Citas bibligráficas
Ancajima, E., (2022). Selección de atributos por métodos de filtrado y wrapper para predecir la satisfacción de usuarios de salud [Universidad Nacional Agraria La Molina]. https://hdl.handle.net/20.500.12996/5703
Ancajima, E., Selección de atributos por métodos de filtrado y wrapper para predecir la satisfacción de usuarios de salud []. PE: Universidad Nacional Agraria La Molina; 2022. https://hdl.handle.net/20.500.12996/5703
@misc{renati/1116272,
title = "Selección de atributos por métodos de filtrado y wrapper para predecir la satisfacción de usuarios de salud",
author = "Ancajima Bohórquez, Edgar Fernando",
publisher = "Universidad Nacional Agraria La Molina",
year = "2022"
}
Data mining techniques (DMT) used for supervised learning generally must consider the large number of attributes in the databases to be analyzed, and many of these attributes are irrelevant and redundant that can distort performance and functionality of these techniques, and therefore their predictive capacity. Research on the subject of feature selection mentions that by selecting a smaller number of features from the total set, it can bring a series of advantages: reduce redundancy, eliminate noise, maximize the relevance of features, reduce computational cost, increase the interpretation and improve the accuracy of the supervised learning classifier. The objective is to present the filtering and Wrapper attribute selection methods that can be applied in supervised data mining techniques for the classification task, obtaining the best subsets of relevant attributes with the highest accuracy rates. Four metrics are applied to select the attributes by filtering (Chi-Square, Information gain, Gain ratio and Relief) and four methods by Wrapper (Best-First, Greedy forward, Greedy backward and Hill climbing) in the National Satisfaction Survey of Health Users–2015. The results applying four TMD to each of the different subsets of attributes selected with the methods of filtering and wrapper, showed the greatest predictive capacities to predict the satisfaction of the users of the attention received from the health services, in the case of binary logistic regression, the Best-First wrapper method with 5 attributes and an accuracy of 88,7%, the C5.0 classification tree with Greedy forward wrapper with 6 attributes and an accuracy of 89,1%, the Naive Bayesian network with Greedy backward wrapper with 16 attributes and an accuracy of 88,3% and the random Forest multiclassifier with Greedy backard wrapper with 16 attributes and an accuracy of 93,0%. The highest AUC for binary logistic regression was with the Greedy forward method with 0,932, the C5.0 classification tree with Greedy forward with 0,891, the Naive Bayesian network with Greedy forward wrapper with 0,9221 and the random Forest multiclassifier with Greedy backard. with 0,941.
Este ítem está sujeto a una licencia Creative Commons Licencia Creative Commons