Clustering Data in Peer-to-Peer Systems

Clustering Data in Peer-to-Peer Systems

Mei Li (Microsoft Corporation, USA) and Wang-Chien Lee (Pennsylvania State University, USA)
Copyright: © 2009 |Pages: 7
DOI: 10.4018/978-1-60566-010-3.ch041
OnDemand PDF Download:
No Current Special Offers


With the advances in network communication, many large scale network systems have emerged. Peer-topeer (P2P) systems, where a large number of nodes self-form into a dynamic information sharing system, are good examples. It is extremely valuable for many P2P applications, such as market analysis, scientific exploration, and smart query answering, to discover the knowledge hidden in this distributed data repository. In this chapter, we focus on clustering, one of the most important data mining tasks, in P2P systems. We outline the challenges and review the start-of-the-art in this area. Clustering is a data mining technique to group a set of data objects into classes of similar data objects. Data objects within the same class are similar to each other, while data objects across classes are considered as dissimilar. Clustering has a wide range of applications, e.g., pattern recognition, spatial data analysis, custom/market analysis, document classification and access pattern discovery in WWW, etc. Data mining community have been intensively studying clustering techniques for the last decade. As a result, various clustering algorithms have been proposed. Majority of these proposed algorithms is designed for traditional centralized systems where all data to be clustered resides in (or is transferred to) a central site. However, it is not desirable to transfer all the data from widely spread data sources to a centralized server for clustering in P2P systems. This is due to the following three reasons: 1) there is no central control in P2P systems; 2) transferring all data objects to a central site would incur excessive communication overheads, and 3) participants of P2P systems reside in a collaborating yet competing environment, and thus they may like to expose as little information as possible to other peers for various reasons. In addition, these existing algorithms are designed to minimize disk access cost. In P2P system, the communication cost is a dominating factor. Therefore, we need to reexamine the problem of clustering in P2P systems. A general idea to perform clustering in P2P systems is to first cluster the local data objects at each peer and then combine the local clustering results to form a global clustering result. Based on this general idea, clustering in P2P systems essentially consists of two steps, i.e., local clustering and cluster assembly. While local clustering can be done by employing existing clustering techniques, cluster assembly is a nontrivial issue, which concerns representation model (what should be communicated among peers) and communication model (how peers communicate with each other). In this chapter, we review three representation models (including two approximate representation models and an exact representation model) and three communication models (including flooding-based communication model, centralized communication model, and hierarchical communication model). The rest of this chapter is organized as follows. In next section, we provide some background knowledge on P2P systems and clustering techniques. The details of representation models and communication models are presented in Section 3. We discuss future trend and draw the conclusion in Section 4 and Section 5, respectively.
Chapter Preview


P2P Systems

Different from traditional client-server computing model, P2P systems have no central control. Each participant (peer) has equal functionality in P2P systems. Peers are autonomous and can join and leave the system at any time, which makes the systems highly dynamic. In addition, the number of peers in P2P systems is normally very large (in the range of thousands or even millions).

P2P systems display the following two nice features. First, they do not have performance bottlenecks and single points of failure. Second, P2P systems incur low deployment cost and have excellent scalability. Therefore, P2P systems have become a popular media for sharing voluminous amount of information among millions of users.

Current works in P2P systems have been focusing on efficient search. As a result, various proposals have emerged. Depending on whether some structures are enforced in the systems, existing proposals can be classified into two groups: unstructured overlays and structured overlays.

Complete Chapter List

Search this Book: