BookPDF Available

CPC: Crime, Policing and Citizenship - Intelligent Policing and Big Data


Abstract and Figures

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 geoinformatics, crime science, computer science and geography 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 analytics and emergent network complexity, in order to uncover patterning and interactions in crime, policing 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 challenging 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 approach to data-driven policing, using daily reported crime statistics, GPS traces of foot and vehicular patrols, surveys of public attitudes and geo-temporal demographic data of changing community structure. The analytic focus takes a spatio-temporal perspective, reflecting the strong spatial and temporal integration of criminal, policing and citizen activities. Street networks are used throughout as a basis for analysis, reflecting 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 ‘intelligent policing’ which embodies the key issues arising in the transition from Big Data into actionable insights. Police intelligence should go beyond current practice, incorporating not only the prediction 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 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 offices in strategic planning, and to provide guidance to frontline officers in daily patrolling. To evaluate police performance, models and tools have been developed to aid identification of areas requiring greater attention, and to analyse the patrolling behaviours of officers. Methods to understand and model confidence in policing have also been explored, suggesting strategies by which confidence in the police can be improved in different population segments and neighbourhood areas. A number of tools have been developed during the course of the project include 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 confidence that we suggest will contribute to the transformation and improvement of the relationship between the public and police.
Content may be subject to copyright.
Crime Po icingCitizenship
111space-time interactions of dynamic networks
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
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
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.
Chadwick Building
University College London
Gower Street
T +44 (0)20 7679 2738
The Crime, Policing and Citizenship (CPC):
Space-Time Interactions of Dynamic Net-
works project was supported by EPSRC
under grant EP/J004197/1.
Executive Summary
– Challenges facing digital policing
– Big Data in policing
– Intelligent data-driven policing
– Our approach
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 ocer tasking
– Routing strategies for foot patrol
Public condence & trust in policing
– Trust and condence in policing
– Small area estimation of public condence
– Geodemographics and public condence
– Predictive mapping
– Map-matching
– Supply and demand
– Online cooperative patrol routing
– Behaviour and activity analysis
Policy implications
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
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, reecting the strong spatial and tem-
poral integration of criminal, policing and citizen
activities. Street networks are used throughout as
a basis for analysis, reecting 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 ofces 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
identication of areas requiring greater attention,
and to analyse the patrolling behaviours of ofcers.
Methods to understand and model condence in po-
licing have also been explored, suggesting strategies
by which condence 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 condence that we
suggest will contribute to the transformation and
improvement of the relationship between the public
and police.
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
The mission of policing is evolving, with issues
such as public condence gaining increased rec-
ognition as important police performance metrics.
The improvement of condence is a key priority for
policymakers such as the Mayor’s Ofce 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
reected in public condence 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% efciency 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 condence over the same period. ‘More for
less’ is the mantra for policing today in London.
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 ofcer 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 condence in the London context.
When combined with geodemographic data, these
Local National
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 inuence 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 insufcient 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 condence.
These issues can only be addressed by using
the datasets available to the police in an integrated
way, modelling and accounting for the signicant
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 efciency 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.
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 specic 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’ identied by
many predictive algorithms, for example, contain
many road sections and it is often unclear exactly
where ofcers 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 efciency, accuracy of
prediction is paramount. This requires renement
of analytical techniques for specic 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 ofcers, 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 ofcers can be coordinated
as efciently as possible.
Finally, we emphasise the role of feedback in the
evaluation and renement 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 ofcers, and
b) whether they had their intended effects. The use
of tracking technology to monitor both compliance
and efcacy can play a vital role in developing an
evidence base.
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, reecting 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 scientic 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 condence
analysis, these form a suite of tools that can be used
to support data-driven policing in an operational
7 priority
confidence Total cost
Figure 2: In London, the Mayor’s Police and Crime Plan
sets ambitious objectives for various aspects of police
Figure 3: The spatio-temporal patterns formed by
crime, policing and citizenship activity form dynamic,
interdependent networks.
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.
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
ofcers at the locations and times at which crime is
most likely focuses policing resources where they
are most needed, resulting in improved efciency,
reduced crime rates and greater public satisfaction.
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).
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 inuence 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).
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
proles can be determined using a machine learning
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 eects. Risky
locations are mapped to aid police in planning patrol
Findings Real-world crime data exhibits clustering
in space and time. Application of a predictive model
identies the locations of future crimes with high
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 specied
time dierence of one another. The blue region indicates
crimes which occur more frequently than would be
expected in the absence of clustering.
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.
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 prole 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 prole varies by location and
crime type; the SEPP model is tailored to the specif-
ic situation.
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 prole 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
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
Figure 3: The crime triggering prole in time and space
for burglaries in Camden.
0 10 20 30 40 50 60 70 80 90
Time (days)
Triggering intensity
Cumulative prob. of triggering a crime
600 400 200 0 200 400 600
Distance (metres)
Triggering intensity
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.
One of the primary benets 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 difcult to identify exactly
where risk lies or where ofcers should patrol. Street
segments, in contrast, are well-dened micro-units,
which allow risk to be estimated with much greater
precision. Furthermore, since ofcers patrol by fol-
lowing dened routes along streets, network-based
risk maps are easier to interpret in practice.
