CrimeVis: An Interactive Visualization System for Analyzing Crime

Data in the State of Rio de Janeiro

Luiz Jos

e Schirmer Silva

, Sonia Fiol Gonz

ales

, Cassio F. P. Almeida

1,2

, Simone D. J. Barbosa

and H

elio Lopes

Departamento de Inform

atica, Pontif

ıcia Universidade Cat

olica do Rio de Janeiro,

Rua Marqu

es de S

ao Vicente 225, 22451-900, Rio de Janeiro, RJ, Brazil

Escola Nacional de Ci

encias Estat

ısticas - ENCE, IBGE, Rua Andr

e Cavalc

anti 106, 20231-050, Rio de Janeiro, RJ, Brazil

Keywords:

Interactive Visualization, Decision Support, Criminal Data, Socieconomic Data.

Abstract:

This paper presents an interactive graphic visualization system for analyzing criminal data in the State of

Rio de Janeiro, provided by the Public Safety Institute of Rio de Janeiro. The system comprises a set of

integrated tools for visualizing and analyzing statistical data on crimes, which makes it possible to extract and

infer relevant information regarding government policies on public safety and their effects. The tools allow us

to visualize multidimensional data, spatiotemporal data, and multivariate data in an integrated manner using

brushing and linking techniques. The paper also presents a case study to evaluate the set of tools we developed.

1 INTRODUCTION

As its cities and their populations have grown, so has

violence increased in Brazil. In Rio de Janeiro, we

observe that the large social and ﬁnancial inequali-

ties, as well as the regional distribution of the popu-

lation, strongly inﬂuence criminal data. The number

of crimes – especially the violent ones – increases ev-

ery year. Containing that increase and ensuring bet-

ter quality of life have become major concerns of the

government and of public safety institutions (Mon-

teiro and Rocha, 2013). Deﬁning efﬁcient public poli-

cies is a challenge for any government, and devising

strategies for combating criminality directly affects

the majority of the vulnerable population.

When thinking about public policy, we need to

face challenges and seize opportunities in dealing

with ‘Big Data‘ (Power, 2014). There is limited room

for making experiments, so we need to devise ways

to analyze heterogeneous, multi-source data to help

make sense of the current situation and the historical

trends that led to it, so we can better support public

policy decision makers in conceiving, planning, and

continuously monitoring of the programs deployed.

This paper describes an interactive visualization

tool developed to support researchers and public pol-

icy makers in analyzing the criminal data of Rio de

Janeiro. We illustrate its usage to evaluate data pro-

vided by the state’s Public Safety Institute (ISP-RJ,

Instituto de Seguranc¸a P

ublica). In total, we ana-

lyzed statistical data of 138 police districts over 12

years.The data obtained were also analyzed together

with socioeconomic data collected and made avail-

able by IBGE, the Brazilian Institute of Geography

and Statistics (IBGE, 2011).

To allow researchers and public policy makers

to analyze hypotheses and evaluate the government

safety policies, we developed CrimeVis to use geo-

referenced and statistical data provided by both ISP-

RJ and IBGE. CrimeVis provides a set of data explo-

ration tools for discovering patterns and correlations

between the analyzed data sets. The main contribu-

tions of this work are:

1. Visual analysis of n-dimensional criminal data,

making it possible to contrast them with socioeco-

nomic variables and relate them to the distribution

of the population of Rio de Janeiro.

2. A set of tools for the analysis of public safety poli-

cies adopted during a certain period of time, and

the relation between criminal data and the social

and economic distribution of the population.

This paper is organized as follows. Section 2 presents

some theoretical background. Section 3 details the

developed tool suite. In Section 4, we present the re-

sults of a user study conducted with people from a

variety of educational backgrounds. Finally, in Sec-

tion 5 we discuss the importance of those results and

Silva, L., González, S., Almeida, C., Barbosa, S. and Lopes, H.

CrimeVis: An Interactive Visualization System for Analyzing Crime Data in the State of Rio de Janeiro.

DOI: 10.5220/0006258701930200

In Proceedings of the 19th International Conference on Enterprise Information Systems (ICEIS 2017) - Volume 1, pages 193-200

ISBN: 978-989-758-247-9

193

their potential impact on state public safety policies.

