Context Abstraction to Improve Decentralized Machine Learning in Structured Sensing Environments

Massinissa Hamidi, Aomar Osmani

Abstract: In Internet of Things applications, data generated from devices with different characteristics and located at different positions are embedded into different contexts. This poses major challenges for decentralized machine learning as the data distribution across these devices and locations requires consideration for the invariants that characterize them, e.g., in activity recognition applications, the acceleration recorded by hand device must be corrected by the invariant related to the movement of the hand relative to the body. In this article, we propose a new approach that abstracts the exact context surrounding data generators and improves the reconciliation process for decentralized machine learning. Local learners are trained to decompose the learned representations into (i) universal components shared among devices and locations and (ii) local components that capture the specific context of device and location dependencies. The explicit representation of the relative geometry of devices through the special Euclidean Group SE(3) imposes additional constraints that improve the decomposition process. Comprehensive experimental evaluations are carried out on sensor-based activity recognition datasets featuring multi-location and multi-device data collected in a structured sensing environment. Obtained results show the superiority of the proposed method compared with the advanced solutions.