From Point to Batch: Advancing Incremental Clustering with Mass-based Dissimilarity

Abstract

Non-incremental clustering algorithms (NICLAs) dealing with dynamic data suffer from issues related to their compute-intensive behaviour. A plausible approach to address these issues lie in transforming the NICLAs into intelligent (incremental) methods that are capable of processing dynamic data. MBSCAN is one such robust NICLA that leverages the idea of data-dependent dissimilarity to find clusters. An incremental version of MBSCAN namely 𝑖𝑀𝑎𝑠𝑠 was designed to handle single-point insertions efficiently. However, due to increase in number of insertions made towards 𝑖𝑀𝑎𝑠𝑠, the algorithm tends to degenerate performance-wise for larger datasets. To address these challenges, we propose a batch-incremental version of MBSCAN known as B𝑖𝑀𝑎𝑠𝑠 (Batch 𝑖ncremental 𝑀𝑎𝑠𝑠-based clustering). Experiments conducted on multiple datasets (real and synthetic) have aptly demonstrated the effectiveness of B𝑖𝑀𝑎𝑠𝑠 over both MBSCAN and 𝑖𝑀𝑎𝑠𝑠. We also provided theoretical perspective about individual scenarios arising out of our proposed approach.