Feminicide & Machine Learning
Feminicide and gender-based violence are critical issues, especially in Latin America and the Caribbean, where laws often lack enforcement and data collection.
Activists bridge data gaps by compiling feminicide incidents from news, facing resource and mental health challenges.
A participatory action research project automates feminicide detection via machine learning to aid activists.
A model trained on media reports achieved 81.1% accuracy in identifying feminicides, showing promise in reducing manual labor.
An interactive feminicide notification system is being co-designed with activists.
Feminicide, defined as the misogynous killing of women by men, includes state complicity through data collection neglect.
Activist movements like Argentina’s “Ni una menos” raise awareness and drive policy change.
Data Feminism principles guide the research, addressing power imbalances in data collection.
Activists' counterdata collection efforts monitor feminicides, highlighting official statistics inadequacies.
The project aims to support activists by developing a machine learning model to classify news articles as feminicides and create an interactive notification application.
The research uses a mixed-methods design, including co-design workshops and qualitative interviews with activists.
The goal is to honor the lives behind the data points and support memorialization practices.
A machine learning system was developed to predict the probability of feminicide from news articles, which aids activists prioritize articles that are more likely to be relevant.
The dataset was created by collecting, labeling, and resolving discrepancies in news articles, resulting in a labeled dataset tailored to feminicide identification.
A multinomial naive Bayes model was trained to predict feminicide probability, showing that sorting articles by predicted probability improves efficiency.
The system architecture uses Media Cloud to pull and sort articles by feminicide likelihood based on user-defined queries.
Extensions include entity extraction, article grouping, and real-time model updating based on user feedback.
The project addresses open questions about the limitations of media reports, language and geography differences, standardization, and participatory methods in machine learning.
The conclusion highlights the development of an automated system to detect feminicides, emphasizing the importance of participatory action research methods.