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 inu-
ence the distribution of crime.
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), reecting the notion that risk is transmitted
down streets rather than uniformly through space.
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
Figure 1: Spatial kernel function implemented in net-
work space (represented by black lines). A kernel is
placed at all locations of previous crimes.
We also developed a novel training protocol for
the algorithm, which identies 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 identied. 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.
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.
Our results indicate that using the street network
as the spatial basis for prospective mapping leads to
signicant 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.
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.
In the majority of work to date, predictive accuracy
the extent to which the locations of future crimes
are correctly identied – 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 signicant issue for police
ofcers: 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.
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 signicance test that allows the difference
in accuracy between two methods to be quantied
Compactness – how concentrated and connected
the areas identied are. Compactness reects 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 reect long-term risk, and those which are
highly variable from day to day (see Figure 2).
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.
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
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.
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 reect differing levels of sparseness in the
underlying patterns.
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 benet. 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.
The framework we have developed allows predictive
approaches to be evaluated in a signicantly 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 identied 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
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
Much crime prevention activity depends, at some
level, on the ability of police ofcers to deter crime
by being physically present in a place. This idea
forms the basis for a signicant 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: ofcers are
sent to the predicted locations in order to discour-
age or interrupt the anticipated crimes.
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 ofcers, 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 ofcer 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.
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.
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 quantied: do streets survive for longer than
they would otherwise have done?
Aims To measure the extent to which routine visible
policing deters crime.
Methods Ocer 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 eect 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.
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
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 signicant.
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 signicant 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
efciency 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 dierence
between observed values and those in the absence
of patrol.
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.
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, trafc, or monitor-
ing activities that purposely avoid major roads.
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 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.
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 signicantly 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.
Path length (metres)
020000 3000010000
least distance
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 signicant improvements in
accuracy when compared to off-the-shelf routing
Our proposed model shows signicant improve-
ments over existing methods in replicating the
routing behaviour of police ofcers. This provides
justication 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
Late Shift
Early Shift
Night Shift
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’.
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 ofcers 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
ofcers who share similar patterns, is of use for both
performance evaluation and for identifying aspects
of behaviour which could be improved.
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
ofcer movements using GPS is one such source,
and provided the basis for our work.
In the proposed framework, APLS traces are rst
map-matched onto the street network to identify
the street segments that the ofcers truly visited.
The places and times at which particularly high lev-
els of ofcer presence were observed are then identi-
ed using a network-based variant of the clustering
algorithm DBSCAN. These regions of space-time
are dened as Spatio-Temporal Regions of Interest
(ST-ROIs), and represent locations which feature
particularly prominently in patrol behaviour
(Figure 1).
Aims To characterise and classify the movement pat-
terns of foot patrol ocers.
Methods Using ocer GPS traces, spatio-temporal
regions of high patrol intensity were identied. Ocer
behaviours were then classied according to their
visits to these locations.
Findings Ocer activity clusters around certain prom-
inent locations, and distinctive behavioural signatures
can be identied in their visits to these.
Figure 1: During the study period, 28 ST-ROIs are
detected for foot patrol ocers by space-time
DBSCAN – these are identied by colour here.
Once ST-ROIs have been identied, individual
ofcers’ activity patterns can be expressed in terms
of their visits to these locations. An individual’s
behaviour is dened as his/her prole 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 proles, ofcers can be partitioned into sub-
groups based on the similarity of their activities in
space and time (see Figure 3).
When the ofcer subgroups are examined in turn,
a qualitative understanding of their characteristic
behaviours can be gained. One group of ofcers,
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
ofcers are seen to deploy for long stints of patrol
after beginning at a Camden station.
The application of our framework has the potential
to offer unprecedented insight into the micro-level
behaviour of police ofcers during routine patrol.
The framework extends traditional ideas of time
budget allocation, and is capable of proling the
activity patterns of ofcers in both space and time
by dening 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 ofcers. Findings such as these could
be used to suggest behavioural modications which
could lead to more efcient 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: Simplied representation of two example
ocers’ movements, showing (a) trajectories in space-
time; (b) simplied representation showing the visiting
sequence of ST-ROIs.
Figure 3: Taxonomy tree showing the clustering of
ocers with dierent patrol patterns (ID numbers
As part of the development of policing, it is natural
for senior ofcers to consider the potential of chang-
es in operational practice to improve the efciency
of policing. These changes might be technological,
such as the adoption of handheld devices, or behav-
ioural, such as the renement of tasking protocols.
Variations in resource levels can impact the ability
of ofcers to provide service, while terrorist attacks
or special events drastically affect the set of respon-
sibilities ofcers 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 difcult task: the
potential for knock-on effects and feedback loops
means that such questions cannot be answered
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 articial 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 articial ‘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.
Aims To provide a tool to explore the impact of chang-
es to police operations.
Methods An agent-based simulation model for ocer
behaviour and tasking was developed, based on in-
sight gained for active ocers.
Findings The model replicates real-world ocer be-
haviour in terms of the spatial distribution of coverage,
suggesting that it has potential use as a real-world
decision support tool.
to incident?
