RISTI - Revista Ibérica de Sistemas e Tecnologias de Informação

Avanços Recentes em Sistemas e Tecnologias de Informação

Machine Learning Models for Classifying Corruption Risk in Public Procurement: The Case of Public Hospitals in Colombia

Resumen La contratación pública es particularmente vulnerable a la corrupción, lo que plantea desafíos significativos para la gestión pública. A nivel mundial, los gobiernos han implementado iniciativas de datos abiertos con el objetivo de promover la transparencia y fortalecer la integridad en la gestión de los recursos públicos. Sin embargo, el éxito de estas iniciativas depende en gran medida de la capacidad para analizar los datos disponibles y detectar patrones que puedan señalar riesgos de corrupción. Este estudio presenta el desarrollo y la evaluación de modelos de aprendizaje automático orientados a analizar y predecir el riesgo de corrupción en procesos de contratación pública. Para ello, se utilizaron datos abiertos de contratación en hospitales públicos colombianos entre 2014 y 2019. Como variable de riesgo, se empleó el porcentaje de contratación directa, ampliamente reconocido como un indicador de posibles irregularidades. Se implementaron y compararon tres algoritmos de aprendizaje automático: Árboles de Decisión, Random Forest y Gradient Boosting. Los resultados evidenciaron un desempeño aceptable, con niveles de exactitud entre el 46% y el 59% y un área bajo la curva ROC que oscila entre el 0.56 y el 0.72.

Abstract Public procurement is particularly vulnerable to corruption, posing significant challenges to public administration. Globally, governments have implemented open data initiatives to promote transparency and strengthen resource management's integrity. However, the success of these initiatives largely depends on the ability to analyze available data and identify patterns that may indicate corruption risks. This study presents the development and evaluation of machine learning models designed to analyze and predict the risk of corruption in public procurement processes. For this purpose, open procurement data from Colombian public hospitals between 2014 and 2019 were used. The percentage of direct contracting, widely recognized as an indicator of potential irregularities, was employed as the risk variable. Three machine learning algorithms were implemented and compared: Decision Trees, Random Forest, and Gradient Boosting. The results demonstrated acceptable performance, with accuracy levels ranging from 46% to 59% and an area under the ROC curve between 0.56 and 0.72.

Análise de metodologias para classificação de deepfakes

Resumo Este trabalho investiga a crescente sofisticação dos deepfakes e a necessidade de aprimorar os métodos de detecção. O objetivo é avaliar o desempenho de diferentes modelos de Deep Learning na classificação de vídeos manipulados, buscando uma acurácia mínima de 80%. Foram analisadas as arquiteturas LSTM, CNN, GRU e CvT, além da aplicação da Lei de Benford como método estatístico auxiliar. A base Celeb-DF V2 foi utilizada por sua fidelidade e complexidade. O modelo LSTM apresentou a melhor acurácia no treinamento, enquanto o GRU obteve melhor desempenho na avaliação prática. Já o CvT mostrou limitações tanto em desempenho quanto em custo computacional. A Lei de Benford mostrou-se promissora, mas inconclusiva sem um parâmetro de correlação. Os resultados reforçam a importância de atualizar continuamente os mecanismos de detecção para acompanhar a evolução dos vídeos sintéticos e mitigar riscos associados à desinformação.

Abstract This study investigates the increasing sophistication of deepfakes and the need to improve detection techniques. The objective is to evaluate the performance of different Deep Learning models in classifying manipulated videos, aiming for a minimum accuracy of 80%. The architectures LSTM, CNN, GRU, and CvT were analyzed, along with the application of Benford’s Law as a statistical aid. The Celeb-DF V2 dataset was selected for its realism and complexity. The LSTM model achieved the highest training accuracy, while GRU showed superior performance during evaluation. The CvT model underperformed in both accuracy and computational cost. Although promising, Benford’s Law yielded inconclusive results without a correlation metric. The findings highlight the importance of continually updating detection methods to keep pace with synthetic video advancements and mitigate risks linked to misinformation.

Application of CNN in Audio Recognition of Isthmus Zapotec

Resumen Este estudio se enfoca en la clasificación de espectrogramas, representaciones visuales del audio para aplicar aprendizaje automático. Los métodos tradicionales, como los MFCCs con clasificadores clásicos, presentan limitaciones en lenguas con pocos recursos, como el zapoteco del Istmo. Modelos avanzados como RNNs y transformers requieren grandes volúmenes de datos, difíciles de obtener en contextos indígenas. Como alternativa, se propone una red neuronal convolucional profunda de 28 capas, entrenada con 10 frases comunes convertidas en espectrogramas y aumentadas artificialmente. El modelo logró un 100% de precisión en entrenamiento y 96.2% en validación. Aunque prometedor, se destaca la necesidad de ampliar el conjunto de datos. El trabajo evidencia el potencial del aprendizaje profundo para mejorar la comunicación intercultural y preservar lenguas indígenas en peligro.

Abstract This study focuses on the classification of spectrograms, visual representations of audio, for the application of machine learning. Traditional methods, such as MFCCs with classical classifiers, have limitations in resource-poor languages such as Isthmus Zapotec. Advanced models, such as RNNs and transformers, require large volumes of data, which are often difficult to obtain in indigenous contexts. As an alternative, a 28-layer deep convolutional neural network is proposed, trained with 10 common phrases converted into spectrograms and artificially augmented. The model achieved 100% training accuracy and 96.2% validation accuracy. Although promising, the need to expand the dataset is highlighted. This work demonstrates the potential of deep learning to improve intercultural communication and preserve endangered indigenous languages.

