Joint outlier detection and variable selection using discrete optimization

In regression, the quality of estimators is known to be very sensitive to the presence of spurious variables and outliers. Unfortunately, this is a frequent situation when dealing with real data. To handle outlier proneness and achieve variable selection, we propose a robust method performing the ou...

Descripción completa

Detalles Bibliográficos
Autores: Jammal, Mahdi, Canu, Stephane, Abdallah, Maher
Tipo de recurso: artículo
Fecha de publicación:2021
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/362112
Acceso en línea:https://hdl.handle.net/2117/362112
https://dx.doi.org/10.2436/20.8080.02.109
Access Level:acceso abierto
Palabra clave:Robust optimization
statistical learning
linear regression
variable selection
outlier detection
mixed integer programming
Programació (Matemàtica)
Intel·ligència artificial
Estadística matemàtica
Classificació AMS::62 Statistics::62J Linear inference, regression
Classificació AMS::62 Statistics::62G Nonparametric inference
Classificació AMS::68 Computer science::68T Artificial intelligence
Classificació AMS::90 Operations research, mathematical programming::90C Mathematical programming
Àrees temàtiques de la UPC::Matemàtiques i estadística::Estadística matemàtica
Descripción
Sumario:In regression, the quality of estimators is known to be very sensitive to the presence of spurious variables and outliers. Unfortunately, this is a frequent situation when dealing with real data. To handle outlier proneness and achieve variable selection, we propose a robust method performing the outright rejection of discordant observations together with the selection of relevant variables. A natural way to define the corresponding optimization problem is to use the ℓ0 norm and recast it as a mixed integer optimization problem. To retrieve this global solution more efficiently, we suggest the use of additional constraints as well as a clever initialization. To this end, an efficient and scalable non-convex proximal alternate algorithm is introduced. An empirical comparison between the ℓ0 norm approach and its ℓ1 relaxation is presented as well. Results on both synthetic and real data sets provided that the mixed integer programming approach and its discrete first order warm start provide high quality solutions.