Exploring Transfer Learning For Crime Prediction

Exploring Transfer Learning for Crime Prediction
Abstract—Crime 1
I. INTRODUCTION 2
II. DATA ANALYSIS 3
III. FRAMEWORK OVERVIEW 4
IV. CONCLUSION 5
REFERENCES 6
2017 IEEE
2017 IEEE International
International Conference
Conference on
on Data
Data Mining
MiningWorkshops
Workshops
Exploring Transfer Learning for Crime Prediction

Xiangyu Zhao Jiliang Tang
Data Science and Engineering Lab
Data Science and Engineering Lab
Michigan State University
Michigan State University
Email: zhaoxi35@msu.edu
Email: tangjili@msu.edu
Abstract—Crime prediction plays a crucial role in addressing

crime, violence, conflict and insecurity in cities to promote determined by time and space, thus spatio-temporal patterns
good governance, appropriate urban planning and management. play a crucial role in crime analysis and have great potential
Plenty efforts have been made on developing crime prediction to help us build crime prediction model.
models by leveraging demographic data, but they failed to In this paper, we explore spatio-temporal patterns for crime
capture the dynamic nature of crimes in urban. Recently, with
the development of new techniques for collecting and integrating
prediction with urban data. To be specific, for temporal-
fine-grained crime-related datasets, there is a potential to obtain spatial patterns, we focus our investigation on (a) intra-region
better understandings about the dynamics of crimes and advance temporal patterns that help us to understand how crime
crime prediction. However, for a city, it is hard to build a uniform evolves over time for a region in a city and (b) inter-region
framework for all boroughs due to the uneven distribution of spatial patterns that suggest the geographical influence among
data. To this end, in this paper, we exploit spatio-temporal
patterns in urban data in one borough in a city, and then leverage
regions in the city. However, for a city, it is hard to build a
transfer learning techniques to reinforce the crime prediction of uniform framework for all boroughs due to the uneven
other boroughs. Specifically, we first validate the existence of distribution of data, e.g., the data of Brooklyn borough is
spatio-temporal patterns in urban crime. Then we extract the much sparser than Manhattan. Motivated by the existence of
crime-related features from cross-domain datasets. Finally we movement of citizens and commuting of public
propose a novel transfer learning framework to integrate these
features and model spatio-temporal patterns for crime transportations between different boroughs, there is a potential
prediction. to exploit transfer learning to make accurate crime prediction
for all boroughs. To this end, in this paper, we exploit spatio-
temporal patterns in urban data in one borough in a city, and
Index Terms—Crime Prediction, Transfer Learning, Spatio-
Temporal Patterns.
then leverage transfer learning techniques to reinforce the
crime prediction of other boroughs.
I. I NTRODUCTION
Recent studies have shown that crime prediction is closely II. DATA A NALYSIS
related to address crime, violence, conflict and insecurity in In this section, we will first introduce the dataset for this
cities to promote good governance, appropriate urban study, and then perform preliminary analysis about spatio-
planning and management[1]. Thus there is an increasing and temporal patterns in crime data.
urgent de- mand for accurate crime prediction. Efforts have
been made on understanding crime prediction model based on A. Data
demographic data [2], [3], [4], i.e., statistical socioeconomic
We collect crime-related data from July 1, 2012 to June 30,
characteristics of a population, but it is still challenging for
2013 in New York City. We divide NYC into 133 disjointed
researchers to accurately predict crime number due to the
regions, and each region is a 2km × 2km grid.
following facts (1) most regions in a city share similar
demographic features that cannot capture the differences • Public Security Data: We collect (1) crime complaint data
between different regions, and that contains the complaint frequencies of multiple types
(2) demographic features are stable over a long-term period of offenses, and (2) Stop-and-Frisk data, which includes a
that cannot capture the dynamics within a region[5]. NYC Police Department practice of temporarily detaining,
Recently, with the development of new techniques for col- questioning, and searching weapons.
lecting and integrating fine-grained urban data, a large amount • Meteorological Data: Meteorology and crime have been
crime-related datasets are recorded such as public safety data, found to be correlated[8]. Hence, we collect meteorological
meteorological data, point of interests data, human mobility data, consisting of weather, temperature, wind strength,
data and 311 public-service complaint data. These datasets precipitation, etc. In total, 30 features are collected from
contain fine-grained information about where and when the NYC meteorological stations every day.
data is collected and provide helpful context information • Point of interests (POIs): The density of POIs can charac-
about crime. Such information allows us to better understand terize the neighborhood functions, which could be helpful
the dynamics of crime and enables us to study spatial for crime prediction. We crawled point of interests from
factors of crime such as geographical influence. Meanwhile, FourSquare. In total, 11 categories of POIs are obtained.
crimi-$31.00
2375-9259/17 nal theories[6], [7] 10.1109/ICDMW.2017.165
© 2017 IEEE DOI suggest that crime distribution is
highly
2375-9259/17 $31.00 © 2017
