Accelerating Fuzzy-C Means Using an Estimated Subsample Size

Area of Science:

• Computer Science
• Data Mining
• Machine Learning

Background:

• Fuzzy c-Means (FCM) clustering is often accelerated using random data sampling.
• Existing methods lack statistical rigor in subsample size estimation, impacting speed and partition quality.
• The influence of subsample size on accelerated clustering performance is not well-understood.

Purpose of the Study:

• Introduce two novel accelerated Fuzzy c-Means (FCM) algorithms: GOFCM and MSERFCM.
• Incorporate a statistical method for estimating subsample size in accelerated clustering.
• Develop and evaluate a new, general stopping criterion for accelerated clustering algorithms.

Main Methods:

• Developed GOFCM, a progressive sampling variant of SPFCM, using statistical subsample size estimation.
• Developed MSERFCM, an improved initialization variant of rseFCM, with statistical subsample size estimation.
• Introduced a novel stopping criterion applicable to various accelerated clustering algorithms.

Main Results:

• GOFCM achieved 4-47x speedup over FCM, outperforming SPFCM on six datasets with partitions within 1% of FCM.
• MSERFCM achieved 5-26x speedup over FCM, producing partitions within 3% of FCM across all datasets.
• The new stopping criterion effectively accelerated SPFCM, yielding partitions comparable to FCM.

Conclusions:

• GOFCM and MSERFCM provide significant speedups for Fuzzy c-Means clustering while preserving partition accuracy.
• Statistical subsample size estimation is crucial for optimizing accelerated clustering performance.
• The novel stopping criterion enhances the efficiency of accelerated clustering methods.