2 THEORETICAL BACKGROUND

In Rio de Janeiro, police districts are distributed in

138 DPs throughout the state. The state government

makes those data available to the population through

ISP-RJ (ISP-RJ, 2013). The data we investigated

span a period from 2003 to 2015. The crimes are

counted as follows: for crimes against the individual

(i.e., homicides, lesions, and threats), they consider

the number of victims; for gun seizures, the number

of guns; for other crimes against property (i.e., rob-

beries and thefts), the number of cases, regardless of

the number of victims in each case. As the data are

georeferenced, we can also generate statistics accord-

ing to the population distribution.

Evaluating and quantifying the impact of violence

on the population of Rio de Janeiro present several

challenges. Understanding the problem is difﬁcult

because some data are inconsistent or unrelated, so

the relation between robberies, homicides, and the so-

cioeconimic distribution of the population is unclear.

Violence naturally varies very much both geograph-

ically and over time. Moreover, it matters whether

a person lives either in a conﬂict territory, at its sur-

roundings, or at a ﬁve kilometer radius from its epi-

center (Monteiro and Rocha, 2013). The ofﬁcial crim-

inal data of Rio de Janeiro are aggregated by city re-

gions: they do not allow us to identify either the pre-

cise location of violence epicenters or the population

of each region. Combining criminal data with socioe-

conomic characteristics can help to explore hypothe-

ses about how criminality evolves and how it relates to

social inequalities. However, no government agency

holds all the data necessary for a comprehensive anal-

ysis. To circumvent this problem, criminal data can

be contrasted with socioeconomic data related to the

region associated to each DP. Those data, obtained

by the Brazilian 2010 census (IBGE, 2011), include

information on the population ethnicity and income,

as well as a classiﬁcation of regions considered sub-

normal, i.e., which lack essential public services and

have irregularly constructed buildings. This way, not

only can we analyze criminal data as indicators of

public safety policies, but also visualize the available

socioeconomic data, which can suggest possible pat-

terns and relations among the data.

2.1 Statistical Data Analysis

The knowledge discovery process requires direct dia-

logue with a domain specialist (de Melo et al., 2015),

in order to decide which questions are relevant and

need support in answering. To help discover patterns

with practical meanings for the user. Different clus-

tering algorithms can be used and evaluated. Algo-

rithms such as k-means and k-medoids (Kaufman and

Rousseeuw, 2009) can be used to identify homoge-

neous groups distributed in the observed source, so

we can evaluate n-dimensional data sets to uncover

possible patterns and correlations. Although those al-

gorithms subdivide the data set efﬁciently, they do not

consider spatial data, i.e., the geographic distribution

of the records. In the context of criminal data anal-

ysis, this is quite limiting. Conversely, the SKATER

(Spatial K’luster Analysis by Tree Edge Removal) al-

gorithm (Assunc¸

ao et al., 2006) partitions a data set

according to the spatial distribution of the data. An-

other relevant algorithm for data analysis is Multidi-

mensional Scaling (MDS) (Lee et al., 2014). With a

large number of variables, it is very difﬁcult to dis-

cover relevant characteristics of how similar objects

are to each other, unless the data can be represented

in a small number of dimensions. MDS is a set of

techniques for analyzing objects in a data set by re-

ducing their dimensionality. The similarity measure

is usually related to a distance matrix. This technique

makes it easy to analyze the distance (i.e., dissimi-

larity) between speciﬁc objects and helps to identify

outliers in clusters.

To help make sense of multidimensional data, dif-

ferent visualization techniques are useful, such as the

parallel coordinates chart can be used. It allows vi-

sualizing multidimensional data through parallel axes

in a 2D chart (Palmas et al., 2014)(Zhou et al., 2008),

where each axis represents an attribute, and a polyline

representing an object intersects each axis at its corre-

sponding attribute value. For instance, in the case of

criminal data visualization, each line represents a po-

lice station, intercepting each axis at its corresponding

attribute values. It is thus possible to analyze informa-

tion of n attributes in a single chart.

