Improving the Fairness of Deep-Learning, Short-term Crime Prediction with Under-reporting-aware Models

TL;DR

This work tackles bias in short-term, place-based crime prediction arising from under-reporting and demographic disparities. It introduces an under-reporting-aware deep learning architecture with two branches: a true-crime predictor using neighbor-convolution on crime and mobility features, and a crime-reporting convolutional gate estimating the under-reporting rate $\\pi_i$ from ACS determinants. Training optimizes predicted reported crimes $z_{i,t}$ to match observed data, while inference uses the true-crime branch to identify next-day hotspots. Across four US cities and two crime types, the approach improves fairness metrics (SP, FPR, FNR, LI) relative to baselines, albeit with a trade-off in accuracy, highlighting the practical balance between equitable predictions and predictive performance.

Abstract

Deep learning crime predictive tools use past crime data and additional behavioral datasets to forecast future crimes. Nevertheless, these tools have been shown to suffer from unfair predictions across minority racial and ethnic groups. Current approaches to address this unfairness generally propose either pre-processing methods that mitigate the bias in the training datasets by applying corrections to crime counts based on domain knowledge or in-processing methods that are implemented as fairness regularizers to optimize for both accuracy and fairness. In this paper, we propose a novel deep learning architecture that combines the power of these two approaches to increase prediction fairness. Our results show that the proposed model improves the fairness of crime predictions when compared to models with in-processing de-biasing approaches and with models without any type of bias correction, albeit at the cost of reducing accuracy.

Figures (4)

• Figure 1: Under-reporting-aware short-term crime prediction with crime-reporting convolutional gate. The 2D feature maps for historical crimes, mobility features and under-reporting determinants are constructed based on the neighboring set for census tract $s_1$ in the same way as shown in Figure \ref{['s3-fig:nbcnn_rearrange']}.
• Figure 2: Arrange the nearest neighbors set for the target census tract $s_1$ and construct the 2D feature map for historical crimes. In the neighboring set of $s_1$, $s_2$ and $s_3$ is the closest to $s_1$; $s_4$ and $s_5$ are the next closest to $s_2$ and $s_3$ respectively; $s_6$ and $s_7$ are the next closest to $s_4$ and $s_5$; $s_8$ and $s_9$ are the next closest to $s_6$ and $s_7$. Similar process is applied to each of the ten mobility features.
• Figure 3: Degrees of unfairness of property crime prediction for four cities (Baltimore, Minneapolis, Austin and Chicago). Results are shown for each fairness metric explained in Section 5.3 and for each protected group. Crime prediction models include under-reporting-unaware model (UU), UU with historical crimes only (UU(C)), UU with individual-based fairness gap regularization, and the proposed under-reporting-aware model (TC).
• Figure 4: Degrees of unfairness of violent crime prediction for four cities (Baltimore, Minneapolis, Austin and Chicago). Results are shown for each fairness metric explained in Section 5.3 and for each protected group. Crime prediction models include under-reporting-unaware model (UU), UU with historical crimes only (UU(C)), UU with individual-based fairness gap regularization, and the proposed under-reporting-aware model (TC).