Neighbour-Aware Cooperation For Semi-Supervised Decentralized Machine Learning

Abstract

Mobile and embedded devices generate vast data, driving interest in decentralized machine learning (DML) for collaborative model training. However, current DML frameworks assume fully annotated data, limiting their applicability in IoT scenarios. We introduce semi-supervised DML, where workers possess partially labelled data within a device-to-device (D2D) network. Existing semi-supervised techniques overlook D2D topology, hindering effective utilization of unlabelled data across workers. To address this, we propose SSD, a framework for semi-supervised DML leveraging D2D cooperation. SSD's key innovation lies in strategic neighbour selection, balancing pseudo-label quality and communication overhead. Workers autonomously select neighbours with high-quality models and similar data distributions, enhancing pseudo-label confidence. Empirical evaluations in real and simulated environments demonstrate SSD's superiority over existing methods, highlighting its efficacy in exploiting D2D cooperation for improved semi-supervised learning in decentralized settings.