Title: Tackling Label Scarcity in Supervised Deep Learning for Big IoT Data: A Review of Semi-Supervised and Weak Supervision Approaches
Authors: Virendra Tank
Volume: 10
Issue: 7
Pages: 157-164
Publication Date: 2026/07/28
Abstract:
The proliferation of Internet of Things (IoT) devices has led to an explosive growth in sensor-generated big data, creating fertile ground for supervised deep learning applications ranging from human activity recognition to network intrusion detection. However, the practical deployment of supervised deep learning in IoT settings is severely constrained by label scarcity: while raw IoT data is abundant, expert-verified ground-truth labels remain scarce, costly, and slow to obtain. This review paper systematically examines the label scarcity problem in the context of deep learning, big data, and supervised learning as applied to IoT systems. It surveys four major families of label-efficient solutions - semi-supervised learning, weak supervision, active learning, and self-training/pseudo-labeling - and analyzes their mechanisms, strengths, and limitations across representative IoT application domains including smart homes, industrial IoT, and IoT network security. A comparative analysis of reported performance gains is presented, followed by a discussion of open challenges such as pseudo-label noise, heterogeneous sensor data, and federated learning constraints. The paper concludes with future research directions, including federated semi-supervised learning, foundation-model-assisted labeling, and edge-native label-efficient architectures for IoT big data analytics.