BiMass: Towards Batch-incremental Clustering using Mass-based Dissimilarity

Abstract

Non-incremental clustering algorithms (NICLAs) dealing with dynamic data may suffer from various performance issues due to its compute-intensive behavior. MBSCAN is one such robust NICLA, the point-wise incremental version (𝑖𝑀𝑎𝑠𝑠) of which was designed to handle such issues. However, with increase in the number of insertions, 𝑖𝑀𝑎𝑠𝑠 gradually degenerates its performance for larger datasets. To address these issues, we propose a batch-incremental version of MBSCAN namely B𝑖𝑀𝑎𝑠𝑠 (Batch 𝑖ncremental 𝑀𝑎𝑠𝑠 based clustering). Experiments conducted over nine real world and synthetic datasets showed the effectiveness of B𝑖𝑀𝑎𝑠𝑠 over both MBSCAN and𝑖𝑀𝑎𝑠𝑠. We theoretically validated our claim addressing individual scenarios arising out of the proposed approach.