Modelos de machine learning para clasificar la cartera en un fondo de pensiones
Fecha
2022-09-22Registro en:
Gil Rubio, R. (2022). Modelos de machine learning para clasificar la cartera en un fondo de pensiones. [Maestría, Universidad Santpo Tomás]. Repositorio institucional.
reponame:Repositorio Institucional Universidad Santo Tomás
instname:Universidad Santo Tomás
Autor
Gil Rubio, Ricardo
Institución
Resumen
The present paper has as objective, the application of different Machine Learning techniques as well as statistical and inferential diagnostics, to propose predictive analysis models that allow to in due time identify, classify and process the companies that are not paying pension contributions to their employees affiliated to the pension fund, and thus to implement different collection strategies to recover contributions owed.
In the process of evaluating the performance of the models, it was possible to show that the Decision Trees technique presents excellent results: it did not require standardization of the data by achieving an excellent percentage of certainty and it quickly and efficiently classified the predictor variable in a database with an adequate number of records. The other techniques showed good results in class type 0, 3 and 4 with percentages above 96.8\% both in completeness and in measure-F, while the performance decreased for Logistic Regression 71.8\% and Support Vector Machines 69.2\% in completeness and Bayesian Networks 18.5\% in measure-F, the above for class type 1. In the Bayesian Networks technique for class type 2 it was reduced by 24.7\% and 29.3\% both in completeness and F-measure and Support Vector Machines at 59.4\% for F-measure. This was addressed with the treatment of unbalanced classes and with the reinforcement or ensemble algorithms.
Class imbalance is a fairly common problem when working with real data; when samples from one or multiple classes are over represented in a data set. There are several areas in which it can occur, such as spam filtering, cancer detection, fraud identification or disease detection. Strategies to deal with class imbalance include minority class up sampling, majority class down sampling, and generation of synthetic training samples using the most commonly used algorithm (SMOTE).
Once the models with the proposed segmentation were evaluated, the strategies were generated that allowed identifying the collection management mechanisms depending on the type of debtor, this ranges from a commercial visit, contact center management for preventive collection or an extract with payment information, for debtors of low criticality, going through a persuasive collection letter, advice at service points or text messages for debtors of medium criticality, to the coercive collection process, embargoes and other measures for debtors who are reluctant to pay.