Go to incident site
(ignore traffic laws) Deal with incident
Wait until Transport
Vehicle arrives
Return to station to
make a report
Figure 1: The decision ow for a simulated ocer in
‘Responding’ status.
In order to explore these questions, we developed an
agent-based model for the behaviour and tasking of
police ofcers. The construction of the model was
informed by consultation with police ofcers, and
incorporated several aspects: the dispatch process,
the travel of ofcers to incidents, and self-directed
patrol behaviour. Simulated ofcers 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.
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
ofcers 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 ofcers
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 signicant improvement over corresponding null
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
rening 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 ocers in
Camden, derived from GPS traces.
Figure 3: The spatial distribution of simulated ocer
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 efciently operationalised,
however, presents a signicant logistical challenge,
especially when police resources are limited and
hotspots are many.
In most operational contexts, hotspot patrolling is
implemented by tasking ofcers 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
efciently as possible, it is necessary to develop a
cooperative routing strategy.
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.
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.
Aims To propose guidelines and measures for the
performance of police patrol and to design a routing
strategy for eective 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
quantied using 5 objective measures, and the novel
routing strategy developed oers substantial improve-
ments in eciency.
Figure 1. From guidelines to BAPS.
Online Bayesian Ant-based Patrolling Strategy
Table 1. Guidelines for police patrol routing strategy
Measure Meaning
Eciency (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 dier-
ent areas (spatial scalability) and with
dierent numbers of patrollers (team
Unpredictability (U) To keep the patrol routes unpredictable
Robustness (R) To remain eective patrolling when
some patrollers are dispatched for
Receive Tasks
Await Tasks
Using pheromones
to record visit history
Bayesian decision-making
(including multiple factors)
Emergency Patrol
Report Task
Efficiency Flexibility Unpredictability Scalability Robustness
Respond to
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 dened 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
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.
This research represents a signicant step in the
development of intelligent patrolling, which has the
potential to substantially improve the operational
efciency of targeted interventions. The perfor-
mance measures introduced allow patrol to be eval-
uated in objective and quantitative terms, meaning
that behaviour can be rened 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.
BAPS 1700 1623 2055 1.037 3.30%
CCPS 2039 2030 497 0.996 4.10%
Change (%) -16.60 -20.05 -- -- --
E- Eciency; F – Flexibility; U – Unpredictability;
S: (Team) Scalability; R - Robustness
Table 2. Comparison of CCPS and BAPS performance
Receive Tasks
Await Tasks
Using pheromones
to record visit history
Bayesian decision-making
(including multiple factors)
Emergency Patrol
Report Task
Efficiency Flexibility Unpredictability Scalability Robustness
Respond to
A growing number of studies report lower levels
of public trust and condence in governments and
their institutions and public services, which inu-
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 condence is shrinking. In the
UK, where the ideological emphasis is on policing
by consent, this is particularly problematicand thus
‘promoting public condence’ in policing has been a
top priority of the government agenda.
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
signicant concerns about the validity of the results
and subsequently the effectiveness of measures and
policies which attempt to improve public trust and
condence in policing.Our research attempts to
address this gap.
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 inuence 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 inuenced by
factors such as cultural or educational background,
prior experiences, misconceptions and the lack of a
shared public understanding of police practices.
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 oers insight into the factors
that inuence 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
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 inuence 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.
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 condence in policing
amongst experts. In other words, the expert mental
models of experts in this topic vary signicantly, 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 signicant gaps and misconceptions, as
well as a limited awareness of engagement activities
and initiatives.
Our analysis suggests that the issue of public trust
continues to pose a signicant 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.
Puc condence in the police is a state in which the
public regard the police as competent and capable
of fullling 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 condent in the police
are more likely to be cooperative, compliant and
crucially to supply the tips which inform proactive
policing operations.
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 condence. Public condence 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 condence intervention strategy.
Aims To estimate and predict public condence in the
police at the neighbourhood level.
Methods A spatio-temporal interaction Bayesian
hierarchical model was developed to model public
condence trends in space and over time.
Findings Public condence in policing is autocorrelat-
ed in space-time. This autocorrelation can be used to
estimate and predict public condence at the neigh-
bourhood level.
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 condence
levels above the London-wide average.
A spatiotemporal Bayesian hierarchical modelling
approach enabled the estimation and prediction
of public condence 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.
Public condence 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
condence levels to vary by neighbourhood and
temporal quarter. It also includes spatial, temporal
and spatiotemporal components, and was used to
estimate and predict public condence levels at the
neighbourhood level (see Figure 3).
This research contributes evidence of the autocorre-
lation of public condence 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 ofcers to design
targeted strategies to improve public condence.
The estimates and predictions obtained can better
enable ofcers to prepare for the proactive pub-
lic condence interventions required to meet the
concerns of the local community. This approach also
allows hot spots and cold spots of public condence
to be better identied and examined.
3D Spatiotemporal variogram of confidence in the police
temporal separation
spatial separation
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
Figure 3: Group of maps of estimated and predicted
public condence levels in London for the period April
2012- March 2013. Dark blue area indicate condence
levels above the London-wide average.
