Human Rights Violations and Machine Learning - Cluster Analysis of Countries using the CIRIGHTS Dataset

Sammanfattning: This master's thesis explores the use of unsupervised machine learning techniques to cluster countries based on their degree of human rights violations. Accordingly, the study evaluates the performance of two clustering methods, K-Means clustering and Latent Class Analysis (LCA), using two cluster validation metrics (Silhouette Coefficient and Dunn Index), as well as an Accuracy measure using the Human Rights index. It analyses the characteristics of clusters and the assignments thereof over four decades to provide compact insights for policymakers. The results, in turn, show that both clustering methods perform equally well, however, LCA is chosen for the bulk of the analysis out of respect for the categorical nature of the data. Consequently, cluster profiling identifies three clusters with varying levels of human rights scores, although, looking at each variable and decade individually, we see that they do not all follow the same order of magnitude that the overall cluster scores suggest. Furthermore, the probability transition matrix shows that, generally, countries do not change significantly over time, in terms of their level of respect for human rights. Finally, policy advice for stable countries involves using cluster 1 as a “gold standard”, incentivizing cluster 2, and taking a proactive approach for cluster 3. In turn, for unstable countries, advice includes incentivizing further improvements for countries that have shown positive progress, understanding reasons for decline, and stabilising and monitoring closely those that have shown fluctuating tendencies. The paper concludes that unsupervised machine learning for detecting human rights violations is useful, efficient, and provides insights into patterns that are not immediately apparent. Furthermore, it is a useful instrument to summarise these patterns in a clear and interpretable way.