UCLA

This dissertation is structured as a three-paper dissertation. The first chapter builds on social network analysis and structural balance theory to analyze, with a novel approach, some of the unintended consequences of Mexico’s kingpin strategy on the network of criminal organizations. The goal of this chapter is threefold: first, to show that the kingpin strategy is associated with the fragmentation of criminal organizations in Mexico; second, criminal organizations developed a set of structurally balanced arrangements before the government waged a war against them and that the kingpin strategy disrupted such arrangements; third, the fragmentation of criminal organizations also produced a process of clustering of violence. The second chapter examines the hazards faced by undocumented Mexican migrants during their U.S.-Mexico border crossings, particularly in the context of violent confrontations between drug trafficking organizations. By employing a negative binomial mixed effects model, difference-in-differences analysis, and unsupervised machine learning techniques, the research aims to discern which demographics within the undocumented migrant population are more susceptible to heightened hazards throughout their journey. The third chapter uses a combination of social network analysis (SNA) and machine learning techniques to predict links in Mexico’s network of criminal organizations. Specifically, it uses four similarity- based algorithms to estimate the likelihood that a link will be formed between two unconnected organizations in the network. Four algorithms are enhanced with the Node2Vec algorithm and a Deep Neural Network Architecture (DNNs) to improve their predictive capabilities. Of the node-similarity indices implemented, the Preferential Attachment algorithm enhanced with both the Node2Vec and the DNN architectures performed the best. When predicting potential future ties, the best performing algorithm was the Jaccard Coefficient enhanced both by the Node2Vec. I argue that this happens because the Jaccard Coefficient captures the immediate neighbor overlap but does not account for longer paths or the global network structure, while the Node2Vec does capture broader topological and community-based features that might not be apparent from direct connections alone.