CGD：超强的跨视图融合分类网络

Title: CGD: Multi-View Clustering via Cross-View Graph Diffusion

Summary

• 现有的多视图处理模型都是先进行表征学习，通过学习得到的表征得到统一的图，再利用该图进行谱聚类。本文考虑将特征通过kNN构图得到每个视图的图，再通过多视图融合迭代公式进行融合扩散。这样，信息是在多个视图中进行扩散的，因此可以学习得到多个视图之间的互补和公共信息。

Problem Statement

• 现有的基于表征学习的多视图处理方法存在诸多挑战，例如：模型庞大、计算量高、需要调参等。
• 图扩散仅仅在单个视图上有应用，现在还没有人将图扩散应用到多个视图中。

Method

• 归一化的扩散流程，可以写成如下优化形式：
  
• 该问题的闭式解为：
  
• 单个视图的扩散公式为：
  
• 将单个视图的扩散公式扩展到多个视图上：
  
• 新的构图方法：
  
• 计算自适应的可变参数，用于权衡原始结构和扩散结构：
  
• 最终将扩散后的多个视图图结构直接求和作为最终的图结构，模型图如下：
  

Evaluation

• 在七个聚类的指标上进行实验评估，多视图数据集如下：
  

Conclusion

• 论文将单个图的扩散公式扩展到多个视图上，取得了显著的效果。

Notes

• Graph based multi-view clustering has been paid great attention by exploring the neighborhood relationship among data points from multiple views.
• Extensive experiments on several benchmark datasets are conducted to demonstrate the effectiveness of the proposed method in terms of seven clustering evaluation metrics.
• It is not uncommon that an object is usually described by multi-view features.
• Multi-view clustering, which partitions these multi-view data into different groups by using the complementary information of multi-view feature sets to ensure that highly similar instances are divided into the same group, is an important branch of multi-view learning.
• In general, most of previous multi-view clustering methods employ graph-based models since the similarity graph can characterize the data structure effectively.

References

• Regularized diffusion process for visual retrieval.
• Diffusion processes for retrieval revisited.
• Mpgraph: multi-view penalised graph clustering for predicting drug-target interactions.
• Late fusion incomplete multi-view clustering.
• Parameter-free auto-weighted multiple graph learning: a framework for multi-view clustering and semi-supervised classification.
• Learning a joint affinity graph for multiview subspace clustering.
• Gmc: graph-based multi-view clustering.
• Graph learning for multiview clustering.
• Multiview consensus graph clustering.