A distributed asynchronous and privacy preserving neural network ensemble selection approach for peer-to-peer data mining

• Authors: Yiannis Kokkinos;Konstantinos G. Margaritis
• Venue: Proceedings of the Fifth Balkan Conference in Informatics
• Year: 2012

This work describes a fully asynchronous and privacy preserving ensemble selection approach for distributed data mining in peer-to-peer applications. The algorithm builds a global ensemble model over large amounts of data distributed over the peers in a network, without moving the data itself, and with little centralized coordination. Only classifiers are transmitted to other peers. Here the test set from one classifier is the train set of the other and vice versa. Regularization Networks are used as ensemble member classifiers. The approach constructs a mapping of all ensemble members to a mutual affinity matrix based on classification rates between them. After the mapping of all members the Affinity Propagation clustering algorithm is used for the selection phase. A classical asynchronous peer-to-peer cycle is continually executed for computing the mutual affinity matrix. The cycle composed of typical grid commands, like send local classifier to a peer k, check for received classifier m in the queue, compute local average positive hits, send results to peer m and send local classifier to a peer k+1. Thus the communication model used is simple point-to-point with send-receive commands to or from a single peer. The approach can also be implemented to other types of classifiers.