Explainable machine learning model for the analysis of violence against women: a dialogue between data and feminist theory

Abstract

Childhood violence is a well-established determinant of multiple adverse outcomes in women’s adult lives and is consistently associated with subsequent experiences of intimate partner violence. This study identifies the predictive factors of exposure to childhood violence among Mexican women using an explainable machine learning approach articulated with the ecological model and feminist theory of gender-based violence. The analysis is based on an integrated dataset derived from the National Survey on the Dynamics of Household Relationships (ENDIREH) 2021 survey and nine complementary sources, comprising 61,205 observations and 67 predictors organized across four ecological levels. A CatBoost model was trained using stratified partitioning, cross-validation based hyperparameter optimization, threshold calibration, and SHAP-based interpretation.

On the test set, the model achieved an AUC of 0.771 (95% CI: 0.760-0.781), with a sensitivity of 0.606 and a specificity of 0.792. The history of violence block accounted for the largest predictive contribution, with witnessed, sexual, and partner-related childhood violence emerging as dominant predictors. These findings support the hypothesis of intergenerational transmission of violence and highlight the potential of explainable models for analyzing complex social phenomena