Alvaro Riascos (Universidad de los Andes)
12 de Noviembre 2020
In this paper we address the accuracy vs. fairness trade-off in state of the art machine learning algorithms for crime prediction. We propose a novel conditional GANs architecture for crime (robberies) prediction in Bogota, capital city of Colombia. The model uses several layers of ConvLSTM neural nets in both the generative and the discriminatory networks. We further condition on past crime intensity maps, weekdays, and holidays. The trained network is able to capture spatio-temporal patterns and outperforms state-of-the-art predictive models such as spatio-temporal Poisson point process, as well as other models trained with the same dataset. Model’s accuracy reaches an area under the Hit Rate – Percentage Area Covered by Hotspots curve of 0.86. However, our predictions suggest that there is a potential bias with heterogeneous effects on vulnerable populations. We address the fairness consequence of this model in low income vs. high income residents by estimating a calibration test conditional to these protected variables. Finally, we introduce a fairness – accuracy balancing technique that quantifies the trade-offs between accuracy and fairness in this type of models. This technique notably reduces bias with a marginal effect on accuracy. The previous model addresses the accuracy vs. fairness trade-off when the protected variable is low income vs. high income populations. However, this distinction is quite artificial in our data set since we use a proxy of income based on administrative data records. Therefore, we consider a data driven urban characterization based on city images that better reflect the living, socioeconomic and environmental conditions of the city. To do this we use state-of-the-art crime prediction models of spatio-temporal clustering including environmental covariates extracted through convolutional networks from street-level images. Our model using image features as covariates outperforms a standard model of self-exciting point process and it’s well calibrated, even before we try to fix it using ex post Platt scaling.