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 ofcers, 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 condence is seen as
an end in itself. Because of this, many forces have
set explicit targets for the improvement of public
condence; the MPS, for example, has been required
to achieve a 20% improvement between 2013 and
In seeking to improve public condence, 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 condence, for
example, or by addressing specic concerns amongst
others. This is extremely challenging, however: the
relationships between condence and socio-demo-
graphic indicators are subject to complex interac-
tions, making generalisation difcult. Furthermore,
aggregate-level relationships are of limited use in
informing strategies for condence improvement:
it is far from clear how the concerns of a particular
age group can be addressed at a city-wide level, for
One way in which the inuence 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
inuences public attitudes. If areas can be proled
in these terms, then efforts to improve public con-
dence can be tailored spatially, according to the
needs of particular neighbourhoods.
Aims To explore and simplify the relationship between
demographic factors and condence.
Methods Survey response data is explored using a ge-
odemographic classication, which identies socially
similar groups.
Findings Geodemographic supergroups display
distinct trends in their attitudes towards police. These
classications can be used to tailor eorts to improve
condence to the needs of particular areas.
A number of geodemographic classications are
available in the UK, including one which is spe-
cic to London and was developed by our research
team: the 2011 London Output Area Classication
(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 specically focussed on
issues of public condence.
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 signicant because it implies that these
groups do indeed display idiosyncratic character-
istics with respect to their perception of police.
Knowledge of an area’s classication therefore
allows inferences to be made about the level of con-
dence of people living there.
The use of geodemographics has the potential to
signicantly improve the precision with which
police efforts to improve public condence 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 classication 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 Classication,
with areas coloured according to geodemographic
Produced by Chris Gale
Contains National Statistics data
© Crown copyright and database right 2016
Contains Ordnance Survey data
© Crown copyright and database right 2016
0 5 10
A - Intermediate
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
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
Effectiveness in crime prevention and protection Community Commitment/ Engagement
Fair Treatment Alleviating Local Anti-Social Behaviour
Feeling informed about local policing
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
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 unied
system for data-driven policing support.
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 signicant 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 ofcers or used in
operational briengs. As well as providing higher
accuracy, the identication of streets allows patrol
activity to be pinpointed to the locations at greatest
Any analysis of police behaviour relies crucially
on knowing exactly where ofcers 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 quantied. The tool can be applied
in policing to produce detailed movement logs, and
also has potential application in other transport-re-
lated settings.
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 ofcers 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
We have produced a tool which allows this
alignment to be visualised and quantied. 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 ofcers in identifying
areas which may be either under- or over-policed.
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 efcient ofcer tasking.
The tool suggests efcient strategies for the
patrolling of specied hotspots, such as those
identied by our predictive tool. The algorithm
can be tuned to particular operational settings –
with ofcer numbers specied, for example – and
accounts for the conguration of the local street
network. By integrating this with existing deploy-
ment systems, ofcer tasking can be supported in an
evidence-based way.
We have implemented our framework for the iden-
tication 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 ofcers.
This is of use for both performance evaluation and
for identifying aspects of behaviour which could be
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 ofcers. This can be used
as a data-driven component of performance evalua-
tion in an operational policing context.
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
Public condence 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. Condence
is fundamentally built upon trust, and the CPC pro-
ject has identied signicant 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.
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 reect
the activity patterns of citizens (since public con-
dence is not shaped exclusively in the residential
setting) as well as the activities of police ofcers.
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 reect
the accountability of individual Borough police
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
4) The small area estimation and analysis of public
condence with geodemographics show that public
condence interventions should be targeted local-
ly and in a way that is specic to geodemographic
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 ofcers. 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.
Borough commanders are constantly updating their
strategies to improve operational efciency under
intense resource constraints. In addition to that
concerning crime occurrence, data on ofcer 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 ofcers understand
the issues are required.
1) Further development of the CPC strategic
planning tools will give the commanding ofcers
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
efcient deployments of ofcers between neighbour-
hoods, over time.
3) Behaviour and activity analysis tools will offer
insight into the micro-level behaviour of police
ofcers 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
Coarse grid-based heatmaps are currently used to
guide frontline ofcers 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 ofcers have reser-
vations about their utility, and are reluctant to use
them. Even when ofcers are open to using predic-
tive maps, they are difcult to use in operational
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 ofcers 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 ofcers 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 efciency of routine patrolling, while also mak-
ing operational practice more resilient to emergency
Policy priority: Frontline ofcers 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 ofcers
can be guided and coordinated by the control centre
in an efcient way, though this will also require
fundamental shifts in behaviour and so may form
part of a wider initiative.
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 ofcers 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 specically set out here will
benet 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
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-
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-
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:
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-
Cheng T and Wicks T (2014). Event
detection using Twitter: a spatio-tem-
poral approach. PLOS ONE, 9(6). DOI:
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:
Davies T and Marchione E (2015).
Event networks and the identica-
tion of crime pattern motifs. PLOS
ONE, 10(11). DOI: 10.1371/journal.
