Train Global, Test Local - Privacy-Preserving Learning of Cost-Effectiveness in Decentralized Systems.

The mandate of citizens for more socially responsible information systems that respect privacy and autonomy calls for a computational and storage decentralization. Crowd-sourced sensor networks monitor energy consumption and traffic jams. Distributed ledgers systems provide unprecedented opportunities to perform secure peer-to-peer transactions using blockchain. However, decentralized systems often show performance bottlenecks that undermine their broader adoption: propagating information in a network is costly and time-consuming. Optimization of cost-effectiveness with supervised machine learning is challenging. Training usually requires privacy-sensitive local data, for instance, adjusting the communication rate based on citizens' mobility. This paper studies the following research question: How feasible is to train with privacy-preserving aggregate data and test on local data to improve cost-effectiveness of a decentralized system? Centralized machine learning optimization strategies are applied to DIAS, the Dynamic Intelligent Aggregation Service and they are compared to decentralized self-adaptive strategies that use local data instead. Experimental evaluation with a testing set of 2184 decentralized networks of 3000 nodes aggregating real-world Smart Grid data confirms the feasibility of a linear regression strategy to improve both estimation accuracy and communication cost, while the other optimization strategies show trade-offs.