Parallel coordinate charts have some limitations,

however. Even with an average-sized data set, it may

suffer from overplotting, making it difﬁcult to identify

characteristics, trends or patterns. Moreover, as the

axes do not have a ﬁxed order, ﬁnding a good order re-

quires heuristics and experimentation. One way to al-

leviate this issue is to use the aforementioned cluster-

ing algorithms to visually group the objects, instead of

visualizing each item in the data set separately (Hein-

rich and Weiskopf, 2013)(Johansson et al., 2005). In

this paper we describe some adopted in CrimeVis for

this problem.

Data sets that include temporal data are ubiqui-

tous and notoriously difﬁcult to visualize efﬁciently,

ICEIS 2017 - 19th International Conference on Enterprise Information Systems

194

especially when they have several dimensions besides

time (Bach et al., 2014), as is the case of the crimi-

nal data in our investigation. To solve this problem

we can decompose the problem into each time step

of recorded data, and represent time in an additional

axis, in a 3D parallel chart.

Parallel coordinates chart can be a powerful tool

when coupled with traditional visualization meth-

ods, such as scatterplots and time series. The syn-

chronous use of multiple views can allow interac-

tive exploration across them, through the techniques

of ”brushing and linking” (Heer and Shneiderman,

2012), where the selection of an attribute in one view

is used to highlight different aspects of the same ob-

ject in another view. The coordinated use of various

views can provide a rich strategy for a domain experts

to analyze patterns in a chart and their projection onto

another one.

2.2 Related Work

We conducted a comparative study of recent solutions

both for the criminal domain and for analyzing sta-

tistical data. To use the application efﬁciently, users

should be able to easily answer questions about the

data, discover interesting patterns, and identify errors

in the data (Heer and Shneiderman, 2012). A con-

textualized analysis made by domain experts can as-

sign meaning to trends, clusters, and outliers identi-

ﬁed in the data set under investigation. In the lit-

erature, we ﬁnd several software for the analysis of

statistical data, be they criminal, meteorological, or

electoral funding data, for example. Chainey et al.

(Chainey et al., 2008) used the Hotspot Mapping tech-

nique to analyze spatial characteristics of criminality.

Their system maps criminal data according to where

the crimes occurred using a geographic information

system (GIS), which allows the analyst to identify

patterns and trends for the analyzed areas. Although

the system is efﬁcient with respect to the distribu-

tion of crime occurrences as related to the population

distribution, it does not consider any socioeconomic

characteristics of the area.

Arietta et al. (Arietta et al., 2014) present a

method for automatically identifying and validating

predictive relationships between the visual appear-

ance of a city and its non-visual attributes (e.g. crime

statistics, housing prices, population density etc.).

They combined Support Vector Regression with Data

Mining techniques. Since each city attribute is asso-

ciated with a location (latitude, longitude), they typi-

cally visualize them as thematic maps. In our system,

thematic maps are also used to visualize the distribu-

tion of an attribute over the DPs.

Similar to our approach, Crime in Chicago

(Rougeux et al., 2012) is a data visualization web tool

to explore crime trends in Chicago’s 50 wards, allow-

ing users to compare crime levels over the years and

across city wards. However, it does not consider so-

cioeconomic attributes, nor

the effects of public polices over neighboring regions.

In the next section we describe CrimeVis, the vi-

sualization system we developed to overcome the lim-

itations we identiﬁed in the tools described here.

3 CrimeVis

We developed CrimeVis to help researchers explore

criminal data made available publicly by the state

government at the ISP-RJ web site. The socioeco-

nomic data are also public and made available by

IBGE (IBGE, 2011). From the latter, we considered

as relevant attributes education, ethnicity, and family

income. We integrated both data sets in a database for

CrimeVis to support a wide range of analyses.

CrimeVis was designed as an interactive graphi-

cal system which integrates several widely used sta-

tistical analysis methods and clustering techniques to

visualize criminal data.The tool architecture is ex-

tensible, so it is possible to add new forms of vi-

sualization and data analysis. Different from the

aforementioned tools, CrimeVis combines different

visualization techniques, presenting diverse data syn-

chronously across different views. It allows domain

experts to make a deeper analysis of the datasets,

considering their correlations and inconsistencies be-

tween them, facilitating the identiﬁcation of patterns.

All data are loaded on demand; therefore, our system

is free from CPU and memory-intensive processes.