Gale C, Singleton A and Longley P
(2014). Does London need a separate
geodemographic classication? 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
Gale C, Singleton A and Longley P
(2015). Proling 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-
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,
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 ofcer’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-
Shen J and Cheng T (2016). A
framework for identifying activity
groups from individual space-time
proles. International Journal of Ge-
ographical Information Science. DOI:
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/
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:
Williams D, Cheng T and Haworth J
(2014). Examining the spatio-temporal
structure of public condence in the
police. GISRUK 2014 Conference (Glas-
gow, UK).
Williams D, Haworth J and Cheng T
(2015). Exploratory spatiotemporal
data analysis of public condence in
the police in London. GISRUK 2015
Conference (Leeds, UK).
Wise S and Cheng T (2015). A model
ofcer: an agent-based model of polic-
ing. GISRUK 2015 Conference (Leeds,
Wise S and Cheng T (2016). How
ofcers create guardianship: an agent-
based model of policing. Transactions in
GIS. DOI: 10.1111/tgis.12173.
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, proling, visualisation and
Our current project portfolio covers four key
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
Twitter @SpaceTimeLab
Crime Po icingCitizenship
111space-time interactions of dynamic networks
... For instance, burglary causality in the morning may differ from the night in urban areas, and crime causality on weekdays may differ from weekends. Moreover, as stated in [7], there is a temporal patterning of crime incidents. The occurrence of a crime actively increases the probability of further incidents in the vicinity [7]. ...
... Moreover, as stated in [7], there is a temporal patterning of crime incidents. The occurrence of a crime actively increases the probability of further incidents in the vicinity [7]. ...
... That is an expected result since by increasing the time window size, more challenging is to build an accurate prediction model. This is in line with the crime theory [7], the occurrence of a crime actively increases the probability of further incidents in the temporal vicinity. The experiments show that the models created are always better for the = 2 time window size and get progressively worst until the = 12 time window. ...
Full-text available
High crime rates have become a public health problem in many important cities, according to World Health Organization. Many researchers have been developing algorithms to predict crime occurrences to tackle this problem. The smart cities' environment can provide us enough ubiquitous data, e.g., traffic flow, human mobility, and Points of Interest (POI) information, to feed those predictive policing algorithms and reflect city dynamics. POIs data provide essential information such as geographical location, category, customer reviews, and busy hours. Recent studies have shown that POI geographical locations are useful for predictive policing. In this paper, we aim at predicting crimes in a delimited region around the POIs of a city with new environmental features. We investigate the relevance of POIs location and the semantic and the temporal features from POIs data in our problem. We also propose and analyze different machine learning approaches to train prediction functions based on these features and conduct experiments on real crime data over multiple years. The experiments demonstrate that the popular time feature is more relevant than the historical information about the number of crimes around a POI, but both information is much less critical than the spatio-temporal information. This work is the first that studies the popular time feature extracted from POIs data and historical criminal information for predictive policing from the authors' knowledge. CCS CONCEPTS • Information systems → Geographic information systems. KEYWORDS predictive policing, point of interest, spatial-temporal systems
... By increasingly adopting data processing methodologies for forecasting purposes, security experts have been influenced by similar big-data applications in the corporate world. In intelligence, counterterrorism, policing and peacekeeping, operations have been transformed by the capabilities of big data and predictive analysis to detect unexpected security-related patterns and identify potential threats [32,33]. Predictive analysis with big data offers security experts the promise of safeguarding the future by anticipating the next terrorist attack and detecting potential crimes before they are committed. ...
... The use of forecasting with big data in the intelligence and counterterrorism arena has, not least because of the Edward Snowdon revelations, sparked intense controversy between advocates of the pursuit and expansion of the deployment of this technology and its critics. Advocates of the use of big data and AI in the intelligence domain argue that their effectiveness in this area has long been proven and that many authorities around the world have used them and have experienced great success [32]. These proponents argue that almost everyone today has a digital footprint that can be traced and analyzed, and therefore much data can be collected through the use of mobile phones, computer systems, apps, social networks, electronic communications, and many other technologies [34]. ...
Life in modern society is increasingly connected by networks that link the world around us and create numerous exciting opportunities, new services and advantages for humanity. Yet concurrently, these underpinning networks have provided routes by which potentially dangerous and harmful incidents can propagate quickly and worldwide. This complexity poses a considerable challenge for risk analysis and forecasting. Conventional methods of risk analysis tend to underestimate the probability and impact of risks (e.g. pandemics, financial collapses, terrorist attacks), as sometimes the existence of independent observations is wrongly assumed and cascading errors that can occur in complex systems are not considered. The purpose of this article is to assess critically the potential of big data to profoundly change the current capability for risk forecasting in diverse areas and the assertion that big data leads to better risk predictions. In particular, the focus is on big data implications for risk forecasting in the areas of economic and financial risks, environmental and sustainable development risks, and public and national security risks. The article concludes that big data and predictive analytics offer substantial opportunities for improving risk forecasting but may not replace the significance of appropriate assumptions, adequate data quality and continuous validation.
... The Edward Snowden revelations have shown how intelligence agencies deploy similar technologies as commercial big data companies, and how they have increasingly mobilized these technologies to find unknown patterns and relations (National Security Agency, 2008;GCHQ, 2011). Yet, big data is not only used by businesses and intelligence agencies, but increasingly by the police, border guards, and even humanitarian actors (Akhgar et al., 2015;Meier, 2015;Cheng et al., 2016). Intelligence, counter-terrorism, policing and counterinsurgency have been transformed by the promise of big data and predictive analytics to uncover unexpected patterns and pinpoint potentially suspect 'needles'. ...
