Content uploaded by Monsuru Adepeju
Author content
All content in this area was uploaded by Monsuru Adepeju on Jun 18, 2016
Content may be subject to copyright.
CPC:
CRIME,
POLICING &
CITIZENSHIP
INTELLIGENT POLICING
AND BIG DATA
Crime Po icingCitizenship
010001110010110111010101011100011
101
g
g
111000010101101011100001
10
p
p
0
p
p
0
11
000
1
0
111space-time interactions of dynamic networks
11
0000
1
0
1
0
1
0
1
0
11
0
1
0
1
0
1
0
1111
000
1
0
1
0
111
000
1
0
11
0000
11
00
RESEARCH DIRECTORS
Tao Cheng
Principal Investigator, UCL SpaceTimeLab
Kate Bowers
Co-Investigator, UCL Crime Science
Paul Longley
Co-Investigator, UCL Geography
John Shawe-Taylor
Co-Investigator, UCL Computer Science
Trevor Adams
Industrial Partner, Head of Data Development
Team, Metropolitan Police Service
RESEARCH TEAM
Toby Davies
UCL SpaceTimeLab & Crime Science
Gabriel Rosser
UCL SpaceTimeLab
Sarah Wise
UCL SpaceTimeLab
Chris Gale
UCL SpaceTimeLab & Geography
Monsuru Adepeju
UCL SpaceTimeLab
Jianan Shen
UCL SpaceTimeLab
Huanfa Chen
UCL SpaceTimeLab
Dawn Williams
UCL SpaceTimeLab
Kira Kempińska
UCL Computer Science & Geography
Artemis Skarlatidou
UCL ExCiteS
Published by UCL SpaceTimeLab, 2016.
Design Atelier Works
ISBN: 978-0-9954939-1-9
To cite this report: Cheng T, Bowers K, Long-
ley P, Shawe-Taylor J, Davies T, Rosser G,
Wise S, Gale C, Adepeju M, Shen J, Chen H,
Williams D, Kempińska K and Skarlatidou A
(2016). CPC: Crime, Policing and Citizenship –
Intelligent policing and big data. UCL Space-
TimeLab: London.
SpaceTimeLab
Chadwick Building
University College London
Gower Street
London
WC1E 6BT
T +44 (0)20 7679 2738
E tao.cheng@ucl.ac.uk
W www.ucl.ac.uk/spacetimelab
The Crime, Policing and Citizenship (CPC):
Space-Time Interactions of Dynamic Net-
works project was supported by EPSRC
under grant EP/J004197/1.
CONTENTS
Executive Summary
PART 1 INTRODUCTION
Background
– Challenges facing digital policing
– Big Data in policing
– Intelligent data-driven policing
– Our approach
PART 2 RESEARCH
Crime prediction & evaluation
– Space-time prediction of crime
– Prediction on street networks
– Evaluation of predictive algorithms
Guardianship & cooperative policing
– Quantifying deterrence
– Police route choice modelling
– Characterising foot patrol behaviour
– Strategic planning for ocer tasking
– Routing strategies for foot patrol
Public condence & trust in policing
– Trust and condence in policing
– Small area estimation of public condence
– Geodemographics and public condence
PART 3 IMPACT
Tools
– Predictive mapping
– Map-matching
– Supply and demand
– Online cooperative patrol routing
– Behaviour and activity analysis
Policy implications
Publications
5
5
6
7
8
10
12
14
16
18
20
22
24
26
28
30
31
31
32
32
3
34
36
2|3
Crime, Policing and Citizenship (CPC) – Space-
Time Interactions of Dynamic Networks has been
a major UK EPSRC-funded research project. It has
been a multidisciplinary collaboration of geoinfor-
matics, crime science, computer science and geog-
raphy within University College London (UCL), in
partnership with the Metropolitan Police Service
(MPS). The aim of the project has been to develop
new methods and applications in space-time analyt-
ics and emergent network complexity, in order to
uncover patterning and interactions in crime, polic-
ing and citizen perceptions. The work carried out
throughout the project will help inform policing
at a range of scales, from the local to the city-wide,
with the goal of reducing both crime and the fear of
crime.
The CPC project is timely given the tremendous
challenges facing policing in big cities nationally
and globally, as consequences of changes in society,
population structure and economic well-being. It
addresses these issues through an intelligent ap-
proach to data-driven policing, using daily reported
crime statistics, GPS traces of foot and vehicular
patrols, surveys of public attitudes and geo-tem-
poral demographic data of changing community
structure. The analytic focus takes a spatio-temporal
perspective, reecting the strong spatial and tem-
poral integration of criminal, policing and citizen
activities. Street networks are used throughout as
a basis for analysis, reecting their role as a key
determinant of urban structure and the substrate on
which crime and policing take place.
The project has presented a manifesto for ‘in-
telligent policing’ which embodies the key issues
arising in the transition from Big Data into action-
able insights. Police intelligence should go beyond
current practice, incorporating not only the predic-
tion of events, but also how to respond to them, and
how to evaluate the actions taken.
Cutting-edge network-based crime prediction
methods have been developed to accurately predict
crime risks at the street segment level, helping
police forces to focus resources in the right places
at the right times. Methods and tools have been
implemented to support senior ofces in strategic
planning, and to provide guidance to frontline of-
cers in daily patrolling. To evaluate police perfor-
mance, models and tools have been developed to aid
identication of areas requiring greater attention,
and to analyse the patrolling behaviours of ofcers.
Methods to understand and model condence in po-
licing have also been explored, suggesting strategies
by which condence in the police can be improved
in different population segments and neighbour-
hood areas.
A number of tools have been developed dur-
ing the course of the project including data-driven
methods for crime prediction and for performance
evaluation. We anticipate that these will ultimately
be adopted in daily policing practice and will play
an important role in the modernisation of policing.
Furthermore, we believe that the approaches to the
building of public trust and condence that we
suggest will contribute to the transformation and
improvement of the relationship between the public
and police.
EXECUTIVE SUMMARY
4|5
INTRODUCTION
CHALLENGES FACING DIGITAL POLICING
Crime and disorder are long-standing societal
problems which continue to blight the emotional
and economic well-being of citizens. Predicting,
preventing and mitigating crime and disorder is
fundamental to the evolution of modern societies
and the cities they live in.
Policing is crucial to public safety. London’s
Metropolitan Police Service (MPS) responds to over
10,000 calls every day, and is also responsible for a
wide range of crime prevention duties. It faces huge
challenges, arising not just from resourcing issues
but also the very way we think about the role of
policing.
The mission of policing is evolving, with issues
such as public condence gaining increased rec-
ognition as important police performance metrics.
The improvement of condence is a key priority for
policymakers such as the Mayor’s Ofce for Po-
licing and Crime, and is seen as a crucial factor in
crime prevention and detection. The fact that recent
achievements in crime reduction have not been
reected in public condence suggests an enduring
‘reassurance gap’ between successful policing and
the public perception of it.
Ongoing reductions in funding are placing
increased strain on police resources: the MPS, for
example, must make 20% efciency savings by
2020. This is not accompanied by any reduction
in targets, and indeed the MPS must also achieve a
20% decrease in crime and a 20% improvement in
public condence over the same period. ‘More for
less’ is the mantra for policing today in London.
BIG DATA IN POLICING
While many aspects of police resourcing are becom-
ing increasingly constrained, however, there is one
respect in which forces are becoming dramatically
richer: access to digital data. Not only is crime re-
corded more comprehensively than ever before, but
technological advances such as GPS tracking offer
unprecedented insight into the way that policing
itself is undertaken. Following the launch of a £2
million scheme in 2010, for example, MPS police
radios record ofcer locations at 5-minute intervals,
while vehicle locations are logged every 15 seconds.
These data sources are complemented by a num-
ber of others which offer insight into the key issues
of public perception and engagement. The Crime
Survey for England & Wales provides information
about the experience of crime, while the MPS Public
Attitude Survey and Victim Survey focus particular-
ly on issues of condence in the London context.
When combined with geodemographic data, these
20
30
40
50
60
70
80
Local National
2011/122010/112009/102008/0922007/082006/072005/062004/052003/042002/032001/02
Percentage perceiving more crime
Figure 1: Trends in the perception of crime levels
show that, paradoxically, people believe that crime is
increasing at a national level, but not in their local area
(Source: British Crime Survey).
sources allow the inuence of the potentially huge
number of factors that shape criminal activity and
public perceptions to be explored.
A number of recent developments, broadly
themed around the concept of ‘predictive policing’,
testify to the potential of data-driven law enforce-
ment. There remain, however, a number of limita-
tions to these techniques, particularly with respect
to their integration in real-world policing. In many
instances these may arise because of a siloed ap-
proach to crime reduction that brings focus to one
element of policing without considering the in-
ter-dependencies between its various aspects.
The success of any algorithmic solution de-
pends crucially on how it translates into actionable
policing: the success of a predictive algorithm, for
example, becomes immaterial if the ensuing police
response is insufcient to prevent the anticipated
crimes. Furthermore, little is understood about
the impact on public perception of geographically
focussed policing strategies; yet this is a crucial con-
cern, given the ever-increasing pressure to improve
public condence.
These issues can only be addressed by using
the datasets available to the police in an integrated
way, modelling and accounting for the signicant
inter-dependencies between them. Criminality and
policing are inter-related phenomena which occur
against a mosaic of different neighbourhood con-
texts. Close examination and interpretation of how
the detailed patterns of police activities relate to the
space-time characteristics of criminal incidents is
required not only for effective policing, but also for
effective reassurance of the public.
Ensuring that data are used intelligently is thus
key to improving the efciency and effectiveness of
operational practice in the era of digital policing.
Innovation in best practice through collaborative
research has a crucial role to play in improving the
exploitation of data, and provides the setting for
this project.
INTELLIGENT DATA-DRIVEN POLICING
The CPC project was set up to develop a fully-inte-
grated approach to data-driven policing, with par-
ticular emphasis on the spatial and temporal charac-
teristics of crime, policing and citizen reassurance.
Measuring, modelling and predicting the interac-
tions between these amounts to an intelligent and
holistic approach to policing in the digital age.
