Citas bibligráficas
Meza, A., (2018). Predicción de fuga de clientes en una empresa de telefonía utilizando el algoritmo Adaboost desbalanceado y la regresión logística asimétrica [Tesis, Universidad Nacional Agraria La Molina]. https://hdl.handle.net/20.500.12996/3245
Meza, A., Predicción de fuga de clientes en una empresa de telefonía utilizando el algoritmo Adaboost desbalanceado y la regresión logística asimétrica [Tesis]. : Universidad Nacional Agraria La Molina; 2018. https://hdl.handle.net/20.500.12996/3245
@mastersthesis{renati/1114629,
title = "Predicción de fuga de clientes en una empresa de telefonía utilizando el algoritmo Adaboost desbalanceado y la regresión logística asimétrica",
author = "Meza Rodríguez, Aldo Richard",
publisher = "Universidad Nacional Agraria La Molina",
year = "2018"
}
The purpose of this research is to apply and to compare the logistic regression model and the Adaboost algorithm in unbalanced data, the purposes of predict the customer churn in a company in the mobile telephony sector. The Adaboost algorithm is based on adaptive learning when training weak classifiers, combining them together to obtain a classifier whose performance is strong. In terms of logistic regression, its modeling was done strictly from a data mining perspective, where the classification is the objective and the performance was evaluated in a validation set. Both techniques were compared using two methods, the first using sampling methods (sub-sampling, oversampling and SMOTE) and the second modifying and / or adjusting the algorithm or function. When working with unbalanced data the classification error rate is inefficient, so the performance measures to choose the best model were accuracy, recall (sensitivity), F-measure, and as a main measure the AUC through ROC curves. When forming logistic models with the sampling methods, the performance measures yielded similar results, the same happened when forming models with the Adaboost algorithm, however when comparing the logistic regression (AUC = 0.86) with the Adaboost algorithm (AUC = 0.93), the latter had the best performance. Regarding the adjustment at the level of algorithm or function, the logistic regression was worked in two ways, the first (Logit Asym) including in the FDA a Kappa value (k) and the second (Power Logit) a Lambda value (λ), in both models the optimal values of k (0.02) and λ (2.5) were identified, in terms of the Adaboost algorithm (Adaboost Asym) the weight of the minority class whose cost of classification was erroneous was adjusted. The comparison of these three adjusted models gave the Adaboost algorithm a higher performance. Finally, cross validation was carried out with 10 iterations for all the models, giving similar results to the retention method. Once all the comparisons and measures of performance are concluded, it is concluded that the optimal model for the prediction of customer leakage in the telephone company is the Adaboost algorithm
Este ítem está sujeto a una licencia Creative Commons Licencia Creative Commons