Comparison of Machine Learning Algorithms for Predicting Crime Hotspots

Xu Zhang,Lin Liu,Jiakai Ji,Luzi Xiao

doi:10.1109/access.2020.3028420

Xu Zhang, Lin Liu + Show 2 more

Open Access

https://doi.org/10.1109/access.2020.3028420

Copy DOI

Abstract

Crime prediction is of great significance to the formulation of policing strategies and the implementation of crime prevention and control. Machine learning is the current mainstream prediction method. However, few studies have systematically compared different machine learning methods for crime prediction. This paper takes the historical data of public property crime from 2015 to 2018 from a section of a large coastal city in the southeast of China as research data to assess the predictive power between several machine learning algorithms. Results based on the historical crime data alone suggest that the LSTM model outperformed KNN, random forest, support vector machine, naive Bayes, and convolutional neural networks. In addition, the built environment data of points of interests (POIs) and urban road network density are input into LSTM model as covariates. It is found that the model with built environment covariates has better prediction effect compared with the original model that is based on historical crime data alone. Therefore, future crime prediction should take advantage of both historical crime data and covariates associated with criminological theories. Not all machine learning algorithms are equally effective in crime prediction.

Highlights

Spatiotemporal data related to the public security have been growing at an exponential rate during the recent years
CRIME PREDICTION WITH MACHINE LEARNING ALGORITHMS The traditional methods usually detect the crime hotspot area from the historical distribution of crime cases, and assume that the past pattern is to be repeated in the future [7], [2]
According to the experimental results, we found that the prediction accuracy of the prediction accuracy of the Long Short-Term Memory (LSTM) model was improved after adding built environment covariates, and the average prediction index-HitRa of 13 experimental periods increased by percentage points increased by 12.8 percentage points, the average prediction index-HitRn of experimental periods increased by percentage points, and the average prediction index-HitEn of 13 experimental periods increased by 10.4 percentage points

Summary

INTRODUCTION

Spatiotemporal data related to the public security have been growing at an exponential rate during the recent years. C. CRIME PREDICTION WITH MACHINE LEARNING ALGORITHMS The traditional methods usually detect the crime hotspot area from the historical distribution of crime cases, and assume that the past pattern is to be repeated in the future [7], [2]. CRIME PREDICTION WITH MACHINE LEARNING ALGORITHMS The traditional methods usually detect the crime hotspot area from the historical distribution of crime cases, and assume that the past pattern is to be repeated in the future [7], [2] This assumption tends to be reasonable for predicting long-term stable crime hotspots. The ability of machine learning algorithm in processing non-linear relational data has been confirmed in many fields, including crime prediction.

EVALUATION INDICATOR

SELECTION OF CRIME TYPES

DATA VISUALIZATION ANALYSIS

CONCLUSION