Optimización de la Precisión en la Detección de Noticias Falsas de política en español mediante la Aplicación de Algoritmos de Optimización en la Regresión Logística

Logistic regression, while widely employed in text classification for fake news detection, shows suboptimal optimization practices in this specific domain. The limited systematic exploration of optimization algorithms—Gradient Descent (GD), Stochastic Gradient Descent (SGD), Mini-Batch Gradient Descent (MBGD), AdaGrad, Adam, and RMSProp—hinders the accurate assessment of their impact on classification metrics. This Curricular Integration Project (CIP) addressed this gap by applying these six algorithms to a logistic regression model for detecting Spanish-language political fake news, following the CRISP-ML methodology. The workflow included: 1) Data engineering to create a custom dataset, 2) Model optimization through hyperparameter tuning of the algorithms, and 3) Evaluation using confusion matrices and performance metrics (Sensitivity, Specificity, Precision, Accuracy, and F1-Score). Results revealed that the SGD-LR variant (Stochastic Gradient Descent) outperformed both the baseline non-optimized logistic regression model (73.7% vs. 80.3% precision) and other evaluated optimizers. This 6.6% improvement highlights how strategic algorithm selection directly enhances classification performance. The study not only validates SGD’s efficacy for this task but also sets a methodological precedent by integrating CRISP-ML into optimization workflows. These findings underscore the necessity of systematic experimentation with optimizers as a critical phase in developing misinformation detection systems, particularly for Spanish-language content where technical studies remain scarce.

• 主要作者: Tene Castillo, Santiago Emanuel (author)
• 格式: bachelorThesis
• 語言: spa
• 出版: 2025
• 主題: MACHINE LEARNING; CLASIFICACION DE TEXTO; AJUSTAR HIPERPARAMETROS
• 在線閱讀: https://dspace.unl.edu.ec/jspui/handle/123456789/32452