NewDynamicDist

NewDynamicDist is a family of adaptive probabilistic models designed to estimate probability distributions from non-stationary data streams. It extends traditional kernel density estimation with online, non-stationary components that can adapt to concept drift by adjusting bandwidths, adding or removing components, and discounting older observations.

The method maintains a dynamic mixture of kernel components. Each component has a location, scale, and weight, which are updated as new data arrive. A forgetting factor reduces the influence of older data, and a drift detector triggers structural updates such as splitting a component when local density grows or merging components when it wanes. The result is a model that can track changing distributions in real time.

Variants include NewDynamicDist-DM, a dynamic mixture version, and NewDynamicDist-NP, a nonparametric variant. The approach supports online density estimation, conditional density, and anomaly scoring, enabling real-time monitoring and decision making in streaming contexts.

Applications include sensor networks, fraud detection, financial surveillance, climate and environmental monitoring, and adaptive control systems. It is positioned as an alternative to static KDE and to Bayesian nonparametrics for non-stationary environments.

Limitations and considerations: performance depends on hyperparameters such as forgetting factor, drift thresholds, and maximum components; computational overhead is higher than static KDE; careful cross-validation or online validation is recommended to avoid overfitting or drift-induced instability.