... Understood as a reconfiguration of time/space, predictive policing is not only a tool for governing 'others' but also for governing 'the self', as police resources have decreased in a neoliberal age (see Mohler, 2014;Cheng et al., 2016). Or as a RAND evaluation of predictive policing puts it, 'Predictive policing is not fundamentally about making crime-related predictions. ...
From ‘connecting the dots’ and finding ‘the needle in the haystack’ to predictive policing and data mining for counterinsurgency, security professionals have increasingly adopted the language and methods of computing for the purposes of prediction. Digital devices and big data appear to offer answers to a wide array of problems of (in)security by promising insights into unknown futures. This article investigates the transformation of prediction today by placing it within governmental apparatuses of discipline, biopower and big data. Unlike disciplinary and biopolitical governmentality, we argue that prediction with big data is underpinned by the production of a different time/space of ‘between-ness’. The digital mode of prediction with big data reconfigures how we are governed today, which we illustrate through an analysis of how predictive policing actualizes between-ness as hotspots and near-real-time decisions.
... It serves to enhance the criminals' perceived risk of detection, thus preventing potential crimes. Moreover, the visible police presence would increase the public's certainty of punishment and the public trust and confidence in policing (Cheng et al. 2016). ...
Conference Paper
Full-text available
This paper deals with the challenges in police patrol routing when multiple patrollers from different police stations. The patrollers should cover each given site at least once, and they should cover balanced route lengths to avoid discontentment and possible work overload. This routing problem is formulated as a Min-Max Multiple Centre Rural Postman Problem here, which is proposed for the first time. This provides a conceptual base toward finding a suitable routing algorithm for police patrol.
Full-text available
Scientists have an enduring interest in understanding urban crime and developing security strategies for mitigating this problem. This chapter reviews the progress made in this topic from historic criminology to data-driven policing. It first reviews the broad implications of urban security and its implementation in practice. Next, it focuses on the tools to prevent urban crime and improve security, from analytical crime hotspot mapping to police resource allocation. Finally, a manifesto of data-driven policing is proposed, with its practical demand for efficient security strategies and the development of big data technologies. It emphasizes that data-driven strategies could be applied in cities due to their promising effectiveness for crime prevention and security improvement.
Ongoing rapid urbanisation has modernised people’s lives but also engendered the increase in traffic congestion, energy consumption, air pollution, urban crimes, road incidents, etc. With the advance in the Internet of Things (IoT) and 5G, massive geotagged and timestamped (spatio-temporal) data have been collected to monitoring urban environment and processes. There are increased interests in developing urban spatio-temporal (ST) forecasting to make cities greener, safer, and smarter. Recently, deep learning (DL) has been used widely in urban environmental monitoring because of its powerful capability in modelling complex and high dimensional data. Its full potential for urban process prediction is yet to develop due to the irregular network-based spatial structure in many urban processes, temporal non-stationarity, ST heterogeneity and data density variation. In addition, real-world applications at city-scale require fast (or near real-time) training and prediction, capable of dealing with abnormal conditions in real-world scenarios (e.g. missing data and non-recurrent events). To address the challenges, this thesis has developed cutting-edge graph DL models to forecast large-scale urban processes on networks. The contributions of this study are summarised from two aspects. First, from a methodological perspective, we use graphs to unify the representation of all ST urban processes, either dense or sparse. A number of novel contributions are then made towards DL technique by expanding and adapting DL to a spatiotemporal framework for different types of network-based ST forecasting tasks. We propose unified DL models with novel spatial or spectral graph convolutions to forecast both directed and undirected dense urban processes on networks, addressing issues including non-stationary temporal dependency modelling, network-structured spatial dependency modelling and ST heterogeneity. We further tackle the data sparsity issue by developing the first graph DL model with an innovative localised weight sharing graph convolution. The proposed models have scalable structures that can produce citywide ST forecasts in a timely and accurate fashion. The modularity of these models allows to deal with missing data and incorporate external factors for robust forecasting under abnormal conditions, which enhances the chance of the models being used in real-time urban applications. From the urban application point of view, various large, real-world urban datasets, including traffic and crime cases, have been employed to validate that the proposed models can outperform various state-of-the-art benchmarks in terms of accuracy and efficiency. The results derived through these forecasting techniques can be used to address many key growth areas in urbanisation, like human mobility, transportation, and public safety, which has the potential to facilitate future policies and improve the well-being of societies.
This article explores the ability to predict contextual variations in the size of a local population using a model developed by fusing local sensed data sources. The purpose is to demonstrate that pervasive computing and data science can improve knowledge and forecasts about people–place interactions, as an alternative to urban simulations that rely upon static administrative statistics and generalized models of behavior. To demonstrate, a simple question is asked: can we forecast the size of a population in an urban open public space and how it will vary due to dynamic environmental and social conditions? To answer the question, a prediction model is developed from a year of daily WiFi device counts and sensed weather conditions.