The CPC project has developed a manifesto for
‘intelligent policing’ which embodies four inter-re-
lated issues that arise in the course of the journey
from data collection to nal policing outcomes.
Each issue raises specic research questions that
have been examined throughout the project.
First, we propose that data-driven tools must be
easy to use and translate straightforwardly into po-
lice action. The outputs of many existing products
are far from intuitive: the large ‘boxes’ identied by
many predictive algorithms, for example, contain
many road sections and it is often unclear exactly
where ofcers should be deployed. To ensure effec-
tive use, tools should be designed with their opera-
tional implementation explicitly in mind.
Our second key principle is that, for tools to be
successful in improving police efciency, accuracy of
prediction is paramount. This requires renement
of analytical techniques for specic policing con-
texts, as well as the selection of appropriate units of
analysis, if police resources are to be directed with
the greatest precision.
Our third point concerns the management of
resources. Many police activities involve the move-
ment and placement of ofcers, either in response
to incidents or in anticipation of them. The volume
and spatial diversity of these demands means that
deciding how these tasks are allocated and struc-
tured is a complex problem, and it is vital to under-
stand how the tasking of ofcers can be coordinated
as efciently as possible.
Finally, we emphasise the role of feedback in the
evaluation and renement of policing strategies. In
order to better understand the effectiveness of new
approaches, it is necessary to know: a) whether as-
signments were properly carried out by ofcers, and
b) whether they had their intended effects. The use
of tracking technology to monitor both compliance
and efcacy can play a vital role in developing an
evidence base.
6|7
OUR APPROACH
The CPC project is fundamentally inter-discipli-
nary, incorporating approaches from geoinformatics,
crime science, geography, computer science and
mathematics. To our knowledge, it is the rst at-
tempt to combine approaches from these elds, and
to address these challenges in a holistic manner.
Many of the methods used in the course of
the project emanate from complexity science. The
research takes an integrated spatio-temporal per-
spective, reecting the strong spatial and temporal
integration of all the three aspects involved: crime,
policing and citizens.
Network science, as a crucial cross-cutting
theme of the project, provides both a framework for
spatio-temporal analysis and a convenient rep-
resentation of urban structure. A key aspect of our
project is that the tools and techniques we have
produced are street network based: we believe this
representation not only allows for greater accuracy,
but is also advantageous from the perspective of
implementation. In addition to this, methods from
Big Data analytics, including machine learning, sta-
tistical analytics and agent-based simulation, have
been used to develop the algorithms and tools.
All of the research has been carried out in close
collaboration with the MPS, which has provided
support in the form of data and guidance concern-
ing current policing challenges. A range of data
sources were used on the project, including incident
logs, tracking data and public surveys. The recent
opening of the JDI Research Laboratory – a £1m
secure data facility – means that these datasets can
be analysed on-site at UCL.
In addition to our scientic ndings, the
strong practical focus of the project has led to the
development of several tools for use in real-world
policing. These include our network-based predic-
tive mapping algorithms and patrol analysis tech-
niques, both of which have been implemented for
real-world use. In conjunction with other tools for
patrol strategy development and public condence
analysis, these form a suite of tools that can be used
to support data-driven policing in an operational
context.
7 priority
crimes
Public
confidence Total cost
20%20%20%
Figure 2: In London, the Mayor’s Police and Crime Plan
sets ambitious objectives for various aspects of police
performance.
Figure 3: The spatio-temporal patterns formed by
crime, policing and citizenship activity form dynamic,
interdependent networks.
SPACE-TIME PREDICTION OF CRIME
BACKGROUND
Years of empirical research have demonstrated
that the occurrence of crime displays a number of
regularities in space and time. Crime occurs dispro-
portionately at certain times and locations, in a way
which can be reconciled with theories of offender
behaviour. The existence of these regularities has
substantial implications for policing, since they im-
ply that crime is to some extent ‘predictable’. If the
regularities in question can be accurately modelled,
this raises the possibility that the spatio-temporal
distribution of crime risk can be forecasted.
MOTIVATION
The idea that crime can be predicted is clearly of
great potential value to policing. In particular, it
provides a basis upon which police can take a proac-
tive approach to resource deployment, intervening
in vulnerable locations before crime occurs. Placing
ofcers at the locations and times at which crime is
most likely focuses policing resources where they
are most needed, resulting in improved efciency,
reduced crime rates and greater public satisfaction.
SPACE-TIME PATTERNING IN
CRIME DATA
At their simplest, the regularities observed for
crime relate to the basic fact that overall levels
or crime vary in space and time: some places are
simply more risky than others. In addition to this,
however, strong interaction effects are frequent-
ly observed. Most prominent among these is the
phenomenon of space-time clustering, which refers
to the tendency of events which are close in time
to also be close in space. This is most distinctively
manifested in ‘near repeat’ victimisation, in which
locations near the site of a recent offence are them-
selves at heightened risk for some time period after
the incident (see Figure 1).
MODELLING CRIME PATTERNS
Two key theoretical concepts have been proposed to
explain the spatial and temporal patterning of crime
incidents. The rst – risk heterogeneity – states that
risk varies because of differences in the underlying
factors which inuence crime: some places are sim-
ply more prone to crime than others. This property
is generally static or slowly changing over time.
The other principle – event dependency – seeks to
explain clustering in particular by stating that the
occurrence of a crime actively increases the proba-
bility of further incidents in the vicinity. This effect
is temporary, but can last for some time (several
weeks in the case of burglary).
AN EARTHQUAKE MODEL APPLIED
TO CRIME
The self-exciting point process (SEPP) model origi-
nated in modelling earthquake aftershocks, but has
recently been applied to describe crime events. The
SEPP model combines the two theories discussed
above into a single framework: crimes occur against
a slowly-varying background risk landscape (risk
heterogeneity) and every crime that occurs has the
potential to trigger further crimes in its vicinity
(event dependence). The background and triggering
proles can be determined using a machine learning
KEY POINTS
Aims To predict where crimes are statistically most
likely to occur over a future time period.
Methods Predictions are based on a model of crime
that combines long- and short-term eects. Risky
locations are mapped to aid police in planning patrol
routes.
Findings Real-world crime data exhibits clustering
in space and time. Application of a predictive model
identies the locations of future crimes with high
accuracy.
0
0.008
0.007
0.006
0.005
0.004
0.003
0.002
0.001
0.000
50 100 150 200 250 300 350 400
Time difference (days)
Event pair density
Figure 1: Relative frequency of pairs of burglary crimes
in Camden that are within 100 metres and the specied
time dierence of one another. The blue region indicates
crimes which occur more frequently than would be
expected in the absence of clustering.
8|9
algorithm, after which they are combined to fore-
cast the appearance of the combined risk landscape
at some future time. Figure 2 shows an example: the
SEPP generates a continuous risk surface, which can
then be transformed onto a square grid, with the
most risky cells highlighted.
APPLYING THE SEPP TO DATA
We applied the SEPP to recorded burglary data
from the London borough of Camden in an 8 month
period commencing in August 2011 (1993 crimes).
The triggering prole is shown in Figure 3 and in-
dicates that a burglary is associated with an elevated
risk within a radius of around 200 metres for a time
period of 60 days. This prole varies by location and
crime type; the SEPP model is tailored to the specif-
ic situation.
IMPLICATIONS
The SEPP model provides a customisable approach
to crime prediction, incorporating two key crimi-
nological theories. It is straightforward to train the
model using historic crime records and the trigger-
ing prole gives insight into the underlying crime
dynamics. The predictive accuracy is higher than
popular alternative methods, suggesting that such
methods offer an effective way to target policing
interventions.
Figure 2: (Top) a continuous prediction heatmap gen-
erated by the SEPP. Darker red indicates higher risk.
(Bottom) the top 10% of grid squares for the same
prediction.
Figure 3: The crime triggering prole in time and space
for burglaries in Camden.
0 10 20 30 40 50 60 70 80 90
Time (days)
Triggering intensity
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
Cumulative prob. of triggering a crime
−600 −400 −200 0 200 400 600
Distance (metres)
Triggering intensity
BACKGROUND
The majority of predictive algorithms developed to
date are either grid-based or produce risk estimates
for regions of a map, such as circles. While it is
certainly possible to predict crime at these levels,
however, there are other representations of space
which may be more appropriate. In particular, using
the street network as the spatial basis for prediction
has a number of potential advantages in terms of
both accuracy and usability.
ADVANTAGES OF NETWORK-BASED
APPROACH
One of the primary benets of using the street net-
work is its spatial granularity. The grid squares typ-
ically used for prediction are relatively large – 250m
x 250m, for example – and contain many roads
and features, making it difcult to identify exactly
where risk lies or where ofcers should patrol. Street
segments, in contrast, are well-dened micro-units,
which allow risk to be estimated with much greater
precision. Furthermore, since ofcers patrol by fol-
lowing dened routes along streets, network-based
risk maps are easier to interpret in practice.
RISK DIFFUSION
There are also reasons to anticipate that the network
will play a role in the spread of crime itself. Much
of the theory on which predictive methods are based
refers to the way in which offenders perceive and
navigate their environment, committing offences
near places that they know or the locations of previ-
ous crimes. Since the network is a key determinant
of urban structure, and of the proximity of places in
particular, it is natural to expect that it will inu-
ence the distribution of crime.
NETWORK-BASED MODEL
To explore the potential of these ideas, we devel-
oped a predictive model of crime which is fully
network-based in terms of both its mechanism and
output. The model builds on previous kernel-based
approaches: a prospective risk ‘surface’ is computed
by summing contributions from nearby previous
crimes, weighted according to how recently they
occurred. The crucial difference in our model is that
the kernels are computed in network space (see Fig-
ure 1), reecting the notion that risk is transmitted
down streets rather than uniformly through space.
KEY POINTS
Aims To develop a street network-based algorithm for
the prediction of urban crime.
Methods A kernel-based approach was adapted to
apply to network space, and used to identify street
segments at high crime risk.
Findings The network-based approach leads to sub-
stantial gains in predictive accuracy, as well as pro-
viding outputs in a form which is convenient for patrol
planning.
