Supervised machine learning algorithms for classifications of gender-based violence in Somalia: a comparison of oversampling techniques

Gender-based violence can include sexual, physical, mental, and economic harm inflicted in public or in private. This violence also has a direct psychological effect, physical and financial consequences, and it has multiple underlying reasons, such as social, economic, cultural, political, and religious aspects. By applying multiple resampling techniques, this study aims to improve the precision and accuracy of supervised machine learning classifications of gender-based violence (GBV) using the SDHS dataset. The class imbalance between GBV-positive and GBV-negative instances makes it very challenging to produce reliable classification machine learning models. To address this issue, oversampling machine learning approaches, including synthetic minority over-sampling technique (SMOTE), adaptive synthetic (ADASYN), and random over-sampling (ROS), were employed to classify the GBV data in Somalia. The logistic regression (LR), decision tree (CART), random forest (RF), naïve Bayes (NB), k-nearest Neighbors (KNN), and support vector machine (SVM) methods were trained and evaluated. In addition, oversampling techniques were employed for improving the imbalanced datasets. Receiver operating characteristic curve (ROC) and the area under the curve (AUC) were used to assess each machine learning classifier and to compare performance on the original GBV dataset. Among the resampling techniques, SMOTE (RF = 0.992, CART = 0.969, and KNN = 0.957) outperformed ADASYN (RF = 0.912, CART = 0.910, and KNN = 0.876) and ROS (RF = 0.920, CART = 0.919, and KNN = 0.880) across almost all evaluation metrics. The classifiers that performed the best were random forest (RF) and classification and regression trees (CART), then k-nearest Neighbors. After resampling the imbalanced dataset, we may therefore conclude that the random forest (AUC = 0.972), CART (AUC = 0.969) and KNN (AUC = 0.957) machine learning classifiers are better at accurately classifying the k-nearest Neighbors dataset. In addition, compared to the other oversampling techniques, SMOTE was used to the machine learning classifiers to balance the imbalanced class distributions in favour of the minority class. In addition, SMOTE with the Mathews correlation coefficient (MCC) outperformed ADASYN and ROS resampling techniques. The MCC values for SMOTE reached their highest values (RF = 0.86, CART = 0.85, and KNN = 0.80), indicating strong overall predictive reliability of the machine learning models. Therefore, the findings of this analysis will assist government and non-government organizations in making policy decisions to GBV risks.