PDBSCAN, A Density-Based Clustering Algorithm

The DBSCAN algorithm is a popular choice for unsupervised clustering due to its ability to identify arbitrarily shaped clusters. However, its effectiveness declines when handling datasets with heterogeneous densities, significant noise, or high dimensionality, primarily due to its sensitivity to the epsilon (ε) and MinPts parameters. To overcome these challenges, this paper presents PDBSCAN, a hybrid clustering approach that incorporates noise filtering, nonlinear dimensionality reduction, adaptive clustering, and post-clustering refinement. Initial preprocessing removes outliers using Local Outlier Factor (LOF), followed by Uniform Manifold Approximation and Projection (UMAP) for structure-preserving dimensionality reduction. Clustering is then performed via HDBSCAN, which adaptively infers cluster structure without fixed parameter input. Remaining unclustered samples are assigned labels using a K-Nearest Neighbors (KNN) model, with final adjustments made through a Random Forest classifier to improve boundary consistency. Experimental evaluation on datasets including R15, D31, Flame, Compound, and Aggregation confirms that the framework achieves over 99.9% accuracy and strong ARI, NMI, and silhouette performance. The method offers scalability, resilience to noise, and is well-suited for real-world tasks in domains such as bioinformatics, anomaly detection, and geospatial analysis