CrimeVis involved prototyping cycles with evalu-

ation with criminality experts. To satisfy their needs,

CrimeVis was iteratively reﬁned according to their

feedback. Figure 1 shows its initial screen.

With CrimeVis, we can analyze patterns of crimi-

nal distribution over a certain period of time, as well

as their social implications. Experts can also answer

complex questions, such as:

Q1. What locations can be considered concentrations

of certain types of crime?

Q2. How have criminality rates evolved over time?

And what is their relation with social character-

istics of the population, if any?

Q3. How can we subdivide the state areas according

to socioeconomic and criminality criteria?

Q4. What are the practical effects of the policy of

deploying Pacifying Police Units (Unidades de

CrimeVis: An Interactive Visualization System for Analyzing Crime Data in the State of Rio de Janeiro

195

Pol

ıcia Paciﬁcadora - UPPs)?

Q5. Are there any inconsistencies in the data made

available by the government?

Figure 1: Initial screen of CrimeVis.

In the next section we describe the software in detail,

discussing the forms of interaction and how we can

answer each of these questions.

3.1 Overview

CrimeVis initially presents users with two views (Fig-

ure 1): a parallel coordinates chart, in which each line

represents the information of a DP; and a map with

the geographic distribution of the DPs, which can be

used to answer Q1.

The initial screen of CrimeVis gives an overview

of how criminality evolved over a given time period

(in our current data set, from 2003 to 2015). There

we can analyze the evolution of criminality rates in

the state and the corresponding socioeconomic data

(Q2). At the right-hand side panel, users can change

the time period to analyze as they wish.

CrimeVis offers various views on demand, such

as: time series chart, MDS projection, scatterplot, and

3D parallel coordinates chart. All the views are syn-

chronized, i.e., the user selections in one view are re-

ﬂected in all the others. This way, it is easier to iden-

tify correlations in the data and ﬁlter speciﬁc infor-

mation simply by clicking on the attributes of one of

the charts and observing how these data are projected

on the others. For example, an attribute such as homi-

cides or scholarship can be added to or removed from

the visualization and all the views are modiﬁed on-

the-ﬂy. CrimeVis allows users to select data manually

or by region, as well as group by crime rate or socioe-

conomic data. Users can load customized data, using

results generated by R scripts in a preprocessing step.

It is also possible to group the data set in clusters

to present them visually. CrimeVis offers three clus-

tering techniques: k-medoids, SKATER, and a com-

bination of MDS and k-medoids. When visualizing

a set of data that considers homicides, thefts, miss-

ing persons, and population income, we can group the

data according to all the attributes (Q3) together or to

each one separately. By knowing in which group each

DP is located, we can analyze properties speciﬁc to

each group, as well as outliers and anomalies (Q5) in

the input data.

CrimeVis allows us to explore multivariate spatio-

temporal data sets and to quickly investigate hypothe-

ses and patterns. The next section describes the clus-

tering techniques implemented in CrimeVis.

3.2 Data Clustering

The three strategies for data clustering in CrimeVis –

k-medoids, SKATER, and MDS+k-medoids – help us

to answer some of the questions, Q3 in particular.

In K-medoids (Kaufman and Rousseeuw, 2009)

(Hartigan and Wong, 1979), we have used the Eu-

clidean distance as a measure of dissimilarity between

the chosen attributes. The ﬁrst step is to choose the

number of clusters. By estimating an optimum aver-

age silhouette width and using the Calinski-Harabasz

index in the K-medoids algorithm, through some ex-

perimentation we have concluded that, for the data set

in question, an adequate number of clusters is ﬁve.

The SKATER algorithm (Assunc¸

ao et al., 2006)

subdivides the data using both the attributes and the

geographic distribution. SKATER constructs a spa-

tially contiguous graph and then creates a minimum

spanning tree based on the pairwise dissimilarities be-

tween the nodes. That tree is pruned and we create a

subset of new clusters. As result, we have a dataset

containing clusters with regions that are adjacent to

each other and having similar crime rates.

