Voting trends within the United Nations

Abstract

The United Nations (UN) is a fundamental worldwide organisation with the sole objective of uniting countries’ common principles in one globalised forum. An important aspect of this organisation are resolutions that are proposed by its members and other members have to vote for these resolutions. The purpose of this research is to mine a dataset that contains previous voting records of countries in UN assemblies and then attempt to predict how each country will vote in UN resolutions and further understand whether a resolution will pass or not as a resolution. The dataset that was used contains data about UN resolutions from 1946 to 2021, including their description and how each country voted for each resolution. The features primarily used from this dataset were the set of texts describing the content of each proposed resolution. This text was then converted into vectors so that it was understood by the models that were able to issue predictions based on this input feature represented as numerical vectors. The models used are the following; Support Vector Machine (SVM), Decision Tree, Random Forest, K-Nearest-Neighbors (KNN), Logistic Regression and Naive Bayes Classifier. Three main objectives were set for this study. The first objective concerns predicting how each country will vote for a resolution using the mentioned text feature and using the mentioned Machine Learning (ML) algorithms above. Five countries were selected and tested on, then, the metric results of these models were compared to identify the best performing models and to what extent. The average accuracy achieved out of all the models implemented was 81%. The second objective revolves around understanding whether a resolution would pass or not as a resolution, by using the descriptions of the resolutions as inputs and as the targets, either a resolution pass (1) or resolution failed to pass (0). The metric scores exceeded 90% from the models implemented. In the last objective, an investigation was conducted on how the probability of a country voting in favour or against a resolution is dependent on time. Hence, a number of different time periods were selected and investigated for the training set of Objective 1, to understand if the models remained valid across different periods. In this objective we also concluded that training models with older data from the dataset did not create a concept drift. The best performing model, which was the SVM model, achieved 85% accuracy.