Dynamic Graph Mamba Model for Crime Count Forecasting Based on Adaptive Geographic Fusion

Wadhah R. Baiee, Lecturer (PhD), University of Babylon, College of IT, Software Dep, Hilla, IraqFollow

ORCID

Wadhah R. Baiee: https://orcid.org/0000-0001-7237-8633

Article Type

Original Study

Abstract

Daily crime counts are hard to predict due to variations in the patterns of crime across neighborhoods and over time, as well as that the patterns vary by offence type. Areas with a shared border can give different information for forecasts than areas with similar crime histories. This study investigates if the joint use of these two relationships would give better prediction than using one or the other, and whether the algorithm can be adapted to the frequent occurrence of zeros and variance greater than the mean. It evaluates the functionality of Dynamic Graph Mamba for Crime Prediction (DG-Mamba-CP), a Mamba2 model with a fixed geographic graph and category-specific correlation graphs based on the previous 60 days. The model predicts the number of incidents for six types of crimes in 77 community areas in Chicago 1, 7, 14 and 28 days ahead. The data from 2018 to 2021 was prepared for training period usage, the data from 2022 was used for the checkpoint selection and conformal calibration, and the data from 2023 was used for rolling evaluation. 40 runs were performed, with eight configurations tested five times with various seeds values taken randomly. The lowest MAE was obtained for the adaptive fusion with a negative binomial objective (Mean Absolute Error (MAE) = 0.918937 ± 0.006547) while the lowest Root Mean Square Error (RMSE) was obtained with the zero-inflated negative binomial variant (RMSE = 1.427633 ± 0.015562). With the same zero-inflated objective, adaptive fusion made 11.56% reduction in MAE compared to temporal-only modeling and geographic-only modeling made 10.45% reduction. The smaller edge of the adaptive fusion over simpler spatial variants was not supported statistically over seeds. The empirical coverage rate for 90% of nominal conformal intervals was 90.04% with a mean interval width of 2.9018. The findings support incorporating geographic relationships into daily crime-count forecasting, while the additional value of adaptive fusion remains uncertain.

Keywords

Crime Forecasting, Mamba2, Spatial graph, Chicago

How to Cite This Article

Baiee, Wadhah R. (2026) "Dynamic Graph Mamba Model for Crime Count Forecasting Based on Adaptive Geographic Fusion," Journal of Intelligent Informatics, Networking, and Cybersecurity: Vol. 2 : Iss. 2 , Article 13.
Available at: https://doi.org/10.65445/3106-1192.1024