Analysis of the fairness of the most common models in Predictive Security:

In this presentation, the notion of fairness in recommender systems is addressed, specifically in the context of the most widely used crime prediction algorithms in our region. The focus is on analyzing and comparing the levels of fairness for each model. To achieve this, a methodology is implemented that adjusts the distribution of the protected variable (i.e., income level) in a simulated study environment that reflects the dynamics of urban growth in Latin American cities. The analysis extends towards the comparison of the fairness results with the overall performance of the models, measured by the EMD (Earth Mover's Distance) metric. The findings reveal an interesting trend: models with higher performance tend to exhibit certain degrees of inequity. In particular, it is observed that models based on kernel density estimation (KDE) tend to exhibit lower levels of equity compared to other approaches. These results shed light on the relationship between the accuracy of models and the fairness they provide in the context of crime prediction. The presentation offers insight into how advances in crime prediction algorithms should not only focus on improving performance, but also on effectively addressing fairness considerations in their design and implementation.