Characterizing Intimate Partner Violence against Married Women in Ethiopia with Machine Learning

Abstract Background Intimate Partner Violence (IPV) is the most common form of violence against women worldwide, perpetrated mostly by intimate partners (men) against women. According to the 2020 United Nations report, 47000 women were killed by their intimate partners or other family members globally. IPV can be categorized as physical, psychological, or sexual violence which may cause a serious and sometimes fatal occurrence. Objective The main goal of this study is to characterize IPV against married women in Ethiopia by developing a machine learning model to 1) predict types of IPV, 2) identify risk factors of IPV, and 3) extract relevant rules for evidence-based strategy and policy formulation to prevent, reduce and control IPV against married women in Ethiopia. Methods The study was conducted following an experimental research approach. The data were gathered from the Ethiopian demographic health survey and preprocessed to make it suitable for the machine learning algorithm. Extreme gradient boosting, Cat boost, Random Forest, and Light GBM with one versus rest class decomposition were employed to develop an IPV prediction model. For building the proposed model a total of 31110 instances with 22 features, and 80/20 training and testing dataset split ratio were used. Result Performance evaluation metrics were used to choose the best-performing algorithm. The overall accuracy of extreme gradient boosting, Cat boost, RF, and Light GBM is 92.57%, 94.39%, 92.94%, and 90.30%, respectively. Feature importance analysis was implemented to identify risk factors of IVP again married women in Ethiopia. Conclusion Cat boost algorithms registered the highest performance with 94.39% accuracy. The most determinant risk factors of intimate partner violence against married women in Ethiopia are women’s household empowerment status, working status, place of residence, first marriage decision, and husband's education level, to mention a few. Relevant rules that may support evidence-driven strategies and policies to prevent, reduce and control IPV against married women in Ethiopia were also extracted from the best performing algorithm.