Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep Learning Based Crime Prediction Models: Experiments and Analysis

Rittik Basak Utsha, Muhtasim Noor Alif, Yeasir Rayhan, Tanzima Hashem, Mohammad Eunus Ali

TL;DR

This study offers a unified, longitudinal comparison of seven deep learning models for crime prediction, addressing spatial, temporal, and categorical dependencies under varied real-world scenarios. By evaluating models such as AIST, HAGEN, and ST-SHN across area, density, and temporal granularity, the authors derive clear guidance on when to favor external features, how to structure spatial and temporal modules, and how regression and classification tasks motivate different architectural choices. The key contributions include a comprehensive experimental framework, a granular ablation analysis of spatial and external components, and actionable recommendations for model design and deployment in practice. The findings have practical impact for city safety analytics, informing practitioners on model selection and design tradeoffs in diverse operational contexts.

Abstract

Crime prediction is a widely studied research problem due to its importance in ensuring safety of city dwellers. Starting from statistical and classical machine learning based crime prediction methods, in recent years researchers have focused on exploiting deep learning based models for crime prediction. Deep learning based crime prediction models use complex architectures to capture the latent features in the crime data, and outperform the statistical and classical machine learning based crime prediction methods. However, there is a significant research gap in existing research on the applicability of different models in different real-life scenarios as no longitudinal study exists comparing all these approaches in a unified setting. In this paper, we conduct a comprehensive experimental evaluation of all major state-of-the-art deep learning based crime prediction models. Our evaluation provides several key insights on the pros and cons of these models, which enables us to select the most suitable models for different application scenarios. Based on the findings, we further recommend certain design practices that should be taken into account while building future deep learning based crime prediction models.

Deep Learning Based Crime Prediction Models: Experiments and Analysis

TL;DR

This study offers a unified, longitudinal comparison of seven deep learning models for crime prediction, addressing spatial, temporal, and categorical dependencies under varied real-world scenarios. By evaluating models such as AIST, HAGEN, and ST-SHN across area, density, and temporal granularity, the authors derive clear guidance on when to favor external features, how to structure spatial and temporal modules, and how regression and classification tasks motivate different architectural choices. The key contributions include a comprehensive experimental framework, a granular ablation analysis of spatial and external components, and actionable recommendations for model design and deployment in practice. The findings have practical impact for city safety analytics, informing practitioners on model selection and design tradeoffs in diverse operational contexts.

Abstract

Crime prediction is a widely studied research problem due to its importance in ensuring safety of city dwellers. Starting from statistical and classical machine learning based crime prediction methods, in recent years researchers have focused on exploiting deep learning based models for crime prediction. Deep learning based crime prediction models use complex architectures to capture the latent features in the crime data, and outperform the statistical and classical machine learning based crime prediction methods. However, there is a significant research gap in existing research on the applicability of different models in different real-life scenarios as no longitudinal study exists comparing all these approaches in a unified setting. In this paper, we conduct a comprehensive experimental evaluation of all major state-of-the-art deep learning based crime prediction models. Our evaluation provides several key insights on the pros and cons of these models, which enables us to select the most suitable models for different application scenarios. Based on the findings, we further recommend certain design practices that should be taken into account while building future deep learning based crime prediction models.
Paper Structure (35 sections, 3 equations, 2 figures, 15 tables)

This paper contains 35 sections, 3 equations, 2 figures, 15 tables.

Table of Contents

  1. Introduction
  2. Deep Learning Models for Crime Prediction
  3. Model Architectures
  4. DeepCrime
  5. MiST: Multi-View Deep Spatial-Temporal Network
  6. CrimeForecaster
  7. HAGEN: Homophily-Aware Graph Convolutional Recurrent Network
  8. ST-SHN: Spatial-Temporal Sequential Hypergraph Network for Crime Prediction with Dynamic Multiplex Relation Learning
  9. ST-HSL: Spatial-Temporal Hypergraph Self-Supervised Learning for Crime Prediction
  10. AIST
  11. Comparative Analysis
  12. Sub-components of Model Architectures
  13. Utilization of Crime and External Datasets
  14. Different Types of Prediction
  15. Experimental Settings
  16. ...and 20 more sections

Figures (2)

  • Figure 1: (a) Spatial Dependency: Crimes can exhibit different patterns for different regions. The communities of Chicago exhibit a wide range of number of crimes in 2019. (b) Temporal Dependency: Crime frequency can be different on different time of the day. The earlier hours of a day have less occurrences of crimes than later in the day in our dataset. (c) Categorical Dependency: One crime can be dependent on another. A heatmap depicting the correlations between the crime categories in our dataset is shown.
  • Figure 2: The Architectural Synopses of Deep Learning Based Crime Prediction Models. (MLP: Multi-Layer Perceptron, FC: Fully Connected NN, LSTM: Long Short-Term Memory, GRU: Gated Recurrent Unit, RNN: Recurrent Neural Network, SAB-LSTM: Sparse Attention-Based LSTM, DCGRU: Diffusion Convolution GRU, GCN: Graph Convolutional Network, GAT: Graph Attention Network, DGCN: Diffusion GCN, GNN: Graph Neural Network, HNN: Hypergraph Neural Network)