densemble

Densemble, sometimes written as d-ensemble, is a term used in machine learning to denote a class of ensemble methods that adaptively combine multiple predictive models. Unlike static ensembles whose member models and their contributions are fixed during deployment, d-ensembles monitor ongoing performance and adjust the composition or weighting of base learners in response to changes in the data distribution, making them suitable for non-stationary environments and data streams.

Core ideas include dynamic selection and dynamic weighting. Dynamic selection chooses the most appropriate subset of base learners for each prediction, while dynamic weighting updates the contribution of each learner over time. Many implementations operate in an online fashion, evaluating learners on recent data, and may incorporate drift detection, error monitoring, and ensemble pruning to maintain efficiency and diversity.

Applications include real-time classification, time series forecasting, sensor data analysis, anomaly detection, and fraud detection, particularly where concept drift occurs. Design considerations include computational cost, latency, robustness to noisy data, and the balance between diversity and accuracy.

In the literature, d-ensemble is discussed as part of broader ensemble and online learning research; related concepts include dynamic ensemble selection and windowed evaluation.