Green crowdsensing by learning inter-auction mappings and non-local graph constraints

In this short paper, we envisage a framework for green crowdsensing that exploits inter-auction correlations, and that incorporates label propagation on graphs using non-local constraints. Mobile crowdsensing (MCS) is a promising paradigm that relies on the pervasiveness of smart, sensor-rich devices for mapping phenomena of interest. Energy management during sensing campaigns is an open problem in building efficient MCS systems. From the perspective of the participants in sensing campaigns, it is convenient to have a platform (which orchestrates the task-participant matching process) that is energy aware, so that battery life and energy demands of various sensors are taken into consideration during task assignment. From the perspective of service demanders, energy-aware platforms are essential towards harnessing the power of MCS without compromising cost and eco-consciousness. Rather than dealing with auctions as individual and separate events in time and space, the framework envisaged in this research formulates green crowdsensing as a random field on irregular graphs, with neighborhood defined in time (consecutive auctions) and in space (auctions in different geographic regions). Then, it solves a binary optimization problem such that the edge weights of these graphs represent inter-auction mappings that can be learned from previous auctions using Long Short-Term Memory networks. These weights also represent non-local smoothness constraints, i.e., it can be established between auctions that are not necessarily neighbors in space. Using the learned inter-auction mappings, we can eliminate redundant auctions, yielding energy reductions during crowdsensing campaigns. We present complete formulations for the objective functions that govern the envisaged framework. We also share preliminary results in the form of a comprehensive dataset-the MAGGIE dataset-that we have built for model learning, as an essential step in the proposed pipeline.