Adaptive Momentum Estimation Algorithms For Non Stationary Data Stream Clustering

Abstract

Background: Real-time systems like network systems, IoT devices, and applications with intelligence produce data streams that continuously change their distribution, making traditional clustering algorithms difficult to deal with in these systems. Traditional static optimization techniques tend to be ineffective when the patterns change and/or concept drift occurs. Methodology: In this research, ocus on a new Adaptive Momentum Estimation Framework for Non-Stationary Data Stream Clustering (AME-NSC) based on the NSL-KDD dataset. It combines preprocessing, window formation in the stream, concept drift detection, adaptive momentum estimation, and dynamic cluster updates. The adaptive momentum module continually adjusts learning parameters to the variations in the stream so that it can adapt to the changing behavior of the data. Results: Experimental work was done with K-Means, CluStream, and DenStream methods. The framework proposed was able to cluster the data with an accuracy of 96.7%, F1 score of 0.96, purity score of 0.94, and silhouette coefficient of 0.89 and reduce execution time to 261 ms for larger stream sizes. The efficacy of the results is shown to be better than the existing approach with regard to the convergence speed and better cluster quality. Conclusion: In the combination of adaptive momentum estimation and concept drift handling greatly improves the stability of clustering and computation in the context of evolving streams. The proposed framework is flexible with respect to the dynamic nature of the data, and it shows good performance for non-stationary stream clustering applications.