Sistema de detección de patrones de fraude con redes neuronales en la Provincia De Los Rios y su insidencia en la telefonia celular, año 2015

In this research work is due to the increase in fraudulent actions and losses of telephone operators worldwide, mainly in mobile telephony. For this reason, a method is proposed for the detection of fraud and anomalies in cellular telephone networks through the use of neural networks, using telephone records as the main source of information. User profiles are created, with which a differential analysis is made to detect changes in consumer consumption patterns and thus generate alarms against possible fraud; With satisfactory results in terms of the efficiency and effectiveness of the applied method. It will address the problem of detecting changes in consumption of cell phone users than normal, the corresponding construction of data structures that represent the recent and historical behavior of each of the users, taking into account the information that contains a call And the complex construction of a function with so many variables input parameterization not always known. While there are different ways of detecting fraud, they all work with peaks of consumption or fixed rules, which do not always indicate abnormal behavior. The solution being used uses unsupervised neural network technology, in particular SOM networks. Once the patterns that define the space of all the calls were obtained, the construction of the user profiles was performed through the development of a frequency distribution of each of the patterns for each profile (CUP and UPH) and the Corresponding alarm detection. The process was based on presenting to the system the calls made in a period of 3 months by the users reported as "high consumption". With each call the user's CUP profile was updated, compared to the UPH profile, obtaining the Hellinger distance (H) between them, and if it exceeded the set threshold, an alarm was triggered. Depending on the refresh frequency parameter of the UPH profile (f), the UPH was updated with the CUP contribution as appropriate. It is worth clarifying that the process of construction and updating was done from the first call of the user; On the other hand the comparison and corresponding detection of the alarm was performed only after the number of calls analyzed for the user passed the minimum amount to build a profile (QL) with sufficient information from the user. At the time of entering the first call of a user, all patterns of CUP and UPH were initialized with the same frequency distribution, assuming that the user had the same tendency to perform any type of a priori call without information. In turn this experience was performed twice: the first one updating the UPH with each call and therefore with a low threshold Hellinger (H) for the detection of alarms because the difference that could be presented between the CUP and UPH profiles was Very small updating the historical profile with each call, it would have to be equal to the current profile. The second experiment was performed by updating the UPH once a day and a high Hellinger (H) threshold to detect important differences that could be considered as behavioral changes.

• Egile nagusia: Pacheco Paliz, Luis Roberto (author)
• Formatua: masterThesis
• Hizkuntza: spa
• Argitaratua: 2015
• Gaiak: Patrones; Redes neuronales
• Sarrera elektronikoa: http://repositorio.uteq.edu.ec/handle/43000/1763