PREDICTION ON STREET NETWORKS
Figure 1: Spatial kernel function implemented in net-
work space (represented by black lines). A kernel is
placed at all locations of previous crimes.
10|11
PROSPECTIVE RISK MAPS
We also developed a novel training protocol for
the algorithm, which identies the optimal kernel
parameters for the model with respect to predictive
accuracy. Using these parameters, the model can be
used to compute a prospective risk surface for crime
on any given day, which, in turn, allows the high-
est-risk streets to be identied. An example of a risk
map produced by this method is shown in Figure
2, which also shows the output of the equivalent
grid-based method for comparison. The difference
between the maps is clear to see: a number of high-
risk streets are located outside the highest-risk cells,
and vice versa.
PREDICTIVE ACCURACY
We compared the predictive performance of the
two approaches by examining their ‘hit rates’ when
applied to the same underlying crime data. In com-
puting these rates, coverage is measured in terms of
network length, thereby controlling for variation in
street length across grid cells. As shown in Figure 3,
the network-based method out-performs the grid-
based equivalent by a factor of between 1.5 and 2
for the majority of coverage levels considered.
IMPLICATIONS
Our results indicate that using the street network
as the spatial basis for prospective mapping leads to
signicant improvements in predictive performance
when compared with grid-based approaches. This
fact alone suggests that such algorithms offer an im-
provement on the existing systems used within op-
erational policing. Given that predictions expressed
in this form also offer a number of practical advan-
tages – the maps are easily understood and especial-
ly conducive to patrol planning – there appears to
be a clear case for the adoption of the network-based
approach throughout predictive policing.
Figure 2: Comparison of predictive risk maps, showing
a) network-based, and b) grid-based approaches. In
both cases, the top 5% most risky units are highlighted.
Figure 3: Comparison of predictive ‘hit rates’ achieved
with network- and grid-based approaches.
BACKGROUND
A vital component of prospective mapping is the
process of evaluating the performance of predictive
methods. As well as quantifying expected perfor-
mance, such analysis provides a means of comparing
the capabilities of different predictive approaches.
Until recently, however, crime analysts have had few
tools at their disposal to aid them in assessing and
comparing the many available prediction methods.
The resulting lack of robust selection criteria has
hampered the adoption of such methods in real
operational environments.
CONTEXT OF THIS STUDY
In the majority of work to date, predictive accuracy –
the extent to which the locations of future crimes
are correctly identied – has been used as the sole
measure of a method’s performance. However, this
alone fails to capture a number of other aspects
which are of material importance to the success of
an approach in a real-world setting. The ease with
which predicted locations (hotspots) can be policed,
for example, is a very signicant issue for police
ofcers: if an intervention cannot be applied, the
predicted crimes cannot be prevented. To address
these issues, we have formulated an evaluation
framework which combines a number of measures
to give a comprehensive view of the performance of
a predictive method in an operational context.
EVALUATION FRAMEWORK
In our evaluation framework, four properties of pre-
dictive maps are considered:
Predictive accuracy – the proportion of crime cap-
tured within the predicted locations. We also incor-
porate a signicance test that allows the difference
in accuracy between two methods to be quantied
statistically.
Compactness – how concentrated and connected
the areas identied are. Compactness reects the
intuitive idea that less dispersed areas are easier to
patrol and therefore operationally preferable (see
Figure 1).
Dynamic variability – the extent to which the
predicted locations change between consecutive
predictions. Our metric measures the timescale
of response to changes in crime patterns, thereby
distinguishing between conservative methods that
simply reect long-term risk, and those which are
highly variable from day to day (see Figure 2).
EVALUATION OF PREDICTIVE ALGORITHMS
KEY POINTS
Aims To develop a robust toolkit for the evaluation of
crime prediction methods.
Methods We propose four metrics which capture
various aspects of performance, with emphasis on
real-world implementation and usability.
Findings Application of the framework reveals
strengths and weaknesses in popular predictive
approaches, and can aid the selection of appropriate
methods in various operational settings.
hotspot
areas
hotspot
areas
Figure 1: Illustration of compactness, with high-risk ar-
eas coloured grey. The hotspots in the left-hand gure
are more compact than those on the right.
Figure 2: Dynamic variability for consecutive days:
grey indicates repeated hotspots, blue indicates
emerging hotspots and red indicates disappearing
hotspots.
12|13
Complementarity – the extent to which different
methods detect the same crimes or provide a sup-
plement to one another is measured by the comple-
mentarity index.
PREDICTIVE METHODS
We used our framework to examine the perfor-
mance of four prominent predictive approaches.
We performed a case study using data from London
borough of Camden, examining three crime types,
chosen to reect differing levels of sparseness in the
underlying patterns.
RESULTS
Our results (see Table 1) demonstrate that the vari-
ous methods considered here each display a number
of strengths and weaknesses with respect to their
operational utility. While some methods are capable
of producing very high predictive accuracy, for ex-
ample, their low compactness scores imply that the
locations they identify may not be easy to patrol.
Furthermore, dynamic variability reveals that some
methods produce predictions which remain essen-
tially unchanged for periods of up to one week.
A number of more general observations can
also be made. There appears to be no correlation
between variability and either accuracy or com-
pactness, suggesting that highly dynamic methods
offer little ultimate benet. Furthermore, comple-
mentarity shows that each method is successful at
identifying, exclusively, a substantial number of
crimes outside the ones jointly captured by the oth-
er methods (see Figure 3), particularly for violent
crimes and burglaries. This suggests that an ensem-
ble method, combining multiple predictions, may
be the best solution in these cases.
IMPLICATIONS
The framework we have developed allows predictive
approaches to be evaluated in a signicantly more
comprehensive way than is currently done, with
particular emphasis on their operational utility. Our
case studies reveal that predictive approaches dis-
play a range of strengths and weaknesses when these
aspects are taken into account, implying that the
choice of method must necessarily involve a num-
ber of trade-offs. Our framework allows these to be
assessed quantitatively, thereby equipping police
analysts with the tools required to select approaches
in an informed way based on operational needs.
Shoplifting Violence Burglary
Undetected: 165
Undetected: 155
Undetected: 9
Figure 3: Venn diagram showing the total number of
crimes identied by each method at a xed coverage
of 20% in Camden.
Table 1: Evaluation metrics for Camden crime predic-
tion at 20% coverage level. Bold indicates the greatest
mean value over 100 days’ prediction.
Crime Type Method Accuracy Hotspot
compactness
Variability
Hit rate CI DVI
Mean SD Mean SD Mean SD
Shoplifting PSTSS 81.3 27.6 0.42 0.04 14.9 11
PKDE 74.3 29.8 0.55 0.04 2.7 1.7
SEPP 91.5 20.1 0.31 0.03 6 2.1
PHotspot 85.1 27 0.37 0.04 19.2 9.2
Violence PSTSS 46.5 20 0.46 0.04 10.8 8
PKDE 51.7 19.5 0.54 0.04 2.6 3.9
SEPP 59.7 19.8 0.12 0.03 4.5 1.6
PHotspot 52.2 19.9 0.32 0.05 21.1 6.7
Burglary PSTSS 34.4 22 0.51 0.05 3.7 5.6
PKDE 38.8 24.2 0.5 0.07 2.3 1.9
SEPP 47.4 26.3 0.02 0.05 1.4 1.2
PHotspot 34.9 23 0.3 0.06 5.3 4.4
BACKGROUND
Much crime prevention activity depends, at some
level, on the ability of police ofcers to deter crime
by being physically present in a place. This idea
forms the basis for a signicant proportion of police
activity, with visible patrol – one of the few tactics
available to police that can be used to address crime
problems on a day-to-day basis – a primary example
of this. In particular, patrol plays a crucial role in
‘predictive policing’ approaches, since it typically
constitutes the primary tactical response: ofcers are
sent to the predicted locations in order to discour-
age or interrupt the anticipated crimes.
MOTIVATION
Despite its importance, however, the question of
whether patrol does indeed deter crime has not been
addressed in quantitative terms: put simply, there is
a lack of evidence concerning whether patrol real-
ly ‘works’. The reason for this is largely technical:
traditionally, there has been no means of systemat-
ically recording the locations of ofcers, meaning
that it has been impossible to test whether their
presence affects crime. The recent proliferation of
GPS technology, however, has removed this barrier
by allowing ofcer movements to be logged in their
entirety. In this part of the project, we used track-
ing data of this type from the MPS to perform the
rst large-scale quantitative study of the deterrent
effect of policing.
TEMPORAL DEPENDENCE
In our initial analysis, we explored the temporal
relationship between patrol visits and calls for
service. Using a technique previously applied to
study repeat victimisation, we examined whether
the two events occurred closer together – or further
apart – than would be expected if they were unre-
lated. Figure 1 shows the intervals between the two
event times for street segments in Camden, with
the black line showing the expected distribution if
the events were independent. As can be seen, short
positive intervals (which correspond to cases where
patrol visits are followed by incidents) occur less
frequently than would be expected, which implies
that police presence does discourage crime.
SURVIVAL ANALYSIS
To test these effects in a more sophisticated way,
we also used an epidemiological approach known
as survival analysis. This involves examining the
‘survival times’ associated with each patrol visit:
the time elapsed until the next crime occurs on the
street segment visited (shown for Camden in Figure
2). By comparing these to what would be expected
if no patrol had occurred, the effectiveness of patrol
can be quantied: do streets survive for longer than
they would otherwise have done?
QUANTIFYING DETERRENCE
KEY POINTS
Aims To measure the extent to which routine visible
policing deters crime.
Methods Ocer GPS traces were used to reconstruct
police locations and movements. These were then
tested for temporal association with incident locations.
Findings Police presence does appear to deter crime,
but the eect is modest and relatively short-lived.
Figure 1: Time elapsed between patrol visits and calls
for service across all 3-hour windows containing ex-
actly one event of each type.
14|15
BASELINE ESTIMATES
Estimating expected survival times is a technical
challenge, however, since the underlying level of
risk varies across segments and times. To address
this, we developed a set of Monte Carlo-based ap-
proaches by which the typical time-to-next-crime
can be estimated for any street segment, at any
time. These times can be compared statistically to
the observed intervals between police visits and
calls.
