Abstract
Abstract. Long-term unemployment (LTU), being unemployed consecutively for a year or more, is disproportionately higher for women in South Africa. The recent literature has demonstrated the application of machine-learning (ML) techniques in facilitating evidence-based decision making within the labour market analysis. However, there is a limitation of the application of decision tree (DT), random forest (RF) and eXtreme Gradient Boosting (XGBoost) models in LTU classification of women in South Africa. This research first investigated the relationship between LTU and socio-demographic features, including age, highest level of education, work experience, marital status and population group, with the application of Pearson correlation coefficient in comparison with Phi coefficient and Chi-squared versus Mutual information. The correlation analysis revealed that savings from previous employment and work experience were the most significant features in explaining LTU of women in South Africa. A baseline model, the binary logistic regression model along with ML models, were trained using these socio-demographic features and the XGBoost outperformed all models in some performance metrics. The performance metrics for XGBoost were accuracy score of 81%, precision of 83%, recall of 97% and AUC of 77%. These findings are consistent with previous studies, demonstrating the importance of ML in studying LTU, compared to traditional statistical models. Thus, this research suggests that considering advanced ML models in studying labour market dynamics can significantly improve the prediction of LTU among women in South Africa.