Adaptive clustering using enhanced DBSCAN: a dynamic approach to optimizing density-based clustering

Abstract

Clustering is a critical unsupervised learning technique for identifying patterns and structures in data. Traditional algorithms, such as density-based clustering non-parametric algorithm (DBSCAN), struggle with datasets characterized by varying densities, overlapping features, and noise, leading to suboptimal clustering quality. To address these limitations, this study introduces an Enhanced Adaptive DBSCAN (ADBSCAN) algorithm that dynamically adjusts the epsilon parameter and leverages silhouette score validation to improve cluster quality. The algorithm was tested on three benchmark datasets representing varying complexities. Findings showed that Enhanced ADBSCAN could find significant clusters, especially in datasets with modest feature overlap. However, datasets with substantial overlap and high-density variations presented difficulties. The findings demonstrate how important parameter selection is and how adaptive techniques that dynamically modify these parameters in response to data properties can greatly improve clustering performance on a variety of data sets. Future studies should concentrate on enhancing adaption mechanisms to better manage overlapping features and varying data density, enhancing the algorithm's resilience and practicality.