AA-DBSCAN: an approximate adaptive DBSCAN for finding clusters with varying densities

Clustering is a typical data mining technique that partitions a dataset into multiple subsets of similar objects according to similarity metrics. In particular, density-based algorithms can find clusters of different shapes and sizes while remaining robust to noise objects. DBSCAN, a representative density-based algorithm, finds clusters by defining the density criterion with global parameters, $$ \varepsilon $$ε-distance and $$ MinPts $$MinPts. However, most density-based algorithms, including DBSCAN, find clusters incorrectly because the density criterion is fixed to the global parameters and misapplied to clusters of varying densities. Although studies have been conducted to determine optimal parameters or to improve clustering performance using additional parameters and computations, running time for clustering has been significantly increased, particularly when the dataset is large. In this study, we focus on minimizing the additional computation required to determine the parameters by using the approximate adaptive $$ \varepsilon $$ε-distance for each density while finding the clusters with varying densities that DBSCAN cannot find. Specifically, we propose a new tree structure based on a quadtree to define a dataset density layer. In addition, we propose approximate adaptive DBSCAN (AA-DBSCAN) and kAA-DBSCAN that have clustering performance similar to those of existing algorithms for finding clusters with varying densities while significantly reducing the running time required to perform clustering. We evaluate the proposed algorithms, AA-DBSCAN and kAA-DBSCAN, via extensive experiments using the state-of-the-art algorithms. Experimental results demonstrate an improvement in clustering performance and reduction in running time of the proposed algorithms.