Αξιολόγηση αλγορίθμων συσταδοποίησης ακολουθιακών χωροχρονικών δεδομένων με χρήση διαφορετικών συναρτήσεων απόστασης
Evaluation of clustering algorithms of sequential spatio-temporal data with various distance functions
View/ Open
Keywords
Συναρτήσεις ομοιότητας ; Μετρικές συναρτήσεις απόστασης ; Συσταδοποίηση ; Αλγόριθμοι ; ΣυστάδεςAbstract
This thesis has as main objective to study and evaluate the performance of clustering algorithms considering sequential spatiotemporal data, each time using a different distance/similarity measure in order to highlight its advantages and disadvantages. Specifically, we apply metric distance functions (Euclidean, Manhattan, Chebyshev EuclideanSTARTEND), as well as non-metric distance functions, based either on dynamic time warping (DTW), or on editing distance on real sequence (EDR) or on longest common subsequence (LCSS). Various trajectories transformations (re-sampling, adding noise and point shift) controlled by two parameters, the rate and distance, are applied to real and synthetic trajectory datasets. For each transformation, the clustering of the original data set and the transformed data sets is evaluated depending on the value of the parameter which is not fixed. The results derived from the extensive experimental study are used to assess the validity of clusters obtained by the Optics clustering algorithm and hierarchical clustering via the Ward method, respectively.