B0566
Title: Robust fuzzy clustering based on quantile autocovariances
Authors: Borja Lafuente-Rego - Universidade da Coruna (Spain) [presenting]
Jose Vilar - Universidade da Coruna (Spain)
Pierpaolo Durso - Universita di Roma La Sapienza (Italy)
Abstract: Three robust versions of the fuzzy C-medoids clustering algorithm for the classification of time series based on comparing estimated sequences of quantile autocovariances (QA) are introduced. Namely, (i) QA-based exponential fuzzy C-medoids clustering, (ii) QA-based fuzzy C-medoids clustering with noise cluster and (iii) QA-based trimmed fuzzy C-medoids clustering. The first one uses a robust metric to neutralize and smooth the effect of outliers, the second one is aimed at detecting outliers and classify them into a noise cluster, and with the third method the model achieves its robustness by trimming away a certain fraction of anomalous time series. The robust fuzzy methods are evaluated in different simulated scenarios, considering different structures of dependence and introducing one or more outliers. All the proposed methods take advantage of the good properties of the QA-based metric, and the results reported from the numerical study outperform the ones obtained with classical fuzzy procedures based on alternative metrics. The usefulness of the proposal is illustrated by a real data application.
Journal