Finally, CrimeVis allows combining the MDS and

the k-medoids algorithms. MDS aims to project the

data in an n-dimensional space so that the distances

between the data points remain approximately the

same (Lee et al., 2014). The algorithm processes a

distance (or dissimilarity) matrix between every data

pair and searches for a projection that minimizes the

cost function. To cluster the data, we modiﬁed the k-

medoids algorithm to consider not only the attributes

of each object, but also the output values of MDS con-

taining an associated weight.

CrimeVis offers the three clustering strategies de-

scribed in this section, combined with charts to help

users to extract relevant patterns.

3.3 Visualization and Inspection

The visualization module provides a set of graphi-

cal tools which include 2D and 3D parallel coordi-

nates charts, scatterplots, time series, and MDS pro-

jections. All views are synchronized to express the se-

ICEIS 2017 - 19th International Conference on Enterprise Information Systems

196

lected subset of objects and attributes of the data set,

in which user actions in a view affect all the others.

The parallel coordinates chart became the central

visualization tool in CrimeVis. It organizes several

data attributes as parallel axes next to one another on

a plane, providing an overview of the relations be-

tween different attributes. Through this visualization,

criminality data can be related to socioeconomic data,

achieving the goal of answering important questions

posed by most researchers in the subject, such as ”Is

there a relation between schooling and crime rates?”

Because traditional parallel coordinates chart may

make it difﬁcult to identify some characteristics due

to the juxtaposition of the polylines, we replaced the

polylines with B

ezier curves. In line with (Heinrich

et al., 2011), the control points of each curve are in-

ﬂuenced by each cluster’s centroid, obtained by one

of the three aforementioned strategies, as well as by

the points in which each curve intercepts each axis.

To build a curve between two adjacent axes x

and

i+1

, we inserted an imaginary central axis between

axes x

and x

i+1

and calculated point C

as the cluster

centroid that would intercept this axis. This point at-

tracts the curves of the corresponding cluster. Auxil-

iary axes are also added to smooth the curve drawing

at a distance d of each axis. Figure 2 compares the

two charts: one with clustered B

ezier curves (on the

left), and a traditional one (on the right).

CrimeVis also provides a 3D parallel coordinates

chart, which is a straightforward extension of the tra-

ditional 2D chart and has the same basic interaction

techniques. Each parallel axis of the original chart is

extended in a third dimension, forming a plane that

represents a 2D scatterplot relating two properties A

and B.For cluster analysis, the 2D chart, even when

using B

ezier curves, may still obscure relevant char-

acteristics of the data, as in some instances we have

investigated. Thus, the 3D chart aims to support the

identiﬁcation of relevant details that would not be eas-

ily noticed in the 2D chart. For each scatterplot, the

values in Z represent an attribute chosen by the user,

and in the Y axis the values are maintained from the

original chart. Figure 3 presents the proposed 3D

chart, in which the geographic distribution of DPs

in the state is chosen for the Z axis. Given the geo-

graphic location of each DP, they were sorted accord-

ing to the following groups: Baixada Fluminense, In-

terior, Grande Niter

oi, and Capital (Q3). The user can

ﬁlter data by clicking on either a speciﬁc set of lines or

a single line. Some degree of transparency is applied

to the lines that were not selected so as to highlight

the data of interest to the user.

Thematic maps can also be used to visualize the

geographic distribution of a certain variable in a par-

ticular period of time, allowing us to answer Q2.

Users can select groups of DP by pressing shift and

clicking on multiple objects or brushing, can hover

over them with the cursor to display their correspond-

ing information in another view, and can zoom in on

a speciﬁed area. CrimeVis can also show a time lapse

animation to visualize how crime evolved over time.

Figure 1 presents a map of homicides in year 2014.

CrimeVis provides three more charts: scatterplot,

time series, and MDS projection. The MDS projec-

tion is nothing more than a scatterplot in which it

is possible to see the data distribution according to

the MDS algorithm and its combination with the k-

medoids algorithm for clustering. Figure 4 presents

the MDS chart. Through this visualization, we can

easily identify outliers, such as a point of cluster 2

(in blue) that seems to be far away from most of the

points belonging to its own group in the Y axis.

CrimeVis offers interactive ﬁlters through which

the user can select a set of objects and observe its

correspondence in another synchronized view. Other

selections can be made through a set of options pre-

sented in the user actions panel, such as the year, area,

and attributes shown in each view. Each ﬁlter opera-

tion modiﬁes the views on-the-ﬂy. For instance, when

selecting a set of lines in a parallel coordinates chart,

the corresponding DPs are also selected in the map of

DPs and in every other active view.

