Estudio comparativo de modelos de agrupamiento para el reconocimiento de patrones de corrupción en contrataciones públicas

Abstract

Corruption is a phenomenon that has been present in societies since ancient times, which is why it has been widely studied from different perspectives to help understand it. Corruption has become an increasingly recurring theme in government administration and to a large extent in public procurement processes. These processes are essential for government management, since through public procurement it seeks to provide any need that the state identifies in order to achieve its objectives. This work presents a comparative study of two models for the recognition of corruption patterns in public procurement data based on clustering algorithms such as Gustafson-Kessel and DBSCAN (Density-based spatial clustering of applications with noise), taking as a reference a framework corruption analysis carried out following an adaptation of the PRISMA methodology (Preferred Reporting Items for Systematic reviews and Meta-Analyzes). For the development of this work, a multidisciplinary literature review is first carried out that allows proposing variables related to the causes of corruption. Second, data on public contracting related to acts of corruption and contracting data from the Electronic Public Procurement System (SECOP) of Colombia are obtained, which are refined and conditioned to form the database that will be used in the study. Third, the experimental methodology is proposed to develop the models. Subsequently, the methodology is executed with the chosen algorithms, its results are analyzed and discussed, and the performance of each one is compared. The results show that the behavior of some variables is in line with the causes found in the analysis framework, for example the variables Surface, GDP and Market; However, other variables such as Population, Education and Contract Amount are not in line with the causes raised in the analysis framework. Taking into account the available data and the proposed methodology, better results were obtained with the model based on the Gustafson-Kessel algorithm, in relation to the expected behavior of the data according to the analysis framework. Regarding the performance of the models, it is observed that the one based on the DBSCAN algorithm uses less than 30% of the time used by the Gustafson-Kessel algorithm.