SURVIVAL RESULTS
In Figure 3, we show the results of a regression
model for the ‘hazard rate’: the instantaneous risk
of crime as time elapses after a patrol event. The
plotted line shows the deterrent effect: its distance
below the zero line shows the decrease in risk, com-
pared with what would be expected in the absence
of patrol. These results imply that patrol does in-
deed have a sustained deterrent effect, but that it is
small and only marginally statistically signicant.
IMPLICATIONS
This research contributes to the evidence base con-
cerning the fundamental policing tactic of visible
patrol. The nding that patrol does act as a deter-
rent is encouraging from the perspective of everyday
policing, but this should be tempered somewhat by
the relatively modest size of the effect. This implies
that the importance often ascribed to patrol may be
overstated, which has signicant potential conse-
quences in terms of how routine policing should be
done. It should be stressed, however, that additional
research is required to address more subtle ques-
tions: how frequently patrol should be applied, and
whether it acts differentially across crime types and
areas. Ultimately, though, research of this type can
be used to inform the design of improved patrol
strategies, via which the police can increase the
efciency with which they prevent crime.
Figure 2: Survival function for vehicle patrol events in
the London borough of Camden.
Figure 3: Survival regression model, showing the
hazard rate as a function of time elapsed since visit -
the ‘vehicle presence’ variable shows the dierence
between observed values and those in the absence
of patrol.
BACKGROUND
The ability to accurately replicate how police
navigate around the city is of crucial importance in
answering what-if questions, such as:
– How does changing the number of vehicles on
patrol affect street coverage?
– How do locations of police stations affect the
time needed to reach crime scenes?
It also gives insights into any regularities in police
behaviour that might impact on their preventive
capabilities. Standard route planning algorithms
fail at this task as they cannot replicate sub-opti-
mal route choices, such as those that are frequently
required in the course of policing.
ARE POLICE ROUTES OPTIMAL?
Standard route planning algorithms are based on
the assumption that drivers are 100% rational in
their behaviour. When given a journey destination,
they would nd possible routes to reach the destina-
tion and always pick the one that incurs the small-
est cost (in terms of distance, time, etc.). However,
as is the case of most drivers, police choose routes
that cannot be fully explained within the rationality
framework. As shown in Figure 1, they tend to fol-
low paths that are slightly longer than the shortest
possible alternatives. This might be explained by
their limited spatial knowledge, trafc, or monitor-
ing activities that purposely avoid major roads.
OUR APPROACH
We proposed an algorithm that is capable of learn-
ing police route choice directly from police-gener-
ated data, hence enabling more realistic simulations
of police movements. We based our route choice
model on the assumption that police plan their
journeys in stages, taking preferred routes through
neighbourhoods en route to destination. This as-
sumption is in line with well-established research
into vehicle route choice, which suggests that driv-
ers’ behaviour is rather suboptimal and that their
route selection takes place in phases, linking loca-
tions and decision points on route to destination.
We extracted preferred routes from police GPS data
and then simulated police journeys as sequences of
the preferred routes.
POLICE ROUTES
Police routing preferences were inferred from large
amounts of police-generated GPS logs using a topic
modelling technique called Latent Dirichlet Allo-
cation. The technique was originally developed to
uncover topics from large collections of documents.
In our setting, it was employed to uncover routes
often traversed by police patrol vehicles. By varying
the desired number of routes, we could investigate
police routing preferences at various scales as shown
in Figure 2.
POLICE ROUTE CHOICE MODELLING
KEY POINTS
Aims To create a realistic model of police route choice.
Methods Route choices were extracted from police
GPS tracks using topic modelling techniques; police
journeys were simulated as sequences of the extract-
ed routes.
Findings Routes predicted using the proposed model
were signicantly closer to actual routes taken by the
police than those predicted by o-the-shelf route plan-
ning algorithms.
Figure 1: Lengths of police journeys in March 2010 and
their shortest alternatives.
Count
Path length (metres)
020000 3000010000
500
1000
1500
2000
2500
3000
3500
4000
40000
actual
Journey
least distance
16|17
SIMULATING POLICE JOURNEYS
We simulated police journeys as sequences of their
preferred routes. This involved representing the
street network in an aggregated form where streets
belonging to the same route were collapsed into a
single node. Links between nodes were drawn if the
nodes contained adjacent streets. An exemplary such
network is shown in Figure 3a. Journeys predicted
using our model, as the one shown in Figure 3b,
incorporated data-driven routing preferences of the
police, which lead to signicant improvements in
accuracy when compared to off-the-shelf routing
algorithms.
IMPLICATIONS
Our proposed model shows signicant improve-
ments over existing methods in replicating the
routing behaviour of police ofcers. This provides
justication for its potential use as a planning tool
for police resourcing. By simulating journeys under
different assumptions, the real-world implications
of potential policy changes can be explored in ad-
vance, allowing police decision-makers to plan in an
evidence-based way.
b) 100 topics
a) 10 topics
Figure 2: Streets in Camden colour-coded by routes
they belong to; routes being inferred by Latent
Dirichlet Allocation by tting (a) 10, (b) 100 topics to
police GPS data.
Figure 3: (a) A police journey simulated using the ag-
gregated street network versus (b) the actual journey
taken.
Simulated
Actual
Late Shift
Early Shift
Night Shift
29
27
26
25
24
23
22
21
20
19
18
17
16
15
13
12
11
10
9
8
7
6
5
4
3
2
1
0
BACKGROUND
The things people do in space and time have long
been a research topic in behavioural and socio-eco-
nomic studies, particularly in relation to highly
dynamic urban environments. In this eld, the term
‘activity pattern’ is used to describe the movements
which groups of people take in the course of their
routine daily activities. These activities are intrin-
sically linked to the places and times at which they
are undertaken, to the extent that it can be said that
‘where, when and how long you stay is who you are’.
ACTIVITY PATTERNS IN POLICING
The study of activity patterns has particular value
for policing since, while coverage is clearly a crucial
operational issue, relatively little is known about
how ofcers make decisions during self-directed
activity. In the context of policing, and foot patrol
in particular, such ndings have potential opera-
tional relevance. Identifying places and time periods
with distinctive patrol patterns, as well as grouping
ofcers who share similar patterns, is of use for both
performance evaluation and for identifying aspects
of behaviour which could be improved.
BEHAVIOURAL CLASSIFICATION
In our research, we have proposed a methodologi-
cal framework for uncovering space-time activity
patterns from individuals’ movement trajectory
data, which can then be used to segregate users into
subgroups according to the similarity between their
patterns. This is designed to be applied to new-
ly-available datasets which collect the ever-changing
position of moving individuals with high spatial
and temporal resolution. The Automatic Personnel
Location System (APLS) used by the MPS to record
ofcer movements using GPS is one such source,
and provided the basis for our work.
SPATIO-TEMPORAL REGIONS
OF INTEREST
In the proposed framework, APLS traces are rst
map-matched onto the street network to identify
the street segments that the ofcers truly visited.
The places and times at which particularly high lev-
els of ofcer presence were observed are then identi-
ed using a network-based variant of the clustering
algorithm DBSCAN. These regions of space-time
are dened as Spatio-Temporal Regions of Interest
(ST-ROIs), and represent locations which feature
particularly prominently in patrol behaviour
(Figure 1).
CHARACTERISING FOOT PATROL BEHAVIOUR
KEY POINTS
Aims To characterise and classify the movement pat-
terns of foot patrol ocers.
Methods Using ocer GPS traces, spatio-temporal
regions of high patrol intensity were identied. Ocer
behaviours were then classied according to their
visits to these locations.
Findings Ocer activity clusters around certain prom-
inent locations, and distinctive behavioural signatures
can be identied in their visits to these.
Figure 1: During the study period, 28 ST-ROIs are
detected for foot patrol ocers by space-time
DBSCAN – these are identied by colour here.
18|19
GROUPING OFFICERS
Once ST-ROIs have been identied, individual
ofcers’ activity patterns can be expressed in terms
of their visits to these locations. An individual’s
behaviour is dened as his/her prole of time alloca-
tion to the ST-ROIs s/he visited, and can be con-
sidered to be a ‘signature’ of patrol activity (Figure
2). By applying a hierarchical clustering approach
to these proles, ofcers can be partitioned into sub-
groups based on the similarity of their activities in
space and time (see Figure 3).
SUMMARY OF THE ACTIVITIES
When the ofcer subgroups are examined in turn,
a qualitative understanding of their characteristic
behaviours can be gained. One group of ofcers,
for example, can clearly be seen to be involved in in-
tensive patrol activity overnight in a location which
is known to be a centre of the night-time economy.
Another group clearly corresponds to a special-
ist operation at a location outside Camden, when
ofcers are seen to deploy for long stints of patrol
after beginning at a Camden station.
IMPLICATIONS
The application of our framework has the potential
to offer unprecedented insight into the micro-level
behaviour of police ofcers during routine patrol.
The framework extends traditional ideas of time
budget allocation, and is capable of proling the
activity patterns of ofcers in both space and time
by dening a new moving behavioural similarity
metric. Since the clustering method explains the se-
mantic meaning of different behaviours – activities
are mapped in terms of visits to particular land-
marks – the ndings are readily interpretable by
commanding ofcers. Findings such as these could
be used to suggest behavioural modications which
could lead to more efcient coverage.