4 EVALUATION AND

DISCUSSION

We conducted a preliminary study to evaluate

CrimeVis and the strategies supported for analyzing

the data. The ﬁrst fully functional prototype was eval-

uated by a group of 24 people, comprising 12 under-

graduate students of different areas, 4 researchers in

the area of criminality, and 8 scientiﬁc visualization

experts. Among the participants, 1 is an expert in

statistics, 3 in the ﬁeld of humanities (1 of whom is

an expert on criminology), and 8 are professional ex-

perts in computer science and scientiﬁc visualization.

No introduction was given of the system for this test.

A set of tasks was given for the users to answer by

interacting freely with CrimeVis, and their answers

were later evaluated. The following sections report

on the evaluation study and its results.

A preliminary study focused on users who do re-

search in public safety. We have conducted an anal-

ysis of the visualization tools, as well as their usabil-

ity. The study collected data through a 7-point Lik-

ert scale questionnaire (1 = completely disagree to 7

= completely agree) for a set of seven tasks related to

CrimeVis: An Interactive Visualization System for Analyzing Crime Data in the State of Rio de Janeiro

197

Figure 2: Clusters in parallel coordinates.

Figure 3: 3D parallel coordinates and their relation with the spatial distribution.

Figure 4: Multidimensional Scaling tool in the CrimeVis.

questions Q1 to Q5. After performing each task, users

were asked to provide their opinion about the ease of

use and usefulness of CrimeVis.

In general, we obtained a positive result in the

questionnaire, especially with respect to ease of use

and interactivity with approximately 70% positive re-

sponses. The main goal of CrimeVis, to support the

analysis of patterns and the understanding of the vi-

sualizations, was the main point investigated. As ex-

pected, the negative feedback was related to the lack

of information on the techniques used and the lack

of a tutorial. Besides, the 3D parallel coordinates

chart was considered in some cases as redundant, be-

cause in most cases the original 2D parallel coordi-

nates chart allowed users to answer the question posed

by the task.

Study participants commented on the efﬁciency

of the data selection techniques and the combination

of interactive visualization with clustering algorithms.

Some of them highlighted this as a strong point of

CrimeVis when compared to other systems. More-

over, according to the study, CrimeVis achieved the

purpose of making it efﬁcient to discover patterns and

correlations in the studied data, which allowed re-

searchers to answer questions such as the ones posed

in Section 3.

The preliminary evaluation of CrimeVis suggests

that the set of tools have achieved its purpose to sup-

port researchers on public safety. Most of the eval-

uated components were deemed easy to understand,

with the exception of a few particular issues. Regard-

ing the interactive controls, the level of understand-

ing was high for the conﬁguration controls (75% of

positive responses) and the selection controls (80%

positive). Five users considered the cluster visualiza-

tion in parallel coordinates plot 2D and 3D difﬁcult

to analyze in certain situations, in which there is jux-

taposition of lines, even when they are replaced by

ezier curves. In addition, one researcher says that

he would need some training in order to easily inter-

pret the chart and to ﬁnd an organization of the axes

that would reveal data attributes more clearly.

Four users reported difﬁculties to select or follow

speciﬁc DPs when analyzing the parallel coordinates

chart with clustering. One user reported that it is dif-

ﬁcult to select and visualize a single DP in the MDS

plot when the plot shows a large number of points.

Another user said that some training is necessary to

use and understand all features of the system. Re-

garding data selection, two users reported that some

legends and information on the data could be clearer.

Despite the reported problems, CrimeVis was well

ICEIS 2017 - 19th International Conference on Enterprise Information Systems

198

accepted by the users, being considered by four spe-

cialists as a strong tool to support research.

4.1 Results

In this section we describe some discoveries in

the data made by public safety researchers using

CrimeVis. These discoveries are related to the Pacify-

ing Police Units (UPPs) program created by the State

of Rio de Janeiro. It is well known that the UPPs had a

large inﬂuence on the criminal activity in recent years.

