P1: On Voter Characterisation In Developing Democracies

Over the last decade, there has been an increase in the use of machine learning techniques for the prediction and analysis of elections. The recent trend has been the use of text mining of social media data to predict election outcomes. In this work, we assess the usefulness of employing census data to understand the voting patterns of South African voters during the 2016 local government elections at the electoral district level. We frame the problem as an unsupervised learning problem. Self-organising maps are utilised due to their visual nature and the resulting accessibility of information to stakeholders. The analysis shows that race, employment status, level of education, internet access and quality of service delivery in local government jurisdictions are key in determining which political party dominates that particular electoral district. These results could be used by the Electoral Commission of South Africa to prevent voter miss-information, especially for those voters who are in remote locations with little access to resources so as to ensure credible, free and fair elections in November 2021. Further, the results also present signals for socioeconomic reforms that can be beneficial for both the electorate and candidates.