Performance Comparison of Dimensional Reduction using Principal Component Analysis with Alternating Least Squares in Modified Fuzzy Possibilistic C-Means and Fuzzy Possibilistic C-Means

Abstract

The clustering method is said to be good if it has resistance to outlier data. One cluster method resistant to outlier data is Fuzzy Possibilistic C-Means (FPCM). FPCM performance on outlier data still has the potential for overlap between cluster members in different clusters, resulting in decreased cluster quality. The Modified Fuzzy Possibilistic C-Means (MFPCM) method is used to modify FPCM in its objective function by inserting updated weight values to increase FPCM performance. In this research, improving the quality of FPCM and MFPCM clusters was carried out by reducing data dimensions through Principal Component Analysis using Alternating Least Squares (PRINCALS) so that members of each cluster do not overlap in the right cluster. The PRINCALS results of the FPCM method have better performance with silhouette values and BSS/TSS ratios of 0.4108 and 60% compared to values without PRINCALS of 0.355 and 43%. The MFPCM method with PRINCALS also performs better, namely 0.4299 and 61%, compared to 0.368 and 42% without PRINCALS. In this study, the performance of MFPCM with PRINCALS or without PRINCALS was better than that of the FPCM method. Overall, PRINCALS can improve the performance of the MFPCM and FPCM methods, resulting in better clusters. PRINCALS in this cluster produce an average silhouette value greater than 0.3 and an average BSS/TSS ratio greater than 50% so that each cluster member is in the right cluster and does not overlap.