Abstract Just as nature’s most cryptic organisms continuously perceive and finely adjust to changes in their surroundings to ensure survival, the proposed model–Crypsis–proactively identifies subtle and hidden shifts within data structures and patterns, adapting in real time to preserve stability and sustain optimal model performance. In today’s ever-evolving landscape of data streams, where information continuously flows from diverse and dynamic sources, we face the persistent challenge of non-stationarity. The swift production of data parallels the unpredictability surrounding the evolving statistical characteristics of the target variable, which our model endeavors to forecast as they dynamically shift over time in unforeseen ways. Model drift, the decline in predictive accuracy due to changing environments and variable relationships, looms. This paper presents an innovative approach designed to tackle the challenge of model drift by focusing on its main underlying causes, including concept drift, concept evolution, and label drift. To evaluate the effectiveness of Crypsis, a series of experiments were carried out, and the results demonstrated steady improvements across both synthetic and real-world benchmark datasets. Overall, the findings show that our method substantially surpasses previous state-of-the-art studies addressing this issue.
Yanni et al. (Tue,) studied this question.