Distributed Classification with Dynamic Communication for Air Quality Sensing

A body of work has focused on distributed task classification in sensor networks. The majority of works in this domain focus on either decision fusion, edge/fog computing (or equivalently, data fusion), split computing (feature fusion) or fully distributed agents. These methodologies, combined with machine learning classifiers, are being applied to problems such as UAV and robot guidance, smart cities and smart grids [1]. In contrast, we propose and define a novel hybrid strategy that makes greedy local decisions at end devices as to whether to operate in a distributed or split computing context. We devise a communication strategy for distributed classification, which balances communication efficiency with classification accuracy. We demonstrate our proposed methodology on a real-world IoT dataset that consists of air-quality sensor readings taken in a home environment. Experimental results depict that our proposed approach balances communication cost and classification accuracy between known optimal values. In addition to this, our proposed strategy can achieve high accuracy at lower cost than traditional methods. We explore in detail the conditions under which our model performs well, and determine some potentially promising avenues for future extension and development.