Since the beginning of this program in 2008, the state

government has created 37 UPPs in the city of Rio

de Janeiro. To better analyze the effect of UPPs, the

researchers divided the set of DPs into groups repre-

senting the sub-regions of the State (Q3), as shown in

Figure 5. Within the State capital (i.e., the city of Rio

de Janeiro, henceforth ‘Rio’), we ﬁnd two groups: the

DPs which received at least one UPP (in red), and the

DPs without UPPs (in green). The other groups are:

Baixada Fluminense (in blue); Grande Niter

oi (in yel-

low); and the Interior of the State (in purple).

Figure 5: Subregions of the State of Rio de Janeiro.

For the analysis, the researchers considered violent

mortality crimes in each group. Violent mortality

comprises crimes of body injury followed by death,

such as homicides, larceny, and police killings. The

study shows that,from 2003 to 2007, crime rates do

not differ much from each other in different regions.

For homicides, they noticed a higher occurrence in

Baixada Fluminense (Q1) in every year. They also

noticed that, from 2008 to 2014, the violent mortal-

ity rates increased, and a major differentiation can be

noted in the behavior of regions, where Baixada Flu-

minense is the more prominent at the end of 2014.

In that region, in 2012, the homicide rates are more

concentrated, and there is a small decrease on aver-

age. Analyzing how crime rates have evolved over

time (Q2), we notice that in 2014 the violent mortality

rates are still concentrated, but with higher values. In

general, violent deaths increased in this region from

2012 to 2014.

The ﬁrst UPP was deployed in December 2008. In

2014, there were 37 UPPs in Rio, covering over 200

communities and an estimated population of 562,691

inhabitants. The expansion of the program raised sev-

eral criticisms questioning its effectiveness in reduc-

ing criminality. In 2015, the Public Safety Institute

(ISP) published a report which showed a decrease in

criminality within the UPPs. However, it did not ex-

amine the neighboring areas of those DPs which re-

ceived UPPs. A program of that nature may inﬂuence

criminality in the surrounding areas as well, and thus

deserves deeper evaluation. The researchers noticed

that in 2008 the regions in Rio with and the neigh-

boring regions without UPPs had similar crime rates.

In 2012, the rates are reduced in two regions, show-

ing that the UPPs may have inﬂuenced the areas of

Rio which had not received a UPP. However, at the

end of 2014, these two regions no longer had simi-

lar behavior. Researchers noticed that, in the regions

without a UPP, there is a larger spread of crime rate

when compared to the regions with a UPP. In the re-

gions with UPPs, homicide rates split into two groups:

those with lower rates correspond mostly to DPs in

the South of Rio, whereas higher rates are found in

DPs in the Northern and Western areas of Rio. The

data show that the program has failed to curb crime

overall. While we have had regions with lower crime

rates, there was a signiﬁcant increase in the Western

parts of Rio and in the Baixada Fluminense from the

year 2012 on. The practical effect of the installation

of UPPs was the spread of crime (Q4).

We found some inconsistencies in the data an-

alyzed (Q5). Analyzing the cluster quality of the

groups for each clustering algorithm used, we con-

cluded that in some cases it is difﬁcult to identify

clusters. To assess the quality of the clustering, we

adopted the silhouette width and the Dunn index (for

which higher values are better; 1 being the best pos-

sible value). Considering 3, 5, 7, and 10 clusters, the

average silhouette width varied from 0.3 to 0.41 for

all techniques and the Dunn index from 0.08 to 0.37.

These values can be related to identiﬁed outliers

in the data, such as DP1. DP1 is located in the central

area of Rio, covering part of the central region and

the island of Paquet

a; it has high rates of homicide

and police murders as shown in Figure 6. This region

has a large ﬂoating population without a signiﬁcant

amount of residents, considering that it is a predomi-

nantly business and commerce area. By analyzing the

data, we can conclude that this region has the high-

est criminal rates given its small number of residents

and the large ﬂoating population, considering that the

rates are calculated by the occurrences of crime per

100 thousand inhabitants. But this is not sufﬁcient

to explain why these rates are so much higher than

in other DPs, considering the gravity of these kinds

of crimes. Further studies are needed to understand

CrimeVis: An Interactive Visualization System for Analyzing Crime Data in the State of Rio de Janeiro

199

criminal behavior in this region.

