Abstract: Abstract: In practical applications, multi-view metric learning has become an effective method for handling multi-view data. However, the incompleteness of multi-view data poses a significant challenge for multi-view metric learning. Although some methods have attempted to address the issue of incomplete multi-view data, they still have the following shortcomings: i) Most methods rely on k-nearest neighbors of existing samples to fill in missing data, which can overlook the unique characteristics of samples or views; ii) They only utilize existing sample representations to calculate neighbors, which cannot fully express the neighbor relationships between samples. To address these issues, a Dual Completion-based Incomplete Multi-View Metric Learning algorithm (DIMVML) was proposed. The algorithm first extracts latent features of each view using a deep autoencoder, then fills in missing samples by combining distribution information of the samples and differences between views; secondly, it fuses the results based on the quality of the completed samples to obtain higher-quality completion results; finally, it optimizes intra-view and inter-view relationships through a loss function. In the experimental analysis, six classic datasets were selected for testing and analysis. The experimental results further demonstrate the effectiveness of the proposed new algorithm.

Key words: incomplete multi-view, metric learning, representation learning, difference, consistency

摘要： 在实际应用中，多视图度量学习成为处理多视图数据的有效方法。然而，多视图数据的不完整性给多视图度量学习带来了巨大挑战。尽管已有一些方法试图解决不完整多视图问题，但它们仍存在以下不足：i) 现有方法大多依赖于已有样本的k近邻来补全缺失数据，容易忽视样本或视图的独特特征；ii) 它们仅利用现有样本表示来计算近邻，无法充分表达样本间的近邻关系。为此，提出了基于双补全的不完整多视图度量学习算法(DIMVML)。该算法首先利用深度自编码器提取各视图的潜在特征，再结合样本的分布信息和视图间的差异信息来补全缺失样本；其次，根据补全样本的质量进行融合，以获得更高质量的补全结果；最后，通过损失函数优化视图内和视图间的关系。实验分析中选取了6个经典数据集进行测试和分析，实验结果进一步表明本文提出新算法的有效性。

关键词: 不完整多视图, 度量学习, 表示学习, 差异性, 一致性