Abstract

Gender violence is an issue of public health which affects women and children globally. According to the UN, “35 percent of women worldwide have experienced either physical and/or sexual intimate partner violence or non-partner sexual violence” (UN Women, s/f). Indeed, WHO found that “almost one-third of all women who have had a relationship have suffered physical or sexual violence at the hands of their partner”. Within the scope of gender violence, femicide is a phenomenon that occurs as a consequence of cycles of violence against a woman, the rates of which continue to grow on a global level. “A total of 87,000 women were intentionally killed in 2017. More than half of them (58 per cent) ̶ 50,000 ̶ were killed by intimate partners or family members More than a third (30,000) were killed by their current or former intimate partner ̶ someone they would normally expect to trust”. 
Meanwhile, we are seeing great advances in AI and the use of machine learning and deep learning for the creation of algorithms for risk prediction. Tools that aim to determine the level of risk of femicide have been developed in Spain and Canada, for example: Viogen, “The Ontario Domestic Assault Risk Assessment” (ODARA), and “Domestic Violence Risk Appraisal Guide” (DVRAG) etc.  When building such tools and considering that risk determination will be carried out by an algorithm, it is pertinent to analyse how the algorithm should be built, how information is collected, how to decide which variables to include or exclude. Also, as the algorithm becomes autonomous thanks to machine learning, the so-called black box plays an important role.  We cannot know the internal workings of the algorithm and the way in which it determines the level of risk. 
Therefore, the question for investigation that arises is: Which variables need to be considered when building algorithms to determine risk in the prevention of gender violence? In order to answer this, an inductive qualitative methodology is used to analyze primary sources, secondary sources and case studies (algorithms). The results show that there is a need to evaluate situational and trigger factors, as well as factors related to the perpetrator, the victim and type of relationship (prior violence, threats of homicide)