MCFS: Min-cut-based feature-selection

In this paper, MCFS (Min-Cut-based feature-selection) is presented, which is a feature-selection algorithm based on the representation of the features in a dataset by means of a directed graph. The main contribution of our work is to show the usefulness of a general graph-processing technique in the...

Descripción completa

Detalles Bibliográficos
Autores: García Vallejo, Carlos Antonio, Troyano Jiménez, José Antonio, Enríquez de Salamanca Ros, Fernando, Ortega Rodríguez, Francisco Javier, Cruz Mata, Fermín
Tipo de recurso: artículo
Estado:Versión enviada para evaluación y publicación
Fecha de publicación:2020
País:España
Institución:Universidad de Sevilla (US)
Repositorio:idUS. Depósito de Investigación de la Universidad de Sevilla
OAI Identifier:oai:idus.us.es:11441/104525
Acceso en línea:https://hdl.handle.net/11441/104525
https://doi.org/10.1016/j.knosys.2020.105604
Access Level:acceso abierto
Palabra clave:Machine learning
Feature selection
Nearest neighbour
Correlations
Max-flow min-cut
Classification
Descripción
Sumario:In this paper, MCFS (Min-Cut-based feature-selection) is presented, which is a feature-selection algorithm based on the representation of the features in a dataset by means of a directed graph. The main contribution of our work is to show the usefulness of a general graph-processing technique in the feature-selection problem for classification datasets. The vertices of the graphs used herein are the features together with two special-purpose vertices (one of which denotes high correlation to the feature class of the dataset, and the other denotes a low correlation to the feature class). The edges are functions of the correlations among the features and also between the features and the classes. A classic max-flow min-cut algorithm is applied to this graph. The cut returned by this algorithm provides the selected features. We have compared the results of our proposal with well-known feature-selection techniques. Our algorithm obtains results statistically similar to those achieved by the other techniques in terms of number of features selected, while additionally significantly improving the accuracy.