A1186
Title: Robust fuzzy clustering for high-dimensional multivariate time series with outlier detection
Authors: Ziling Ma - King Abdullah University of Science and Technology (KAUST) (Saudi Arabia) [presenting]
Angel Lopez Oriona - King Abdullah University of Science and Technology (KAUST) (Saudi Arabia)
Ying Sun - KAUST (Saudi Arabia)
Hernando Ombao - King Abdullah University of Science and Technology (KAUST) (Saudi Arabia)
Abstract: Fuzzy clustering offers a natural approach to modeling partial memberships. However, most existing algorithms were developed for static, low-dimensional data. Such methods struggle with multivariate time series (MTS), where challenges include temporal dependence, unequal sequence lengths, moderate-to-high dimensionality, and frequent contamination by noise or artifacts. To overcome these challenges, RFCPCA is introduced, a robust fuzzy clustering framework that is, to the best of knowledge, the only approach specifically tailored to MTS that integrates (i) membership-informed subspace learning through common principal component analysis, (ii) the ability to handle unequal lengths and moderately high dimensions, and (iii) robustness against outliers via trimming, exponential down-weighting, and a noise cluster. This unique combination allows RFCPCA to capture latent temporal structure while remaining stable under contamination and informative about ambiguous or atypical series. Simulated and real Electroencephalogram data demonstrate that RFCPCA not only improves clustering accuracy compared to related methods, but also provides a more reliable characterization of uncertainty and atypical structure in MTS.