Systematic review of virtual games in numerical and scientific knowledge for grades 1 to 5

Resumen El presente trabajo indaga como los juegos en formato virtual contribuyen al desarrollo cognitivo en las áreas de conocimiento matemático y científico dirigidos a grados de 1° a 5° de primaria. Se realizó una revisión de literatura entre el periodo (2017 - 2022) en tres revistas referentes en publicaciones de tecnología y educación. Como resultado, se identificaron que la gran mayoría de estudios usaron diseños de investigación cuasi experimental y frente a las situaciones mediadas por TIC hubo una tendencia hacia los video juegos generalistas. El análisis de los estudios (diseños de investigación - videojuegos) evidenció que el uso de las TIC se relaciona a los aspectos educativos, dejando de lado un poco, la discusión acerca del desarrollo cognitivo.

Abstract The present work investigates how games in virtual format contribute to cognitive development in the areas of mathematical and scientific knowledge aimed at grades 1 to 5 of primary school. A literature review was carried out during the period (2017 - 2022) in three leading technology and education publications journals. As a result, it was identified that the vast majority of studies used quasi-experimental research designs, and in the face of ICT-mediated situations, there was a tendency towards generalist video games. The analysis of the studies (research designs - video games) showed that the use of ICT is related to educational aspects, leaving aside the discussion about cognitive development.

Early Music Education and Technology: A Systematic Review

Resumen La educación musical en la primera infancia y la integración de recursos tecnológicos han despertado un interés creciente por sus potenciales beneficios en el desarrollo cognitivo, la motivación y la interacción social. Con el objetivo de analizar los efectos de la tecnología aplicada a la práctica musical en niños de 2 a 6 años, se realiza una revisión sistemática en las bases de datos Web of Science y Scopus. Se identificaron 151 registros, se seleccionaron ocho artículos que cumplían con los criterios de inclusión, con una muestra total de 697 participantes. Los resultados sugieren que la incorporación de herramientas tecnológicas en la práctica musical colectiva favorece la creatividad, la motivación y el aprendizaje autónomo. La escasez de investigaciones empíricas limita la generalización de los hallazgos, lo que evidencia la necesidad de ampliar la producción científica y de avanzar hacia un modelo educativo híbrido que integre equilibradamente tecnología y métodos tradicionales.

Abstract Early childhood music education and the integration of technological resources have sparked growing interest due to their potential benefits for cognitive development, motivation, and social interaction. To analyze the effects of technology applied to musical practice in children aged 2 to 6, a systematic review was conducted in the Web of Science and Scopus databases. One hundred and fifty-one records were identified, and eight articles that met the inclusion criteria were selected, with a total sample of 697 participants. The results suggest that the incorporation of technological tools into collective musical practice fosters creativity, motivation, and autonomous learning. The scarcity of empirical research limits the generalization of the findings, highlighting the need to expand scientific production and move toward a hybrid educational model that smoothly integrates technology and traditional methods.

Caracterização espectral da cobertura do solo do Parque Estadual do Cocó a partir de imagens CBERS-4ª

Resumo Os parques urbanos são essenciais para a biodiversidade e o bem-estar humano. O Parque Estadual do Cocó (PEC), em Fortaleza, Brasil, sofre alta pressão antropogênica, exigindo monitoramento contínuo. Este estudo caracterizou espectralmente suas coberturas com imagens do sensor WPM do satélite CBERS-4A, usando uma cena de 24 de julho de 2024. Foram analisadas assinaturas espectrais das principais classes, com base na classificação do MapBiomas, e calculados os índices NDVI, SAVI e EVI para avaliar a vegetação. Os resultados mostraram que o sensor WPM discrimina eficazmente as coberturas, com assinaturas distintas para manguezais, restingas, água e áreas urbanas. O manguezal domina (50,78%), seguido pela restinga herbácea (27,24%) e áreas urbanizadas (10,78%), evidenciando pressão periférica. O EVI superou o NDVI em sensibilidade em áreas de alta biomassa, evitando a saturação. As imagens CBERS-4A provaram ser ferramentas eficazes e econômicas para monitorar ecossistemas urbanos complexos, fornecendo dados essenciais para a conservação e gestão sustentável do PEC.

Abstract Urban parks are essential for biodiversity and human well-being. The Cocó Stake Park (PEC) in Fortaleza, Brazil, faces significant anthropogenic pressure, necessitating continuous monitoring. This study spectrally characterized land cover using WPM sensor imagery from the CBERS-4A satellite, based on a scene from July 24, 2024. Spectral signatures of major land cover classes were analyzed, supported by MapBiomas classification, and vegetation indices (NDVI, SAVI, and EVI) were computed to assess vegetation condition. Results demonstrated that the WPM sensor effectively discriminates land cover types, with distinct spectral signatures for mangroves, restinga, water bodies, and urban areas. Mangroves dominate the park (50.78%), followed by herbaceous restinga (27.24%) and urbanized areas (10.78%), highlighting substantial peripheral pressure. EVI outperformed NDVI in sensitivity to structural variations in high-biomass zones, avoiding saturation. CBERS-4A imagery proves to be a cost-effective and reliable tool for monitoring complex urban ecosystems, providing critical data for informed decision-making in the sustainable conservation and management of the PEC.