CrimeForecaster: Crime Prediction by Exploiting the Geographical Neighborhoods’ Spatiotemporal Dependencies

Abstract

Crime prediction in urban areas can improve the allocation of resources (e.g., police patrols) towards a safer society. Recently, researchers have been using deep learning frameworks for urban crime forecasting with better accuracies as compared to previous work. However, these studies typically partition a metropolitan area into synthetic regions, e.g., grids, which neglects the geographical semantics of a region, nor captures the spatial correlation across the regions, e.g., precincts, neighborhoods, blocks, and postal division. In this paper, we design and implement an end-to-end spatiotemporal deep learning framework, dubbed CrimeForecaster, which captures both the temporal recurrence and the spatial dependency simultaneously within and across regions. We model temporal dependencies by using the Gated Recurrent Network with Diffusion Convolution modules to capture the cross-region dependencies at the same time. Empirical experiments on two real-world datasets showcase the effectiveness of CrimeForecaster, where CrimeForecaster outperforms the current state-of-the-art algorithm by up to 21%. We also collect and publish a ten-year crime dataset in Los Angeles for future use by the research community.

Z. Lin—The work was conducted and completed while the author was an intern at USC.

Appendix A Impact of the Adjacent Neighborhoods’ Information

Fig. 6.

Forecasting results for all crimes on three different neighborhoods in Los Angeles and Chicago. (Los Angeles: SO – Sherman Oaks; EP – Echo Park; SC – Studio City. Chicago: AG – Auburn Gresham; FP – Fuller Park; RD: RiverDale)

We have observed that the spatial correlation would help with the crime prediction in Sect. 1. Here we want to know experimentally about if the neighborhood information helps to predict the crimes for a target individual neighborhood. Formally, for each neighborhood $r_i$, we would like to know if its neighborhoods’ information helps to predict the crime event in the next time slot given the previous records. Therefore, we train the same classification models on two different feature settings: a) using the previous crime occurrences of $r_i$ itself as features, and b) using the previous crime occurrences of both $r_i$ itself and the aggregates of $r_i$’s neighborhoods.

$$\begin{aligned} f\big (y_{i,l}^{K+1}|(y_{i,l}^{1}, \dots , y_{i,L}^{K})\big ) \quad \text {V.S.}\quad f\big (y_{i, l}^{K+1}|\text {agg}(\{Y_{\mathcal {N}(1)}, \dots , Y_{\mathcal {N}(M)}\})\big ) \end{aligned}$$

where $Y_{\mathcal {N}(1)}=(y_{\mathcal {N}(1),l}^1, \dots , y_{\mathcal {N}(1),L}^K)$, and L is the number of crime categories. Here f is a prediction function, which can be classification models, deep learning models and etc. $\mathcal {N}(m)$ denotes an adjacent neighbor of $R_i$, i.e., $W_{i,{\mathcal {N}(m)}} = 1$. In this experiment, we use classical machine learning classifiers: Logistic Regression (LR), K-Nearest Neighbor (KNN), Extra Tree (ET), Decision Tree (DT) and Random Forest (RF) classifiers.

In both Fig. 6 (a) and Fig. 6 (b), we can see that the performance of simple models improves a lot by simply appending its neighborhoods’ information. In the next following sessions, we explored other ways of dealing with the spatial information to make the model applicable to predict the crime at multiple locations, such as proposing MiST* by adding network embedding learned from LINE [25] or using the diffusion graph convolution in CrimeForecaster. CrimeForecaster shows great superiority over other methods. We show our experiment results in the Sect. 5.