Measuring the Unmeasurable—A Study of Domestic Violence Risk Prediction and Management

ABSTRACT

The prevention of domestic violence (DV) have aroused serious concerns in Taiwan because of the disparity between the increasing amount of reported DV cases that doubled over the past decade and the scarcity of social workers. However, most common collaborations for DV prevention, between academic researchers and advocacy groups or governments outsourcing, often fail to produce effective prevention strategies. Hence, Data for Social Good Initiative (D4SG) worked with Taipei City Government to improve the efficiency of DV prevention and risk management on two levels—project collaboration level and data level. On the project collaboration level, we adopted a platform strategy and utilize public-private partnership (PPP) to connect and empower change agents across data silos from pilot runs to actual project execution. On the data level, we helped social workers differentiate the risk level of new cases by building a repeat victimization risk prediction model using random forest method with the 2015 data from Taipei City government’s DV database. The accuracy and F1-measure of our model were 96.3% and 62.8%. This projects’ PPP approach and quantification method successfully improved DV prevention process. These methodologies have also been applied to other work fields including firework prediction, emergency healthcare management as a paradigm.

KEYWORDS:

• Data for social good
• domestic violence prevention
• public–private partnership
• random forest model
• risk management

Acknowledgment

We deeply appreciate the partnership of the Department of Social Welfare of Taipei City and Taipei City Center for Prevention of Domestic Violence and Sexual Assault for providing the data and to DSP Inc. for offering strategic advice and technical support. We would also like to thank the social workers’ team from TPDVPC, Ying-Yi Chang, Yungji-Chih Huang, Chien-Sheh Chou, Hui-Chuan Lin, and Meng-Chuan Hsieh for contributing DV domain knowledge and assisting us in data correction. We also want to recognize the tireless work of Ya-Yun Chen, Brian Pan, and Tonyq Wang.