1、Abstract
作者想解决什么问题,method是什么,answer是什么
2、Introduction
为什么要研究,该研究进行到哪一阶段,理论依据是什么
3、Conclusion
未来的研究着重点在哪?有无可改进的方法
4、Table Or Fig
直观捕捉有用信息
5、 Related work
之前的研究进展如何
6、Algorithm
尽可能理解每一步表示什么
1. Abstract
Question:Spectral clustering algorithm‘’s performance may degrade due to noise existing in each individual view or inconsistencies between heterogeneous features.
Method:We propose to fuse multi-view information in a partition space ,which enhances the robustness of Multi-view clustering.
1) generate multiple partitions and integrate them to find a shared partition.
2) unifies graph learning,generation of basic partitions ,and view weight learing.
2. Introduction
In many real-world problems,data are collected from different sources.To process these kinds of data , a number of multi-view learning algorithms have been developed.
多视图方法的两种基本假设:consistency and complementarity(共识性和互补性)
to achieve a better performance by integrating compatible and complementary information from different views.
近些年多视图聚类算法可以分为以下几类:(详细介绍)
1)co-training style algorithms
2) kernel-based methods
3) graph-based methods
4) subspace clustering based methods
在实际生活中数据是存在噪音或存在缺失,这就有了一个关键问题图学习是容易受这些因素的影响,从而导致结果质量不佳,鲁棒性较弱。
We propose to fuse partitions to improve the model robustness.We adaptively learn a weight for each view to control its contribution to the final clustering.The final clustering is achieved through a purposely designed weighting mechanism imposed on basic partitions.
Main contributions:
1) propose to fuse multi-view information in a partition space.
2) model which iteratively learns a graph for each view, a partition for each view, a weight for each partition, and a consensus clustering.
3) 实验结果表明该模型是有效率的,在有噪声的数据下鲁棒性好。
3. Conclusion
1. 提出了一种novel的多视图子空间聚类方法,旨在探索分区空间中的多视图信息以此来增强模型的鲁棒性。
2. 模型把图学习、谱聚类、加权学习和分区融合整合到一个统一的框架,每一部分都是调优的。
3. 实验结果也表明了该模型的有效性和鲁棒性
4. Proposed methodology
与MVSC相比,本方法将图的构建与谱聚类结合在一起。每个视图都生成一个分区。Multiple partitions allow us to manipulate multi-view information in partition space,which enhances the model robustness.
![在这里插入图片描述](https://img-blog.csdnimg.cn/20210712083859700.png
由于噪声与各种特征的影响,在一般情况下分区矩阵是不容易被识别的为此引入了加权机制,
w的值越大表明这个分区对共识聚类的贡献越大
分区融合的目标函数
Y是共识聚类的指示矩阵
Partition Fusion multi-view Subspace Clustering(PFSC) can be formulated as :
输入数据X,PFSC会将聚类的指示矩阵输出,因此PFSC是一个端到端的聚类。
S,F,w,Y的优化
将原始问题分为四个子问题,采用坐标轮换的方法,优化其中一个的时候,将其他三个固定。
算法如下:
最后附上论文地址:https://www.sciencedirect.com/science/article/pii/S0020025521000712
代码地址:https://github.com/LyuJC/PFSC