The predictive hotspot mapping of sparse spatio-temporal events (e.g., crime and traffic accidents) aims to forecast areas or locations with higher average risk of event occurrence, which is important to offer insight for preventative strategies. Although a network-based structure can better capture the micro-level variation of spatio-temporal events, existing deep learning methods of sparse events forecasting are either based on area or grid units due to the data sparsity in both space and time, and the complex network topology. To overcome these challenges, this paper develops the first deep learning (DL) model for network-based predictive mapping of sparse spatio-temporal events. Leveraging a graph-based representation of the network-structured data, a gated localised diffusion network (GLDNet) is introduced, which integrating a gated network to model the temporal propagation and a novel localised diffusion network to model the spatial propagation confined by the network topology. To deal with the sparsity issue, we reformulate the research problem as an imbalance regression task and employ a weighted loss function to train the DL model. The framework is validated on a crime forecasting case of South Chicago, USA, which outperforms the state-of-the-art benchmark by 12% and 25% in terms of the mean hit rate at 10% and 20% coverage level, respectively.
Full-text available
In urban areas, crime and disorder have been long-lasting problems that spoil the economic and emotional well-being of residents. A significant way to deter crime, and maintain public safety is through police patrolling. So far, the deployment of police forces in patrolling has relied mainly on expert knowledge, and is usually based on two-dimensional spatial units, giving insufficient consideration to the underlying urban structure and collaboration among patrol officers. This approach has led to impractical and inefficient police patrol strategies, as well as a workload imbalance among officers. Therefore, it is of essential importance to devise advanced police patrol strategies that incorporate urban structure, the collaboration of the patrol officers, and a workload balance. This study aims to develop police patrol strategies that would make intelligent use of the street network layout in urban areas. The street network is a key component in urban structure and is the domain in which crime and policing take place. By explicitly considering street network configurations in their operations, police forces are enabled to provide timely responses to emergency calls and essential coverage to crime hotspots. Although some models have considered street networks in patrolling to some extent, challenges remain. First, most existing methods for the design of police districts use two-dimensional units, such as grid cells, as basic units, but using streets as basic units would lead to districts that are more accessible and usable. Second, the routing problem in police patrolling has several unique characteristics, such as patrollers potentially starting from different stations, but most existing routing strategies have failed to consider these. Third, police patrolling strategies should be validated using real-world scenarios, whilst most existing strategies in the literature have only been tested in small hypothetical instances without realistic settings. In this thesis, a framework for developing police patrol strategies based on the urban street network is proposed, to effectively cover crime hotspots, as well as the rest of the territory. This framework consists of three strategies, including a districting model, a patrol routing strategy for repeated coverage, and a patrol routing strategy for infrequent coverage. Various relevant factors have been considered in the strategy design, including the underlying structure of the street network and the collaboration among patrollers belonging to different stations. Moreover, these strategies have been validated by the patrolling scenarios in London. The results demonstrate that these strategies outperform the current corresponding benchmark strategies, which indicates that they may have considerable potential in future police operations.
Full-text available
Crime is a complex phenomenon, emerging from the interactions of offenders, victims, and their environment, and in particular from the presence or absence of capable guardians. Researchers have historically struggled to understand how police officers create guardianship. This presents a challenge because, in order to understand how to advise the police, researchers must have an understanding of how the current system works. The work presents an agent-based model that simulates the movement of police vehicles, using a record of real calls for service and real levels of police staffing in spatially explicit environments to emulate the demands on the police force. The GPS traces of the simulated officers are compared with real officer movement GPS data in order to assess the quality of the generated movement patterns. The model represents an improvement on existing standards of police simulation, and points the way toward more nuanced understandings of how police officers influence the criminological environment.
Full-text available
This paper presents the methodology that has been used to create the 2011 Area Classification for Output Areas (2011 OAC). This extends a lineage of widely used public domain census only geodemographic classifications in the UK. It provides an update to the successful 2001 OAC methodology, and summarizes the social and physical structure of neighbourhoods using data from the 2011 UK Census. We also present the results of a user engagement exercise that underpinned the creation of an updated methodology for the 2011 OAC. The 2011 OAC comprises 8 Supergroups, 26 Groups and 76 Subgroups. Finally, we present an example of the results of the classification in Southampton.
