AdaptiveSSL: A Unified Semi-Supervised Learning Framework for Robust Classification via Adaptive Uncertainty Calibration and Dynamic Labeling

In industries like telecommunications and banking, where labeled data is scarce and mispredictions cost millions, semi-supervised learning (SSL) and active learning (AL) offer promising solutions but often falter due to overconfident predictions, redundant sampling, and inflexible pseudo-labeling thresholds. Traditional approaches tackle these issues separately, limiting their effectiveness in dynamic, real-world settings. This paper introduces a scalable SSL framework called AdaptiveSSL that integrates Wasserstein distance for uncertainty calibration, multi-resolution hashing for efficient diversity-driven sample selection, and Lagrangian optimization for adaptive pseudo-labeling thresholds. By iteratively refining confidence estimates, ensuring representative sampling, and dynamically adjusting thresholds, our system achieves robust churn prediction with AUC scores up to 0.9326, outperforming conventional SSL and AL methods by up to 10%, while maintaining execution times under 90 seconds on large datasets. Tested across four real-world churn datasets, this framework delivers a practical, industry-ready tool for data-efficient classification, paving the way for broader adoption in resource-constrained expert systems.