1903 PO
1911 PO
1924 PO
1811 PO
1812 PO
1939 PO
1949 PO
1957 PO
1963 PO
1970 PO
1971 PO
1972 PO
1973 PO
1978 PO
1979 PO
1987 PO
1993 PO
1996 PO
1818 PO
1819 PO
2013 PO
2014 PO
1821 PO
2034 PO
2040 PO
2043 PO
2046 PO
2058 PO
1825 PO
2094 PO
2096 PO
1828 PO
2099 PO
2114 PO
2134 PO
2145 PO
2147 PO
2171 PO
2177 PO
5811 PO
5812 PO
2200 PO
2207 PO
2208 PO
2209 PO
2218 PO
2244 PO
2258 PO
2267 PO
2268 PO
2276 PO
2281 PO
2286 PO
2287 PO
2288 PO
1847 PO
2295 PO
6806 PO
2300 PO
2302 PO
2308 PO
6906 PO
2311 PO
2328 PO
7132 PO
2341 PO
2347 PO
2352 PO
2358 PO
2366 PO
2367 PO
2369 PO
2398 PO
2410 PO
2415 PO
2419 PO
2425 PO
2426 PO
2428 PO
1861 PO
2432 PO
2435 PO
2456 PO
2457 PO
8470 PO
8473 PO
8474 PO
8477 PO
8478 PO
8486 PO
8487 PO
8489 PO
2480 PO
2500 PO
2501 PO
2505 PO
8907 PO
8912 PO
8913 PO
8914 PO
8919 PO
8920 PO
8937 PO
2512 PO
8940 PO
8967 PO
2515 PO
8972 PO
8973 PO
8974 PO
8986 PO
2517 PO
8989 PO
8991 PO
9015 PO
2520 PO
9020 PO
2522 PO
2528 PO
2547 PO
1873 PO
2559 PO
2564 PO
1806 PO
2606 PO
2609 PO
1882 PO
1894 PO
1897 PO
1857 PO
1872 PO
Figure 2: Simplied representation of two example
ocers’ movements, showing (a) trajectories in space-
time; (b) simplied representation showing the visiting
sequence of ST-ROIs.
Figure 3: Taxonomy tree showing the clustering of
ocers with dierent patrol patterns (ID numbers
randomised).
BACKGROUND
As part of the development of policing, it is natural
for senior ofcers to consider the potential of chang-
es in operational practice to improve the efciency
of policing. These changes might be technological,
such as the adoption of handheld devices, or behav-
ioural, such as the renement of tasking protocols.
Variations in resource levels can impact the ability
of ofcers to provide service, while terrorist attacks
or special events drastically affect the set of respon-
sibilities ofcers must handle. These cases represent
material changes to the way that policing is done.
A crucial task in assessing the potential value of
such changes is estimating their likely impact on
performance, so that decisions can be taken in an
informed way. The complexity of policing practice,
however, means that this is a very difcult task: the
potential for knock-on effects and feedback loops
means that such questions cannot be answered
simply.
AGENT-BASED MODELLING
One analytical technique which has substantial
potential value in such situations is agent-based
modelling. Agent-based modelling is a computer
simulation method in which articial populations
are endowed with realistic behavioural rules and
situated within an environment, within which
they interact with one another. In simple terms,
it involves the simulation of articial ‘worlds’, the
development of which can be examined with unlim-
ited granularity. By changing the behavioural rules
or parameters, the potential impact of these changes
can be explored: the method provides a means of ex-
amining ‘what if’ questions in a very detailed way.
In the context of policing, such a simulation can be
used to explore the impact of potential operational
changes on performance indicators such as response
time and coverage.
STRATEGIC PLANNING FOR OFFICER TASKING
KEY POINTS
Aims To provide a tool to explore the impact of chang-
es to police operations.
Methods An agent-based simulation model for ocer
behaviour and tasking was developed, based on in-
sight gained for active ocers.
Findings The model replicates real-world ocer be-
haviour in terms of the spatial distribution of coverage,
suggesting that it has potential use as a real-world
decision support tool.
Despatched
to incident?
Yes
No
Patrol
Go to incident site
(ignore traffic laws) Deal with incident
Need
Transport
Vehicle?
Need
Transport
Vehicle?
Wait until Transport
Vehicle arrives
Return to station to
make a report
No
No
Yes
Yes
Figure 1: The decision ow for a simulated ocer in
‘Responding’ status.
20|21
SIMULATING TASKING
In order to explore these questions, we developed an
agent-based model for the behaviour and tasking of
police ofcers. The construction of the model was
informed by consultation with police ofcers, and
incorporated several aspects: the dispatch process,
the travel of ofcers to incidents, and self-directed
patrol behaviour. Simulated ofcers act according
to decision trees (see Figure 1) and the aim of the
model is provide a comprehensive simulation of the
day-to-day behaviour of a police force.
CAMDEN CASE STUDY
In order to establish that the model produced
realistic behaviour, and to examine its potential as
a policy tool, we applied the model in the context
of the London borough of Camden. We used re-
al-world data on calls for service, so that simulated
ofcers were required to respond to calls exactly as
they were in the real world. To establish validity, we
examined whether the movement patterns gener-
ated were similar to those of real Camden ofcers
in the period in question. The distribution of road
usage – a distinctive feature of real-world patrol –
was broadly in agreement (see Figures 2 and 3), and
a signicant improvement over corresponding null
models.
IMPLICATIONS
Modelling tools based on agent-based simulation
have considerable potential as policy tools because
they provide a means to explore extremely complex
systems in an objective and transparent way. Our re-
sults suggest that our model is a useful representa-
tion of real world policing, and that it therefore
represents a valid basis for such a tool. As well as
rening the model in order to achieve closer corre-
spondence with reality, we will use it to explore a
number of real-world policy questions relating to
potential changes in policing practice.
Figure 2: The distribution of activity for real ocers in
Camden, derived from GPS traces.
Figure 3: The spatial distribution of simulated ocer
behaviour.
BACKGROUND
Police patrol occupies a central place in crime
control efforts. In particular, hotspot patrolling – in
which effort is focussed on small geographical units
with high crime intensity – has gained particular
prominence as a means of deterring potential crime
and increasing public perception. The issue of how
such patrols can be most efciently operationalised,
however, presents a signicant logistical challenge,
especially when police resources are limited and
hotspots are many.
MOTIVATION
In most operational contexts, hotspot patrolling is
implemented by tasking ofcers to randomly rotate
between hotspots. However, randomised strategies
have a number of shortcomings, as they omit the
peculiarities and challenges of daily police patrol,
such as the desire to minimise the average time lag
between two consecutive visits to hotspots, while
also coordinating multiple patrollers and imparting
unpredictability to patrol routes. To address these
issues, and to ensure that patrol is performed as
efciently as possible, it is necessary to develop a
cooperative routing strategy.
MEASURES
Before any new strategy can be developed, it is
necessary to express the objectives of police patrol
in quantitative terms. However, there is a lack of
systematic investigation of the relevant measures.
In the rst stage of this research, we explored the
question of what makes a good police patrol routing
strategy, and proposed a number of guidelines and
measures, shown in Table 1.
ONLINE BAYESIAN ANT-BASED
PATROLLING STRATEGY
We developed a routing strategy, Bayesian Ant-
based Patrolling Strategy (BAPS), according to the
proposed guidelines (Figure 1). The main idea of
BAPS is to use ‘pheromones’ to record visit history
on hotspots, which can then be used to prioritise de-
mand. The algorithm uses Bayesian decision-mak-
ing to incorporate multiple factors, and performs
one-step routing each time.
ROUTING STRATEGIES FOR FOOT PATROL
KEY POINTS
Aims To propose guidelines and measures for the
performance of police patrol and to design a routing
strategy for eective foot patrol.
Methods A 'pheromone'-based algorithm was com-
bined with Bayesian decision-making to produce a
cooperative routing strategy.
Findings The performance of police patrol can be
quantied using 5 objective measures, and the novel
routing strategy developed oers substantial improve-
ments in eciency.
Figure 1. From guidelines to BAPS.
Online Bayesian Ant-based Patrolling Strategy
Table 1. Guidelines for police patrol routing strategy
Measure Meaning
Eciency (E) To cover hotspots regularly and fairly
Flexibility (F) To cover hotspots that have varying
degrees of weights
Scalability (S) To be applicable to patrolling in dier-
ent areas (spatial scalability) and with
dierent numbers of patrollers (team
scalability)
Unpredictability (U) To keep the patrol routes unpredictable
Robustness (R) To remain eective patrolling when
some patrollers are dispatched for
emergencies
Receive Tasks
One-step
routing
Await Tasks
Guidelines
Using pheromones
to record visit history
Bayesian decision-making
(including multiple factors)
BAPS
Control
Centre
Emergency Patrol
Report Task
Completed
Efficiency Flexibility Unpredictability Scalability Robustness
Patrol
Respond to
Emergency
22|23
CASE STUDY
We examined the performance of BAPS using a be-
spoke multi-agent modelling framework (see Figure
2), with real-world crime data from the London
borough of Camden used as a case study. Crime
hotspots were dened as the 5% of segments with
the highest crime density level (see Figure 3), and
were divided into 5 levels. An existing determinis-
tic patrolling strategy known as CCPS was used as a
benchmark.
RESULTS
The performance of BAPS is compared against that
of CCPS in terms of the 5 key metrics (See Table
2). In the experiments, 30 patrollers started from
6 stations. In the normal scenario, they only pa-
trolled, whereas in the emergency scenario, they also
responded to emergency calls when needed. The
results show that BAPS outperforms CCPS across all
5 performance measures.
IMPLICATIONS
This research represents a signicant step in the
development of intelligent patrolling, which has the
potential to substantially improve the operational
efciency of targeted interventions. The perfor-
mance measures introduced allow patrol to be eval-
uated in objective and quantitative terms, meaning
that behaviour can be rened in an evidence-based
way. By adopting a probabilistic framework, the co-
operative BAPS strategy introduced achieves signf-
icant performance improvement, suggesting that it
has the potential to form the basis for an effective
real-world protocol. Future work will consider how
this can be implemented in a practical context, and
integrated alongside vehicle patrol.
Figure 2. Schematic diagram of BAPS.
Figure 3. Camden hotspot map.
E F U S R
BAPS 1700 1623 2055 1.037 3.30%
CCPS 2039 2030 497 0.996 4.10%
Change (%) -16.60 -20.05 -- -- --
E- Eciency; F – Flexibility; U – Unpredictability;
S: (Team) Scalability; R - Robustness
Table 2. Comparison of CCPS and BAPS performance
Receive Tasks
One-step
routing
Await Tasks
Guidelines
Using pheromones
to record visit history
Bayesian decision-making
(including multiple factors)
BAPS
Control
Centre
Emergency Patrol
Report Task
Completed
Efficiency Flexibility Unpredictability Scalability Robustness
Patrol
Respond to
Emergency
BACKGROUND
A growing number of studies report lower levels
of public trust and condence in governments and
their institutions and public services, which inu-
ence public engagement, co-operation and partici-
pation which all exist at the heart of democratic in-
stitutions. Public trust in policing is no exception;
evidence shows that condence is shrinking. In the
UK, where the ideological emphasis is on policing
by consent, this is particularly problematicand thus
‘promoting public condence’ in policing has been a
top priority of the government agenda.