IEEE DOI
1150
1158
• Human Mobility: Human mobility provides useful each region in the near future of other boroughs in the city
information, such as residential stability, which is related to police departments, and then assists them take actions to
to urban crime. We extract three features from this source, prevent crimes.
i.e., check- ins from the POI dataset, and pick-up & drop-off
points from the taxi trajectories dataset, which denote the
number of people arriving at or departing from the target
region.
• 311 Public-Service Complaint Data: 311 is NYC’s gov-
ernmental non-emergency service number, allowing people
to complain about things by making a phone call, which
shows the citizens’ dissatisfaction with government service,
thus it is highly related with crime.
B. An Analysis on spatio-temporal Pattern
In this subsection, we investigate spatio-temporal patterns
in crime data. To be specific, we focus on: (i) intra-region
temporal patterns and (ii) inter-region spatial patterns. We
could summarize the observations from our preliminary study
as follows:
For a region, we observe intra-region temporal patterns:
(1) for two adjacent time slots, they are likely to share
similar crime numbers; and (2) with the increase of Fig. 1. The Overall of the crime prediction system.
differences between two time slots, the crime difference has
the propensity to increase. IV. C ONCLUSION
For different regions within a borough, we note inter-region In this paper, we propose a novel framework STF, which
spatial patterns: (1) two spatially close regions have similar captures temporal-spatial patterns and leverages transfer learn-
crime numbers; and (2) with the increase of spatial distance ing for crime prediction. STF leverages cross-domain urban
between two regions, the crime difference tends to increase; datasets, e.g., public security data, meteorological data, point
and (3) for regions from different boroughs, they do not of interests (POIs), human mobility data and 311 complaint
follow above mentioned two inter-region spatial patterns. data.
The above observations provide the groundwork for our There are several interesting research topics, such as (1)
proposed transfer learning framework for crime prediction. Where: mining where is the hotspot for a specific category
III. F RAMEWORK O VERVIEW of crime such as robbery, (2) When: knowing which hours
are safe or unsafe for certain crimes, (3) Who: mining who
Figure 1 shows that the framework of our method com- are likely be criminals through user portrait on online social
prises of three major components: (i) feature extraction, (ii) network, (4) How: identifying the process of how an offense
transfer learning based framework, and (iii) crime prediction, happens, and (5) Why: exploring why an offense happens at a
as follows. specific time in a specific region. We will leave these as future
Step 1: Feature Extraction. We first extract features from investigation directions.
multiple datesets, such as crime complaint dataset, stop-and-
frisk dataset, meteorology, Point of Interests (POIs), human REFERENCES
mobility and 311 complaint dataset. Then we fuse these [1] C. Couch and A. Dennemann, “Urban regeneration and sustainable devel-
features into feature matrices. opment in britain: The example of the liverpool ropewalks partnership,”
Cities, vol. 17, no. 2, pp. 137–147, 2000.
Step 2: the proposed strategy STF. Based on feature [2] I. Ehrlich, “On the relation between education and crime,” in Education,
matrices and the historical crime numbers, we propose an income, and human behavior. NBER, 1975, pp. 313–338.
[3] B. P. Kennedy, I. Kawachi, D. Prothrow-Stith, K. Lochner, and V. Gupta,
approach to learn the model parameters for each region in “Social capital, income inequality, and firearm violent crime,” Social
each time slot for one borough. We exploit a spatio-temporal science & medicine, vol. 47, no. 1, pp. 7–17, 1998.
multi-task learning strategy to develop the predictive model, [4] J. Braithwaite, Crime, shame and reintegration. Cambridge University
Press, 1989.
which includes (1) intra-region temporal patterns that the [5] H. Wang, D. Kifer, C. Graif, and Z. Li, “Crime rate inference with big
crimes change smoothly over adjacent time and change over data,” in Proceedings of the 22nd ACM SIGKDD International
consecutive weeks periodically within a region, and (2) inter- Conference on Knowledge Discovery and Data Mining. ACM, 2016, pp.
635–644.
region spatial patterns that two spatially close regions tend to [6] L. E. Cohen and M. Felson, “Social change and crime rate trends: A
have similar crime numbers. Then, based on the parameters of routine activity approach,” American sociological review, pp. 588–608,
one borough, we leverage transfer learning techniques to train 1979.
[7] D. B. Cornish and R. V. Clarke, The reasoning criminal: Rational choice
the parameters W of other boroughs in the city. perspectives on offending. Transaction Publishers, 2014.
Step 3: Crime Prediction. Based on the well trained [8] E. G. Cohn, “Weather and crime,” British journal of criminology, vol. 30,
parameters W, our framework provides crime prediction of no. 1, pp. 51–64, 1990.
1159
1151

Exploring Transfer Learning For Crime Prediction

Uploaded by

Copyright:

Available Formats

Exploring Transfer Learning For Crime Prediction

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Exploring Transfer Learning For Crime Prediction

Uploaded by

Copyright:

Available Formats

Exploring Transfer Learning for Crime Prediction

Exploring Transfer Learning for Crime Prediction

Abstract—Crime prediction plays a crucial role in addressing

You might also like