[论文解读] Community Detection in Networks: The Leader-Follower Algorithm

Devavrat Shah, Tauhid Zaman|arXiv (Cornell University)|Nov 2, 2010

Complex Network Analysis Techniques参考文献 11被引用 51

一句话总结

本文提出了一种名为领导者-追随者算法的新型社区检测方法，通过利用网络中心性来区分领导者（跨社区连接者）与追随者（社区内节点），从而在无需先验知识的情况下自动检测社区数量。该方法通过关注社区内部结构而非外部图割，相较于谱聚类表现更优，在具有自然社交结构的网络中实现了精确的社区恢复。

ABSTRACT

Natural networks such as those between humans observed through their interactions or biological networks predicted based on various experimental measurements contain a wealth of information about the unobserved structure of the social or biological system. However, these networks are inherently noisy in the sense that they contain spurious connections making them seemingly dense. Therefore, identifying important, refined structures such as communities or clusters becomes quite challenging. Specifically, we find that the popular, traditional method of spectral clustering does not manage to learn refined community structure. The primary reason for this is that it is based upon external community connectivity properties such as graph-cuts. Motivated to overcome this limitation, we propose a community detection algorithm, called the leader-follower algorithm, based upon identifying the natural internal structure of the expected communities. The algorithm uses the notion of network centrality in a novel manner to differentiate leaders (nodes which connect different communities) from loyal followers (nodes which only have neighbors within a single community). Using this approach, it is able to learn the communities from the network structure. A salient feature of our algorithm is that, unlike the spectral clustering, it does not require knowledge of number of communities in the network; it learns it naturally. We show that our algorithm is quite effective. We prove that it detects all of the communities exactly for any network possessing communities with the natural internal structure expected in social networks. More importantly, we demonstrate its effectiveness in the context of various real networks ranging from social networks such as Facebook to biological networks such as an fMRI based human brain network. 1

研究动机与目标

• 解决在社交网络和生物网络等固有噪声网络中检测精细社区结构的挑战。
• 克服谱聚类的局限性，后者依赖于图割等外部连接度量，无法恢复细粒度的社区结构。
• 开发一种基于自然内部结构识别社区的方法，特别是利用网络中心性区分领导者与追随者。
• 实现在无需输入社区数量的情况下自动检测社区数量。
• 在包括Facebook和基于fMRI的人脑网络在内的多样化真实世界网络中验证算法的有效性。

提出的方法

• 该算法新颖地应用网络中心性，根据节点的连接模式将其分类为领导者或追随者。
• 领导者被定义为连接多个社区的节点，而追随者则是其邻居仅属于单一社区的节点。
• 通过迭代识别领导者，并将追随者分配给其邻居所在的社区，基于内部凝聚力形成社区。
• 该方法在执行过程中动态学习社区数量，避免了对先验指定的依赖。
• 其依赖于网络的结构性质，而非图割或模块度优化等外部度量。
• 该算法设计用于在具有社会系统中预期的自然内部社区结构的网络中实现精确的社区恢复。

实验结果

研究问题

• RQ1 通过聚焦于内部结构属性而非外部连接度量，社区检测算法能否比谱聚类更准确地识别社区？
• RQ2 一种方法如何在无需作为输入提供的情况下自动确定社区数量？
• RQ3 基于中心性的方法在多大程度上能够区分领导者与追随者，以实现精确的社区识别？
• RQ4 该算法在真实世界噪声网络（如社交网络和脑网络）中是否保持高准确性？
• RQ5 在具有自然内部社区组织结构的网络中，该算法能否精确恢复所有社区？

主要发现

• 领导者-追随者算法在具备社交网络中预期的自然内部结构的网络中，能够精确检测出所有社区。
• 与谱聚类不同，该算法不依赖于图割等外部度量，从而能够更好地恢复精细的社区结构。
• 该算法在执行过程中自动学习社区数量，消除了对先验指定的需求。
• 它在真实世界网络（包括Facebook和基于fMRI的人脑网络）中表现出色，证实了其实际有效性。
• 该方法依赖于网络内部结构，使其在存在噪声和虚假连接的情况下表现出更强的鲁棒性。