MOTIVATION
An increasing number of studies and surveys at-
tempt to investigate and measure trust and con-
dence in policing.Not only is there no conceptual
clarity between the two terms and how these are
used in the policing context, but there is also a
complete lack of a theoretical and empirical frame-
work on which investigations are based.This raises
signicant concerns about the validity of the results
and subsequently the effectiveness of measures and
policies which attempt to improve public trust and
condence in policing.Our research attempts to
address this gap.
THE ROLE OF THE TRUSTEE ATTRIBUTES
Trust is a complex phenomenon, which is highly
context-dependent and incorporates a number of
factors. In many models, including that used by our
team – see Figure 1 – an individual’s willingness
to trust is dependent on both their own personal
attributes and those of the trustee. The inuence of
these factors is of particular relevance in the con-
text of policing. It is natural to assume that trust
perceptions are not universal, and are inuenced by
factors such as cultural or educational background,
prior experiences, misconceptions and the lack of a
shared public understanding of police practices.
TRUST AND CONFIDENCE IN POLICING
KEY POINTS
Aims To provide a theoretical and empirical framework
for investigating and improving public trust in policing.
Methods A mental models approach was applied in
conjunction with interviews conducted in London.
Findings The approach oers insight into the factors
that inuence trust, and reveals a number of miscon-
ceptions amongst both experts and the public. These
can be used to inform strategies to improve trust.
Figure 1: Trust Model
24|25
THE MENTAL MODELS APPROACH
The mental models approach provides a framework
for uncovering the decision processes of individu-
als. It is based on the idea that human memory is
organised into ‘schemas’, which describe how an
individual perceives a problem, depending strong-
ly on previous personal experiences, knowledge,
activating stimuli, and other factors. Depending on
their complexity, mental models may consist of one
or more schemas, which inuence decisions and the
way that individuals perceive the world. The mental
models approach provides a systematic way to un-
cover these, the concepts that are included and the
vocabulary people use to describe them. In addition,
it can also reveal misconceptions and knowledge
gaps that should be corrected.
LONDON STUDY
We conducted 25 interviews with both experts
and members of the public in London (with almost
half of the subjects from the borough of Barking
& Dagenham and the rest from wider London).
Analysis of our preliminary results shows – perhaps
surprisingly – that there is no common and shared
understanding of trust and condence in policing
amongst experts. In other words, the expert mental
models of experts in this topic vary signicantly, as
do the issues that they perceive to matter most to
the public. Furthermore, the trustee attributes that
are used in various surveys to measure trust (e.g.
perceived legitimacy, altruism) are not necessarily
present in the expert mental models captured in our
study. In addition to this, the public mental models
also reveal signicant gaps and misconceptions, as
well as a limited awareness of engagement activities
and initiatives.
IMPLICATIONS
Our analysis suggests that the issue of public trust
continues to pose a signicant challenge to police,
in terms of both their understanding of it and the
success of engagement activities. Nevertheless,
the mental model approach provides a number of
insights into the issues that the public expect to be
addressed in order to improve trust. Further anal-
ysis will help to systematise the trustee attributes
that are important to the public and generate a list
of guidelines for improving trust. Furthermore,
the framework can be used to inform the design of
questions targeting key attributes, which can then
be incorporated into a survey that will more effec-
tively measure public trust.
BACKGROUND
Puc condence in the police is a state in which the
public regard the police as competent and capable
of fullling their roles. This results from the police
being effective in dealing with crime and anti-social
behaviour, as well as fair treatment of and engage-
ment with the community. The British model of
policing is underpinned by a philosophy of “polic-
ing by consent” whereby the police are empowered
by the common consent of the public. The public
observe the law as a result of their approval, respect
and affection for the police rather than compliance
being motivated by fear. In this context, public con-
dence in the police is a key component of effective
policing. Persons who are condent in the police
are more likely to be cooperative, compliant and
crucially to supply the tips which inform proactive
policing operations.
MOTIVATION
The Metropolitan Police Public Attitudes Survey
(PAS) collects data on the experiences and percep-
tions of Londoners with respect to crime, policing
and anti-social behaviour. While the most robust
survey of its kind in the world, the PAS is not
designed for use at the neighbourhood level. Im-
provements are required to support the neighbour-
hood level policing initiatives which are central to
improving public condence. Public condence in
the police varies across geographical space and over
time (see Figure 1). Understanding these patterns at
the local level is an important step in developing a
targeted condence intervention strategy.
SMALL AREA ESTIMATION OF PUBLIC CONFIDENCE
KEY POINTS
Aims To estimate and predict public condence in the
police at the neighbourhood level.
Methods A spatio-temporal interaction Bayesian
hierarchical model was developed to model public
condence trends in space and over time.
Findings Public condence in policing is autocorrelat-
ed in space-time. This autocorrelation can be used to
estimate and predict public condence at the neigh-
bourhood level.
010205Kilometers
Legend
River Thames
Confidence above MPS average
Confidence below MPS average
Quarter 29 Quarter 30
Quarter 31 Quarter 32
Figure 1: Group of maps of measured public con-
dence levels in London for the period April 2012-
March 2013. Dark blue areas indicate condence
levels above the London-wide average.
26|27
SMALL AREA ESTIMATION
A spatiotemporal Bayesian hierarchical modelling
approach enabled the estimation and prediction
of public condence in the police at the small area
level. This approach allows trends to be explored in
space-time and neighbourhood level intelligence to
be obtained from sparse sample survey data.
RESULTS
Public condence was found to exhibit spatiotem-
poral dependence. In Figure 2, we show a space-
time variogram, which illustrates the strength of
the dependence in space and time. Neighbourhoods
up to 1 kilometre away in space and two quarters
away in time are correlated. A Bayesian hierarchical
model was then developed which allowed public
condence levels to vary by neighbourhood and
temporal quarter. It also includes spatial, temporal
and spatiotemporal components, and was used to
estimate and predict public condence levels at the
neighbourhood level (see Figure 3).
IMPLICATIONS
This research contributes evidence of the autocorre-
lation of public condence in space and time. This
autocorrelation was leveraged in a spatiotemporal
model to produce estimates and predictions at a lev-
el which can be useful in enabling ofcers to design
targeted strategies to improve public condence.
The estimates and predictions obtained can better
enable ofcers to prepare for the proactive pub-
lic condence interventions required to meet the
concerns of the local community. This approach also
allows hot spots and cold spots of public condence
to be better identied and examined.
3D Spatiotemporal variogram of confidence in the police
gamma
temporal separation
(quarter)
spatial separation
(metres)
010205Kilometers
Legend
River Thames
Confidence above MPS average
Confidence below MPS average
Quarter 29 Quarter 30
Quarter 31 Quarter 32
Figure 2: 3D representation of a space-time variogram
which describes the space-time structure of public
condence
Figure 3: Group of maps of estimated and predicted
public condence levels in London for the period April
2012- March 2013. Dark blue area indicate condence
levels above the London-wide average.
BACKGROUND
It has been acknowledged for some time now that
the relationship between the police and the rest of
society is an important factor in the prevention of
crime. Evidence suggests that the public’s con-
dence in the police affects, among other things, the
willingness of citizens to obey the law and cooperate
with ofcers, and is therefore a crucial issue for po-
lice legitimacy. In addition to this, recent years have
seen an increasing recognition that perception of the
police is an important aspect of public wellbeing,
with the result that increasing condence is seen as
an end in itself. Because of this, many forces have
set explicit targets for the improvement of public
condence; the MPS, for example, has been required
to achieve a 20% improvement between 2013 and
2016.
IMPROVING PUBLIC CONFIDENCE
In seeking to improve public condence, it is natu-
ral that the police should seek to tailor their efforts
to particular sections of the population, by focus-
ing on groups with particularly low condence, for
example, or by addressing specic concerns amongst
others. This is extremely challenging, however: the
relationships between condence and socio-demo-
graphic indicators are subject to complex interac-
tions, making generalisation difcult. Furthermore,
aggregate-level relationships are of limited use in
informing strategies for condence improvement:
it is far from clear how the concerns of a particular
age group can be addressed at a city-wide level, for
example.
GEODEMOGRAPHICS
One way in which the inuence of multiple so-
cio-demographic factors can be brought together
is through the use of geodemographics. Geodemo-
graphics is “the analysis of people by where they
live” and is used to identify socially similar groups
on the basis that similar people are more likely to
live within the same locality, have similar lifestyles,
and share similar views. Using this approach, pop-
ulations can be grouped according to their various
characteristics, providing an overview of any par-
ticular neighbourhood which can be used to make
inferences about the individuals within it. This is
of clear value for policing, since it provides a means
to assess how the combination of measured factors
inuences public attitudes. If areas can be proled
in these terms, then efforts to improve public con-
dence can be tailored spatially, according to the
needs of particular neighbourhoods.
GEODEMOGRAPHICS AND PUBLIC CONFIDENCE
KEY POINTS
Aims To explore and simplify the relationship between
demographic factors and condence.
Methods Survey response data is explored using a ge-
odemographic classication, which identies socially
similar groups.
Findings Geodemographic supergroups display
distinct trends in their attitudes towards police. These
classications can be used to tailor eorts to improve
condence to the needs of particular areas.
28|29
LONDON OUTPUT AREA
CLASSIFICATION
A number of geodemographic classications are
available in the UK, including one which is spe-
cic to London and was developed by our research
team: the 2011 London Output Area Classication
(LOAC; see Figure 1). Using this as a basis, we
examined spatial and temporal trends in the public
perception of the MPS, as measured by the Public
Attitude Survey (PAS). This is a monthly cross-sec-
tional survey designed to elicit the public’s percep-
tions of policing needs, priorities and experiences,
with a number of questions specically focussed on
issues of public condence.
