Machine learning approaches to modelling marriage duration in South Africa

Abstract

Marriage is one of the most important social institutions, as it greatly influences the stability of families, population, and socio-economic well-being. The changing trends of marriages and divorces in South Africa confirm the importance of predictive analysis on the stability of marriages. Marriage duration prediction is a challenging task, as it is affected by a number of factors. Recent developments in machine learning techniques provide new opportunities to model this problem. This experiment used the following seven different classifications: Random Forest, Gradient Boosting, XGBoost, Logistic Regression, Support Vector Machine, K Nearest Neighbours, and Decision Tree, which aimed to forecast the categories of marriage length (short, medium, and long). The models used the following metrics: accuracy, precision, recall, F1-score, macro ROC-AUC, and the ROC curve. The results showed that ensemble-based approaches performed better than conventional classifiers. Gradient Boosting had the best performance in macro ROC AUC and was best at discriminating classes, especially for long and medium-length marriages. Random Forest had the best overall accuracy and was best at balancing classes. XGBoost had a similar performance to that of Random Forest in predicting. SVM had a high precision value but a lower recall value. Decision Trees and KNN had poor performance. Gradient Boosting was found to be the best-performing method for multi-class prediction of duration of marriage and achieved a balance between sensitivity and specificity. The results obtained from these studies prove that ensemble learning methods can be effectively utilized for modelling complex societal events.