空间中的公共表示的过程中，不同视图的样本数量不同而导致的尺寸不一致问题。实验结果表明，我们的方法在包括图像和文档在内的各种实际数据集上均优于五种最新方法。我们提出了一种改进的加权特定于视图的实例样本映射策略，以解决在学习统一的潜在子空间中的公共表示的过程中，不同视图的样本数量不同而导致的尺寸不一致问题。实验结果表明，我们的方法在包括图像和文档在内的各种实际数据集上均优于五种最新方法。