GROUP-LEVEL TRENDS
Our results show that substantial differences can be
observed between LOAC supergroups in terms of
their attitudes towards the police. In Figures 2 and
3, we show the trends over time in the responses
to 5 key questions for 2 distinct geodemographic
groups: ‘urban elites’ and ‘multi-ethnic suburbs’.
This is signicant because it implies that these
groups do indeed display idiosyncratic character-
istics with respect to their perception of police.
Knowledge of an area’s classication therefore
allows inferences to be made about the level of con-
dence of people living there.
IMPLICATIONS FOR POLICING
The use of geodemographics has the potential to
signicantly improve the precision with which
police efforts to improve public condence can be
focussed. The fact that spatiotemporal variation can
be explained in these terms suggests that classi-
cation systems such as the LOAC provide a simple
and convenient basis on which to target initiatives.
Since each classication corresponds to a stylised
socio-demographic ‘portrait’ of the kind of person
who lives there, interventions can additionally be
designed in ways which are tailored to the needs of
particular groups.
Figure 3: Trends in PAS responses for the LOAC
supergroup ‘multi-ethnic suburbs’.
Figure 2: Trends in PAS responses for the LOAC
supergroup ‘urban elites’.
Figure 1: The 2011 London Output Area Classication,
with areas coloured according to geodemographic
supergroups.
Produced by Chris Gale
@geogale
Contains National Statistics data
© Crown copyright and database right 2016
Contains Ordnance Survey data
© Crown copyright and database right 2016
0 5 10
Kilometres
A - Intermediate
Lifestyles
B - High Density and High
Rise Flats
C - Settled Asians
D - Urban Elites
E - City Vibe
F - London Life-Cycle
G - Multi-Ethnic Suburbs
H - Ageing City Fringe
0
10
20
30
40
50
60
70
80
90
100
April 2006
July 2006
October 2006
January 2007
April 2007
July 2007
October 2007
January 2008
April 2008
July 2008
October 2008
January 2009
April 2009
July 2009
October 2009
January 2010
April 2010
July 2010
October 2010
January 2011
April 2011
August 2011
October 2011
December 2011
April 2012
July 2012
October 2012
January 2013
April 2013
July 2013
PERCENTAGE OF POSITIVE RESPONSES
MONTH
Effectiveness in crime prevention and protection Community Commitment/ Engagement
Fair Treatment Alleviating Local Anti-Social Behaviour
Feeling informed about local policing
0
10
20
30
40
50
60
70
80
90
100
April 2006
July 2006
October 2006
January 2007
April 2007
July 2007
October 2007
January 2008
April 2008
July 2008
October 2008
January 2009
April 2009
July 2009
October 2009
January 2010
April 2010
July 2010
October 2010
January 2011
April 2011
August 2011
October 2011
December 2011
April 2012
July 2012
October 2012
January 2013
April 2013
July 2013
PERCENTAGE OF POSITIVE RESPONSES
MONTH
Effectiveness in crime prevention and protection Community Commitment/ Engagement
Fair Treatment Alleviating Local Anti-Social Behaviour
Feeling informed about local policing
The CPC project has at its core a strong practical
focus, and an essential component of the project is
therefore the translation of research into real-world
tools. We have developed a number of our research
outputs into stand-alone tools suitable for use by
practitioners. These perform a number of tasks,
including both predictive policing and performance
evaluation functions. They are designed with police
end-users in mind, and act as prototypes for the
eventual large-scale deployment of these systems in
real-world police forces. They are integrated to-
gether in a suite of tools, which provides a unied
system for data-driven policing support.
PREDICTIVE MAPPING
On the basis of our research into crime prediction,
we have developed a predictive tool which can be
used to prospectively identify the locations of future
crimes in a real-world setting. The tool imple-
ments our network-based predictive algorithm, as
described in Prediction on street networks, which we
found to offer signicant improvements in predic-
tive performance when compared with other algo-
rithms currently available.
The tool takes recorded crime data as its input,
and applies a learning algorithm to determine the
optimal parameters for crime prediction. These can
then be applied to identify the streets at greatest
risk at future times – typically one day ahead – for
a range of crimes. The tool produces maps which
display the highest risk streets against a map of the
area, which can be provided to ofcers or used in
operational briengs. As well as providing higher
accuracy, the identication of streets allows patrol
activity to be pinpointed to the locations at greatest
risk.
TOOLS
30|31
MAP-MATCHING
Any analysis of police behaviour relies crucially
on knowing exactly where ofcers have travelled.
Although GPS technology facilitates this to an
extent, reconstructing paths through the network is
far from straightforward: GPS readings are subject
to errors, and low sampling rates mean that inter-
polating between signals can be problematic. To
address this, we have developed an algorithm for
this ‘map-matching’ task which is packaged as a
generically-available tool.
Given any GPS tracking data as its input, our
tool infers the most likely route through the net-
work, adjusting for reading errors and characteris-
tics of the streets involved. It does this probabilis-
tically, so that the likelihood that the inferred route
is correct can be quantied. The tool can be applied
in policing to produce detailed movement logs, and
also has potential application in other transport-re-
lated settings.
SUPPLY AND DEMAND
The geospatial processing techniques we have de-
veloped can be applied to police GPS data to infer
exactly which routes through the street network
have been taken by ofcers in the course of patrol.
This provides a complete record of police activity,
and can be taken as a measure of how the ‘supply’
of policing – in the sense of its visual presence – is
distributed. By comparing this to the distribution
of ‘demand’, such as calls for service, the extent
to which supply and demand are aligned can be
explored.
We have produced a tool which allows this
alignment to be visualised and quantied. The tool
reads live tracking data collected by the police, and
– for any analysis period chosen by the user – shows
which areas and streets have received levels of patrol
that are disproportionate to the volume of incidents
occurring there. This can be used as a decision
support tool for commanding ofcers in identifying
areas which may be either under- or over-policed.
PATROL ROUTING
As described in Routing strategies for police patrol,
our research has sought to address the issue of
patrol routing in policing: a critical problem in the
translation from data-driven insight into operation-
al procedure. We have implemented the dynamic
patrol routing strategy we developed as a tool which
can be used in real-world policing environments to
support efcient ofcer tasking.
The tool suggests efcient strategies for the
patrolling of specied hotspots, such as those
identied by our predictive tool. The algorithm
can be tuned to particular operational settings –
with ofcer numbers specied, for example – and
accounts for the conguration of the local street
network. By integrating this with existing deploy-
ment systems, ofcer tasking can be supported in an
evidence-based way.
BEHAVIOUR AND ACTIVITY ANALYSIS
We have implemented our framework for the iden-
tication of activity groups from individual GPS
traces as a tool that can be used to analyse the be-
haviour and activity patterns of foot patrol ofcers.
This is of use for both performance evaluation and
for identifying aspects of behaviour which could be
improved.
The tool can visualise high-volume GPS traces
in space and time, identify the places and time peri-
ods of high patrol intensity, as well as grouping of-
cers who share similar patterns. It can also provide
summaries of the time spent on different activities
by different type of police ofcers. This can be used
as a data-driven component of performance evalua-
tion in an operational policing context.
32|33
The path to the following policy responses leads
from the research questions and analytics that have
been developed throughout the CPC project. Each
requires innovation in the creation, maintenance
and analysis of data resources that are already a
by-product of routine policing activities. In some
instances the policy responses also require linkage of
data arising out of policing in environments that are
secure, while other responses also entail the supple-
mentation or even partial replacement of conven-
tional survey procedures with other instruments to
monitor the ways in which reassurance policing is
received.
POLICY GOAL: IMPROVE PUBLIC
CONFIDENCE IN POLICING
Public condence in policing is most obviously
shaped by the likelihood of being a victim of crime,
but depends also on the visibility of policing,
the victim experience, and the ways in which the
social and human capital in our communities can
be harnessed in partnership activities. Condence
is fundamentally built upon trust, and the CPC pro-
ject has identied signicant gaps in trust between
and among the stakeholders in both communities.
The issues that matter most to the public are per-
ceived differently by experts, and limited awareness
of engagement activities and initiatives have been
revealed by the public during our interviews.
Recommendations:
1) Work needs to be done on both sides to increase
the public’s understanding of police work and foster
better relationships, through better communication
and greater community engagement.
2) This requires re-thinking the design and imple-
mentation of public attitude surveys to better reect
the activity patterns of citizens (since public con-
dence is not shaped exclusively in the residential
setting) as well as the activities of police ofcers.
The current public attitude survey conducted by
MOPAC should be fully reviewed in order to ac-
commodate events such as ‘signal crimes’. The new
design should remain representative of the popula-
tion that the MPS serves, but should better reect
the accountability of individual Borough police
forces.
3) Redesigned public attitude surveys should be
augmented with the conscientious use of social me-
dia sources, carefully reweighted to ensure that the
attitudes and views of non-users are appropriately
represented.
4) The small area estimation and analysis of public
condence with geodemographics show that public
condence interventions should be targeted local-
ly and in a way that is specic to geodemographic
types.
Policy priority: This requires reappraisal of existing
survey instruments, development of new ones, and
greater efforts to explain methods and achieve ‘buy
in’ from front line ofcers. The new tools and data
resources should be free to use as part of the Open
Data movement that is shaping similar initiatives in
other areas of public service delivery.
POLICY GOAL: REDUCE COSTS
AND IMPROVE EFFICIENCY
(“DO MORE FOR LESS”)
Borough commanders are constantly updating their
strategies to improve operational efciency under
intense resource constraints. In addition to that
concerning crime occurrence, data on ofcer activity
patterns can provide valuable feedback and evalu-
ation of performance, which should be considered
with higher weight in their decision-making. To fa-
cilitate this, simple tools to help ofcers understand
the issues are required.
Recommendations:
1) Further development of the CPC strategic
planning tools will give the commanding ofcers
of individual police forces a clear picture of how
routine activities function and how emergencies can
be coped with. The same tools will also be of use in
making budgetary or procurement decisions.
