Distributed privacy-preserving P2P data mining via probabilistic neural network committee machines

This work describes a probabilistic neural network (PNN) committee machine for Peer-to-Peer data mining. The pattern neurons of the PNN committee are composed of locally trained class-specialized regularization network Peer classifiers. The training takes into account the asynchronous distributed and privacy-preserving requirements that can be met in P2P systems. The Peer classifiers are first trained in parallel based on their local data. While no local data exchange is possible among them, the peers can exchange their classifiers in the form of binaries, or agents. Then an asynchronous distributed computing P2P cycle is executed to construct a mutual validation matrix. The train set of one Peer becomes the validation set of the other and only average rates are returned back. From this matrix we demonstrate that it is possible to perform weight based ensemble selection of best peer members for every class and in this way to find class-specialized Peer modules for the committee machine.