AiGAS-dEVL: An adaptive incremental neural gas model for drifting data streams under extreme verification latency

Nowadays, the rapid generation of data and the high costs of labeling (in most cases, reliant on human supervision) often lead to partially labeled data streams, particularly in scenarios characterized by extreme verification latency (EVL), where supervision becomes indefinitely unavailable. Additionally, streaming data can exhibit non-stationarities, known as concept drift, requiring models to incrementally adapt to evolving data patterns. This paper addresses the simultaneous occurrence of these two issues, where concept tracking and change adaptation mechanisms must operate without supervision. To this end we propose AiGAS-dEVL (Adaptive Incremental neural GAS model for drifting Streams under Extreme Verification Latency), a novel approach utilizing growing neural gas to characterize concept distributions in a data stream over time. By analyzing the behavior of prototypical representations, our method identifies changes in concept behavior and informs adaptation strategies to accommodate these shifts. Experimental results on various synthetic datasets demonstrate that AiGAS-dEVL outperforms baseline models, achieving an average prequential error improvement of 39.5% across datasets of varying drift characteristics, showcasing enhanced adaptability while maintaining a straightforward and interpretable instance-based adaptation strategy.