A Nature-Inspired Concept Drift Adaptation Method for Industrial Data Stream Regression

The industrial domain is characterized by many regression problems. Machine learning (ML) solutions to such problems that perform well and are robust against concept drift (CD) under real-world conditions are scarce. This paper proposes a novel model-agnostic method for ensembles, inspired by principles from evolutionary biology. It leverages these to preserve fit and significant concepts from the data stream and to constantly consider newly emergent, potentially short-lived concepts to challenge aforementioned ones. Introduced as dynamic member rotation and weight update algorithmics, this aids to robustly and explicitly adapt to various CD types as well as balance statistic modeling stability and plasticity. The method outperforms popularly employed state-of-the-art approaches on two real-world case studies by a considerable margin. With its capability of respecting practically feasible resource constraints, economizing ground truth consumption and enabling the straightforward use of tried-and-tested ML models, it is particularly valuable for productive use in data streams.