Incremental Density-Based Clustering on Multicore Processors.

Area of Science:

• Data Mining and Machine Learning
• Database Systems
• Computational Science

Background:

• Density-based clustering is crucial for data analysis but struggles with dynamic datasets.
• Reclustering from scratch is inefficient when databases frequently change.
• Existing incremental methods face challenges with update overheads and scalability.

Purpose of the Study:

• To introduce IncAnyDBC, a novel parallel incremental data clustering approach.
• To address the challenge of efficiently updating clustering results in frequently changing databases.
• To improve upon existing methods by reducing update overheads and enhancing efficiency.

Main Methods:

• IncAnyDBC processes data changes in bulks, unlike batch processing in other methods.
• It maintains an object node graph to incrementally update clusters by propagating changes.
• The approach selectively examines meaningful objects for exact or approximate clustering results under time constraints.
• Parallelization on multicore CPUs is achieved by processing objects in blocks.

Main Results:

• IncAnyDBC significantly reduces update overheads by processing changes in bulks.
• The object node graph enables efficient incremental updates with localized change propagation.
• The method achieves exact DBSCAN results or approximations efficiently, outperforming existing techniques.
• Experimental results on large datasets demonstrate IncAnyDBC's superior performance and scalability.

Conclusions:

• IncAnyDBC offers an efficient and scalable solution for incremental data clustering in dynamic environments.
• The bulk processing and object node graph strategies effectively minimize update costs.
• The parallel implementation ensures high performance on multicore architectures, making it suitable for large-scale data analysis.