How to Simulate Outliers with the Desired Properties

[EN] Deviating multivariate observations are used typically to test the performance of outlier detection methods. Yet, the generation of outlying cases itself usually appears as a secondary methodological step in methods comparison. In the literature, outliers are defined using certain distribution...

Descripción completa

Detalles Bibliográficos
Autores: González-Cebrián, A., Arteaga, Francisco, Folch-Fortuny, Abel, Ferrer, Alberto|||0000-0001-7244-5947
Tipo de recurso: artículo
Fecha de publicación:2021
País:España
Institución:Universitat Politècnica de València (UPV)
Repositorio:RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Idioma:inglés
OAI Identifier:oai:riunet.upv.es:10251/182472
Acceso en línea:https://riunet.upv.es/handle/10251/182472
Access Level:acceso abierto
Palabra clave:Outliers
Squared prediction error
Hotelling&apos
s T^2
Simulation
PCA
Matlab
ESTADISTICA E INVESTIGACION OPERATIVA
Descripción
Sumario:[EN] Deviating multivariate observations are used typically to test the performance of outlier detection methods. Yet, the generation of outlying cases itself usually appears as a secondary methodological step in methods comparison. In the literature, outliers are defined using certain distribution parameters which differ from those of the clean or reference data. However, these parameters change among authors, leading to a lack of a standard and measurable definition of the characteristics simulated outliers. This makes the comparison between methods hard and its results dependent on the procedure followed to simulate the data. In order to set a standard procedure, a framework to simulate outliers is defined here. Since it is based on certain specifications for both the Squared Prediction Error (SPE) and Hotelling's T2 statistics from a Principal Component Analysis (PCA) model, tuning them becomes a simple and efficient task. This procedure has been implemented in a set of Matlab functions.