Mixture-based Clustering Non-gaussian Data with Fixed Bounds

0462336 - ÚTIA 2017 RIV BG eng C - Konferenční příspěvek (zahraniční konf.)
Nagy, Ivan - Suzdaleva, Evgenia - Mlynářová, Tereza
Mixture-based Clustering Non-gaussian Data with Fixed Bounds.
Proceedings of 2016 IEEE 8th International Conference on Intelligent Systems. Sofia: IEEE, 2016, 265-271, s. 265-271. ISBN 978-1-5090-1353-1.
[2016 IEEE 8th International Conference on Intelligent Systems IS'2016. Sofia (BG), 04.09.2016-06.09.2016]
Grant CEP: GA ČR GA15-03564S
Institucionální podpora: RVO:67985556
Klíčová slova: mixture-based clustering * recursive mixture estimation * mixture of uniform distributions * data-dependent pointer
Kód oboru RIV: BB - Aplikovaná statistika, operační výzkum
http://library.utia.cas.cz/separaty/2016/ZS/suzdaleva-0462336.pdf

This paper deals with clustering non-gaussian data with fixed bounds. It considers the problem using recursive mixture estimation algorithms under the Bayesian methodology. Such a solution is often desired in areas, where the assumption of normality of modeled data is rather questionable and brings a series of limitations (e.g., non-negative, bounded data, etc.). Here for modeling the data a mixture of uniform distributions is taken, where individual clusters are described by mixture components. For the on-line detection of clusters of measured bounded data, the paper proposes a mixture estimation algorithm based on (i) the update of reproducible statistics of uniform components; (ii) the heuristic initialization via the method of moments; (iii) the non-trivial adaptive forgetting technique; (iv) the data-dependent dynamic pointer model. The approach is validated using realistic traffic flow simulations.
Trvalý link: http://hdl.handle.net/11104/0262262