Conference Paper
Full-text available
Detecting crime patterns as they emerge in both space and time can enhance situational awareness amongst security agents and prevent epidemics of crimes in potential problematic areas (Neill and Gorr, 2007). Amongst others, space-time scan statistics (STSS) (Kulldorff et al. 2005) and space -time kernel (Nakaya and Yano, 2010) have been widely used in crime analysis. Stemmed on strong statistical theories, the STSS could provide the significance of the purported crime clusters, and this continues to gain huge popularities for crime hotspot analysis (LeBeau, 2000; Neill and Gorr, 2007; Uittenbogaard and Ceccato , 2011; Cheng and Williams, 2012; Gao et al. 2012). All these works applied STSS to crime clusters detection in a retrospective manner where all clusters within certain frame of time are detected. The approach was found to be very effective for historic analysis of crime outbreaks and near-repeat victimization. However, most STSS-based hotspot analyses were conducted at either region-wide and/or at monthly temporal granularity. This is not appropriate for city-based policing, which requires detailed spatial (local or micro) and temporal (daily) analysis. Few studies have actually attempted prospective detection of clusters with the aim of capturing their growth (emergence) in both space and time simultaneously so as to facilitate early prevention of the phenomenon in question. This was however seen only in epidemiology where outbreaks of diseases were detected employing this approach using the over-the-counter drug sales in Allegheny County from 2/13/04 - 2/12/05 (Neill et al., 2005). However, there is no quantitative evaluation of the significance of the emerging patterns and the rapidness of their emergence. Therefore, the aim of this research is to explore a prospective detection of emerging crime patterns at detailed spatial and temporal scales so as to facilitate proactive policing. In particular, we use the permutation STSS for the detailed crime emerging pattern detection and evaluate their significance as well as the rapidness of detection, by comparing the results with that of retrospective analysis.
Conference Paper
Full-text available
This study examines the existing metrics used in evaluating the effectiveness of area-based crime hotspots for operational policing. We identified some of the limitations of the metric (i.e. Area-to-Perimeter (AP) ratio) used for measuring compactness of hotspots and then proposed a new improved metric called " Clumpiness Index (CI) ". The case study of London Metropolitan police crime dataset features the prediction of 3 different crime types using two different crime predictive methods. The effectiveness of the hotspots was then measured using both AP ratio and CI. The comparison of the results clearly shows that CI is a better metric for measuring the effectiveness of crime hotspots for operational policing.
Conference Paper
Improving public confidence in the police is one of the most important issues for the London Metropolitan Police Service (Met). Public confidence varies over geographic space and changes over time. Spatiotemporal analysis and modelling becomes more manageable with a thorough understanding of the underlying spatiotemporal autocorrelation structure of the phenomena under scrutiny. In this study, exploratory spatiotemporal analysis is conducted on repeated cross-sectional survey data from the Metropolitan Police Public Attitude Survey. This confirmed the presence of second order nonstationarity in public perceptions of the Met police.
This is a welcome addition to any academic library.
Conference Paper
In this small paper we aim at presenting a framework of conceptual representation and clustering analysis of police officers’ patrol pattern obtained from mining their raw movement trajectory data. This have been achieved by a model developed to accounts for the spatio-temporal dynamics human movements by incorporating both the behaviour features of the travellers and the semantic meaning of the environment they are moving in. Hence, the similarity metric of traveller behaviours is jointly defined according to the stay time allocation in each Spatio-temporal region of interests (ST-ROI) to support clustering analysis of patrol behaviours. The proposed framework enables the analysis of behaviour and preferences on higher level based on raw moment trajectories. The model is firstly applied to police patrol data provided by the Metropolitan Police and will be tested by other type of dataset afterwards.
Conference Paper
The main objective of this research is to propose a method for group movement pattern generalisation and classification. To this end, DBSCAN is used on stay points for POI identification. Then, movement features are extracted and selected for the behaviour classification. A kernel-based Support Vector Machine method is developed to infer the working types of the officers based on the selected features depicting their movement histories. By analysing the geo-tagged police data, we demonstrate how this method can be used to reveal user information, especially interest information based on their POIs and spatial-temporal movement patterns.
Crime is a complex phenomenon, emerging from the interactions of offenders, victims, and their environment, and in particular from the presence or absence of capable guardians. Researchers have historically struggled to understand how police officers create guardianship. This presents a challenge because, in order to understand how to advise the police, researchers must have an understanding of how the current system works. The work presents an agent-based model that simulates the movement of police vehicles, using a record of real calls for service and real levels of police staffing in spatially explicit environments to emulate the demands on the police force. The GPS traces of the simulated officers are compared with real officer movement GPS data in order to assess the quality of the generated movement patterns. The model represents an improvement on existing standards of police simulation, and points the way toward more nuanced understandings of how police officers influence the criminological environment.
Many physical and sociological processes are represented as discrete events in time and space. These spatio-temporal point processes are often sparse, meaning that they cannot be aggregated and treated with conventional regression models. Models based on the point process framework may be employed instead for prediction purposes. Evaluating the predictive performance of these models poses a unique challenge, as the same sparseness prevents the use of popular measures such as the root mean squared error. Statistical likelihood is a valid alternative, but this does not measure absolute performance and is therefore difficult for practitioners and researchers to interpret. Motivated by this limitation, we develop a practical toolkit of evaluation metrics for spatio-temporal point process predictions. The metrics are based around the concept of hotspots, which represent areas of high point density. In addition to measuring predictive accuracy, our evaluation toolkit considers broader aspects of predictive performance, including a characterisation of the spatial and temporal distributions of predicted hotspots and a comparison of the complementarity of different prediction methods. We demonstrate the application of our evaluation metrics using a case study of crime prediction, comparing four varied prediction methods using crime data from two different locations and multiple crime types. The results highlight a previously unseen interplay between predictive accuracy and spatio-temporal dispersion of predicted hotspots. The new evaluation framework may be applied to compare multiple prediction methods in a variety of scenarios, yielding valuable new insight into the predictive performance of point process-based prediction.