2) Tools for the analysis of supply and demand will
identify where scarce resources in patrolling might
be more effectively redeployed, and to achieve more
efcient deployments of ofcers between neighbour-
hoods, over time.
POLICY IMPLICATIONS
34|35
3) Behaviour and activity analysis tools will offer
insight into the micro-level behaviour of police
ofcers during routine patrol, which might be used
to identify innovative (but undocumented) front
line practices that might be more widely adopted.
Conversely, less effective activities could be phased
out over time.
Policy priority: The prototype tools developed in
CPC need to be completed and implemented as vi-
able operational products so that they can be widely
adopted.
POLICY GOAL: CRIME REDUCTION
AND PREVENTION
Coarse grid-based heatmaps are currently used to
guide frontline ofcers to predicted high risk areas,
but the grid frames are neither intuitive nor appro-
priate for patrols conducted along street segments.
Partly as a consequence, many ofcers have reser-
vations about their utility, and are reluctant to use
them. Even when ofcers are open to using predic-
tive maps, they are difcult to use in operational
situations.
Recommendations:
1) The results of the CPC research provide road
network-based crime predictions for different crime
types at the street segment level, which can be used
to guide police ofcers to the right place at right
time for crime prevention. The power of the pre-
dictive approach can be evaluated using empirical-
ly-based accuracy evaluation measures.
2) Integration of our patrolling tool within an
online central control system has the potential to
provide guidance to frontline ofcers in selecting
areas and routes to patrol, thereby complementing
and supporting existing policing procedures. Its use
in this way can contribute to the improvement of
the efciency of routine patrolling, while also mak-
ing operational practice more resilient to emergency
situations.
Policy priority: Frontline ofcers should be encour-
aged to accept the concept of intelligent patrol, and
to use the network-based predictive maps to sup-
plement their own experience of where and when
it is best to patrol. The system of online patrolling
might also be developed so that frontline ofcers
can be guided and coordinated by the control centre
in an efcient way, though this will also require
fundamental shifts in behaviour and so may form
part of a wider initiative.
IN SUMMARY
Policing is facing great challenges and opportuni-
ties. A fundamental paradigm shift is required to
capitalise upon the value of Big Data for intelligent
policing. Intelligence is not only about prediction,
but also how to act on it, and how to evaluate the
action. This requires behaviour change from sen-
ior and frontline ofcers alike in order to use the
insights that can be gained from police data, as part
of wider adoption of evidence-based policing. Adop-
tion of the tools developed here requires training,
but also a greater openness to the core organising
concepts of predictive policing. We believe that
the policy implications specically set out here will
benet digital policing, not only in London, but in
other large cities across the UK and beyond.
Adepeju M and Cheng T (2014).
Prospective space-time scan statistics
(STSS) for crime prediction. GISRUK
2014 Conference (Glasgow, UK).
Adepeju M, Cheng T, Shawe-Taylor J
and Bowers K (2015). A new metric of
crime hotspots for operational policing.
GISRUK 2015 Conference (Leeds, UK).
Adepeju M, Rosser G and Cheng T
(2016). Novel evaluation metrics for
sparse spatio-temporal point process
hotspot predictions – a crime case
study. International Journal of Geographi-
cal Information Science. DOI:10.1080/13
658816.2016.1159684.
Bowers K and Guerette R (2014). Ef-
fectiveness of situational crime preven-
tion. In G Bruinsma and D Weisburd
(eds.), Encyclopedia of Criminology and
Criminal Justice. Springer Verlag, New
York. DOI: 10.1007/978-1-4614-
5690-2_553.
Chen H, Cheng T and Wise S (2015).
Designing daily patrol routes for po-
licing based on ant colony algorithm.
ISPRS Annals of Photogrammetry, Remote
Sensing and Spatial Information Sciences,
II-4/W2. DOI: 10.5194/isprsan-
nals-II-4-W2-103-2015.
Cheng T and Adepeju M (2013).
Detecting emerging space-time crime
patterns by prospective STSS. In Pro-
ceedings of the 12th International Confer-
ence on GeoComputation, Wuhan, China,
May 2013.
Cheng T and Adepeju M (2014). Modi-
able Temporal Unit Problem (MTUP)
and its effect on space-time cluster
detection. PLOS ONE, 9(6). DOI:
10.1371/journal.pone.0100465.
Cheng T, Haworth J, Anbaroglu B,
Tanaksaranond G and Wang J (2014).
Spatiotemporal data mining, In MM
Fisher, P Nijkamp (eds.), Handbook of
Regional Science, Springer Verlag, Berlin
Heidelberg. DOI: 10.1007/978-3-642-
23430-9_68.
Cheng T and Wicks T (2014). Event
detection using Twitter: a spatio-tem-
poral approach. PLOS ONE, 9(6). DOI:
10.1371/journal.pone.0097807.
Davies T and Bowers K (2015). Quan-
tifying the deterrent effect of police
patrol via GPS analysis. GISRUK 2015
Conference (Leeds, UK).
Davies T and Johnson SD (2014).
Examining the relationship between
road structure and burglary risk via
quantitative network analysis. Journal
of Quantitative Criminology, 31(3). DOI:
10.1007/s10940-014-9235-4.
Davies T and Marchione E (2015).
Event networks and the identica-
tion of crime pattern motifs. PLOS
ONE, 10(11). DOI: 10.1371/journal.
pone.0143638.
Gale C, Singleton A and Longley P
(2014). Does London need a separate
geodemographic classication? GIS-
RUK 2014 Conference (Glasgow, UK).
Gale C, Singleton A and Longley P
(2016). Creating the 2011 Area Classi-
cation for Output Areas (2011 OAC).
Journal of Spatial Information Science, in
press.
Gale C, Singleton A and Longley P
(2015). Proling burglary in London
using geodemographics. GISRUK 2015
Conference (Leeds, UK).
Johnson SD, and Bowers K (2013).
Near repeats and crime forecasting. In
G Bruinsma and D Weisburd (eds.),
Encyclopedia of Criminology and Criminal
Justice. Springer Verlag, New York.
DOI: 10.1007/978-1-4614-5690-
2_210.
Kowalska K, Manley E and Shawe-Tay-
lor J (2015). Routing regions: ex-
tracting spatial patterns from drivers’
preferences. Plurimondi. An International
Forum for Research and Debate on Human
Settlements, 7(15).
Kowalska K, Shawe-Taylor J and Long-
ley P (2015). Data-driven modelling
of police route choice. GISRUK 2015
Conference (Leeds, UK).
Lai J, Cheng T and Lansley G (2015).
Spatio-temporal patterns of passengers’
interests at London tube stations. GIS-
RUK 2015 Conference (Leeds UK).
Rosser G (2015). Crime prediction
with the self-exciting point process.
Technical report.
Rosser G and Cheng T (2014). Point
process models for prospective crime
analysis. GISRUK 2014 Conference
(Glasgow, UK).
Rosser G and Cheng T (2015). A
self-exciting point process model for
predictive policing: implementation
and evaluation. GISRUK 2015 Confer-
ence (Leeds, UK).
Shen J and Cheng T (2015). Group be-
haviour analysis of London foot patrol
police. GISRUK 2015 Conference (Leeds,
UK).
Shen J and Cheng T (2015). Clus-
tering analysis of London police foot
patrol behaviour from raw trajectories.
In Proceedings of the 13th International
Conference on GeoComputation, Richardson,
USA, May, 2015.
Shen J and Cheng T (2015). Cluster-
ing analysis of ofcer’s behaviours in
London police foot patrol activities.
ISPRS Annals of Photogrammetry, Remote
Sensing and Spatial Information Sciences,
II-4/W2. DOI: 10.5194/isprsan-
nals-II-4-W2-143-2015.
Shen J and Cheng T (2016). A
framework for identifying activity
groups from individual space-time
proles. International Journal of Ge-
ographical Information Science. DOI:
10.1080/13658816.2016.1139119.
Singleton A and Longley P (2015). The
internal structure of Greater London: a
comparison of national and regional ge-
odemographic models. Geo: Geography
and Environment, 2(1). DOI: 10.1002/
geo2.7.
Skarlatidou A, Cheng T and Haklay M
(2013). Guidelines for trust interface
design for public engagement Web
GIS. International Journal of Geograph-
ical Information Science, 27(8). DOI:
10.1080/13658816.2013.766336.
Williams D, Cheng T and Haworth J
(2014). Examining the spatio-temporal
structure of public condence in the
police. GISRUK 2014 Conference (Glas-
gow, UK).
Williams D, Haworth J and Cheng T
(2015). Exploratory spatiotemporal
data analysis of public condence in
the police in London. GISRUK 2015
Conference (Leeds, UK).
Wise S and Cheng T (2015). A model
ofcer: an agent-based model of polic-
ing. GISRUK 2015 Conference (Leeds,
UK).
Wise S and Cheng T (2016). How
ofcers create guardianship: an agent-
based model of policing. Transactions in
GIS. DOI: 10.1111/tgis.12173.
PUBLICATIONS
SpaceTimeLab is a multi-disciplinary
research centre at UCL Department of Civil,
Environmental and Geomatic Engineering.
It brings together researchers from a diverse
set of elds, in geomatics, GIScience,
geography, computer science, crime science,
mathematics, social science, and transport.
SpaceTimeLab’s mission is to generate
actionable insights from geo-located and time-
stamped data for government, business and
society. Using integrated space-time thinking,
we develop theories, methods and platforms
for prediction, proling, visualisation and
simulation
Our current project portfolio covers four key
themes:
1. Transport and Mobility
2. Security and Policing
3. Business Intelligence
4. Environmental Resilience
We are also developing applications in the
following areas with new partners:
5. Health
6. Economics
Website www.ucl.ac.uk/spacetimelab
Twitter @SpaceTimeLab
Crime Po icingCitizenship
010001110010110111010101011100011
101
g
g
111000010101101011100001
10
p
p
0
p
p
0
11
000
1
0
111space-time interactions of dynamic networks
11
0000
1
0
1
0
1
0
1
0
11
0
1
0
1
0
1
0
1111
000
1
0
1
0
111
000
1
0
11
0000
11
00
