Information Technology Reference
In-Depth Information
systems distribute the physical memory among the nodes but provide a shared
global address space on each processor.
Fig. 12.4. Parallel and distributed association rules mining algoithms(Zaki, 1999)
Figure 12.4 shows where each parallel association rules mining methods fall
in the design space(Zaki, 1999). Distributed memory methods form the dominant
platform, and a mix of data- and task- parallel approaches have been explored.
However, all schemes use static load balancing, or very limited dynamic load
balancing. The main design issues in distributed memory methods are minimizing
communication and evenly distributing data for good load balancing.
David Cheung and his colleagues proposed the Fast Distributed Mining
(FDM) algorithm for association rules (Cheung et al, 1996). The main difference
between parallel and distributed data mining is the interconnection network
latency and bandwidth. David Cheung and his colleagues discover that there exist
valuable properties between local and global large data sets. One should
maximally take advantages of such properties to reduce the number of messages
to be passed and confine the substantial amount of processing to local sites. Fast
Distributed Mining is developed three editions: FDM-LP, FDM-LUP, FDM-LPP.
They all had similar structure, but different pruning algorithms. FDM-LP only
