A0546
Title: Robust linkage methods for functional data clustering
Authors: Tianbo Chen - Anhui University (China) [presenting]
Abstract: Clustering, an essential component of data mining and machine learning, serves as a statistical tool for classifying the data unsupervised. Among hierarchical clustering techniques, Ward's linkage method measures the incremental sum of squares errors (SSE, or graphically, the diameter of a cluster) when two clusters are merged. However, traditional linkage methods exhibit limitations in handling outliers and contaminations within datasets, compromising their partitioning capabilities. The aim is to introduce two robust Ward-like linkage methods for functional data clustering by only taking the most central curves into account. The cluster diameter is defined as the width of the band delimited by the most central curves selected by modified band depth and magnitude-shape outlyingness measure. Results from simulations and EEG data analysis demonstrate the superior performance of the proposed methods over conventional Ward's linkage and centroid linkage, particularly when different types of outliers and contaminations are presented.
