Using unsupervised machine learning to find profiles of domestic abuse perpetrators

In this article we use unsupervised machine learning to discover hidden structures and patterns in a longitudinal police dataset of domestic abuse suspects, to provide a police force with an overarching or ‘baseline’ picture of how domestic abuse manifests locally. 3 algorithms were used to analyse 12 variables in a longitudinal dataset of over 40,000 suspects, organising them into discreet “clusters” or profiles with common characteristics and highlighting the differences and continuities between these. The quantitative findings, which highlighted clusters of abuse that had not previously been ‘on the radar’ of domestic abuse services in the specific force area, were then contextualised through qualitative interviews with a range of stakeholders to help identify priorities for intervention and further research. Our study shows how cutting-edge quantitative methods can be applied to improve understanding of prevalence and features of police-recorded abuse; draw attention to previously under-addressed types of abuse; serve as the groundwork for further, more in-depth research; and provide an evidence-base for local decision-making.