Naslov
A hybrid dissimilarity measure for mixed-type data clustering

Autori
Perišić, Ana

Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, sažetak, znanstveni

Izvornik
Book of Abstracts BIOSTAT 2021 - 25th Int. Scientific Symposium on Biometrics / Jazbec, Anamarija ; Pecina, Marija ; Sonicki, Zdrenko ; Šimić, Diana ; Vedriš, Mislav ; Sović, Slavica - Zagreb : Hrvatsko biometrijsko društvo, 2021, 18-18

Skup
BIOSTAT 2021 - 25th International Scientific Symposium on Biometrics

Mjesto i datum
Poreč, Hrvatska, 08.09.2021. - 10.09.2021

Vrsta sudjelovanja
Predavanje

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
clustering, mixed data, hybrid dissimilarity measure, robust

Sažetak
One of the greatest challenges in clustering mixed-type data is finding the adequate distance function between objects. Most distance metrics work either with continuous only or categorical-only data, but in applications, however, mixed-type data are prevalent in many real-world applications. Hybrid distance methods involve selecting a distance function that can accommodate mixed-type variables where a popular hybrid distance function is Gower’s distance. This work presents a hybrid dissimilarity measure for mixed-type data where distances are calculated conditional on the feature type. The proposed dissimilarity measure is established as a normalized linear combination of distances following the principles of Gower’s coefficient calculation. For numerical features, distances are calculated by applying a modified winsorized Huber loss, while for categorical features, a distance measure based on variable entropy is incorporated. The established measure is robust to outliers, skewed and sparse data, and can handle unbalanced categorical features and highly skewed numerical features.

Izvorni jezik
Engleski

Znanstvena područja
Matematika, Interdisciplinarne prirodne znanosti

POVEZANOST RADA