基于重构残差保持的深度矩阵分解的多视图聚类

Multi-View Clustering Based on Deep Matrix Factorization with Local Residual Preserving

摘要: 多视图聚类是机器学习中的一个基本问题，近年来，基于矩阵分解用于学习多视图数据获得广泛的关注，并且取得了不错的效果。然而，我们发现现有的多视图聚类方法大多主要关注多视图数据所蕴含的重要属性，即一致性和互补性，而忽略了属于单个视图的特定统计属性。为此，本文提出一个矩阵深度矩阵分解的局部残差保持的多视图聚类算法(LRPDMF)，我们通过图嵌入的方式，捕捉到每个视图的特定统计属性，保证每一层的数据重构，相似的数据点具有相似的重建残差。在融合阶段的，我们利用这些学习到的特征，由于不同的视图可能会有不同的权重，我们利用自适应的方式将他们逼近到一个相似低维空间。在不同的数据集上，我们的LRPDMF与现在最先进的方法比较，获得不错的实验效果。

Abstract: Multi-view clustering is a fundamental problem in machine learning, and in recent years, the use of matrix factorization-based for learning multi-view data has gained widespread attention and achieved good results. However, we find that most of the existing multi-view clustering methods mainly focus on the important attributes embedded in multi-view data, i.e., consistency and complementarity, while ignoring the specific statistical properties belonging to individual views. To this end, in this paper, we propose a local residual preserving multi-view clustering algorithm with deep matrix factorization (LRPDMF), where we capture the specific statistical properties of each view by means of graph embedding, which ensures that the data is reconstructed at each layer and similar data points have similar reconstructed residuals. In the fusion phase, we utilize these learned features, and since different views may have different weights, we approximate them to a similar low-dimensional space using an adaptive approach. On different datasets, our LRPDMF obtains good experimental results when compared with the state-of-the-art methods.