Satrap: Data and Network Heterogeneity Aware P2P Data-Mining Authors: Hock Hee Ang, Vivekanand Gopalkrishnan, Anwitaman Datta, Wee Keong Ng, Steven C. H. Hoi Keywords: Distributed classification, P2P network, cascade SVM
Abstract:
Distributed classification aims to build an accurate classifier by learning from distributed data while reducing computation and communication cost. A P2P network where numerous users come together to share resources like data content, bandwidth, storage space and CPU resources is an excellent platform for distributed classification. However, two important aspects of the learning environment have often been overlooked by other works, viz., 1) location of the peers which results in variable communication cost and 2) heterogeneity of the peers’ data which can help reduce redundant communication. In this paper, we examine the properties of network and data heterogeneity and propose a simple yet efficient P2P classification approach that minimizes expensive inter-region communication while achieving good generalization performance. Experimental results demonstrate the feasibility and effectiveness of the proposed solution.
Among the other claims for Satrap:
- achieves the best accuracy-to-communication cost ratio given that data exchange is performed to improve global accuracy.
- allows users to control the trade-off between accuracy and communication cost with the user-specified parameters.
I find these two the most interesting.
In part because semantic integration, whether explicit or not, is always a question of cost ratio and tradeoffs.
It would be refreshing to see papers that say what semantic integration would be too costly with method X or that aren’t possible with method Y.