Figure 6: Crime rates of the DP1 and their location.

5 CONCLUSION

CrimeVis offers a set of tools to support researchers

on public safety. The overview provided by the tool

allows users to easily discover patterns and analyze

trends in the data being investigated. The software

is still undergoing testing to be deployed and widely

used by researchers in the ﬁeld. Our preliminary stud-

ies showed that CrimeVis is efﬁcient when it is nec-

essary to analyze a data set for a speciﬁc time period.

The users could easily establish relations between the

data and identify trends and patterns through interac-

tive analysis of the data. Moreover, the brushing and

linking technique allows us to select and ﬁlter infor-

mations more easily, being a powerful technique to

answer questions relevant to how the relation between

different data attributes. With CrimeVis, we can ana-

lyze not only groups, but also individual areas using

the map of DPs, in which it is possible to interpret the

evolution of a certain attribute over time.

REFERENCES

Arietta, S. M., Efros, A. A., Ramamoorthi, R., and

Agrawala, M. (2014). City forensics: Using visual

elements to predict non-visual city attributes. Visu-

alization and Computer Graphics, IEEE Transactions

on, 20(12):2624–2633.

Assunc¸

ao, R. M., Neves, M. C., C

amara, G., and

da Costa Freitas, C. (2006). Efﬁcient regionalization

techniques for socio-economic geographical units us-

ing minimum spanning trees. International Journal of

Geographical Information Science, 20(7):797–811.

Bach, B., Dragicevic, P., Archambault, D., Hurter, C., and

Carpendale, S. (2014). A review of temporal data vi-

sualizations based on space-time cube operations. In

Eurographics conference on visualization.

Chainey, S., Tompson, L., and Uhlig, S. (2008). The utility

of hotspot mapping for predicting spatial patterns of

crime. Security Journal, 21(1):4–28.

de Melo, B. M., Guimaraes, J. L., de Castro, A. S., San-

tos, C. A., Nascimento, D. M., and Del Pino Lino, A.

(2015). Criminal data mining: A case studyin crim-

inal observatory tapaj

os. In Information Systems and

Technologies (CISTI), 2015 10th Iberian Conference

on, pages 1–6. IEEE.

Hartigan, J. A. and Wong, M. A. (1979). Algorithm as 136:

A k-means clustering algorithm. Applied statistics,

pages 100–108.

Heer, J. and Shneiderman, B. (2012). Interactive dynam-

ics for visual analysis. Proceedings of the 26th ACM

Conference on Hypertext and Social Media, 10(2):30.

Heinrich, J., Luo, Y., Kirkpatrick, A. E., Zhang, H.,

and Weiskopf, D. (2011). Evaluation of a bundling

technique for parallel coordinates. arXiv preprint

arXiv:1109.6073.

Heinrich, J. and Weiskopf, D. (2013). State of the art of par-

allel coordinates. STAR Proceedings of Eurographics,

2013:95–116.

IBGE (2011). Censo demogr

aﬁco 2010: Caracter

ısticas da

populac¸

ao e dos domic

ılios: resultados do universo.

IBGE: Indicadores Sociais.

ISP-RJ (2013). Notas metod

ologicas do instituto de

seguranc¸a p

ublica do rio de janeiro.

Johansson, J., Ljung, P., Jern, M., and Cooper, M. (2005).

Revealing structure within clustered parallel coordi-

nates displays. In Information Visualization, 2005.

INFOVIS 2005. IEEE Symposium on, pages 125–132.

IEEE.

Kaufman, L. and Rousseeuw, P. J. (2009). Finding groups

in data: an introduction to cluster analysis, volume

344. John Wiley & Sons.

Lee, J. H., McDonnell, K. T., Zelenyuk, A., Imre, D., and

Mueller, K. (2014). A structure-based distance met-

ric for high-dimensional space exploration with mul-

tidimensional scaling. Visualization and Computer

Graphics, IEEE Transactions on, 20(3):351–364.

Monteiro, J. and Rocha, R. (2013). Tr

aﬁco de drogas e

desempenho escolar no rio de janeiro. IBRE - Notas

ecnicas.

Palmas, G., Bachynskyi, M., Oulasvirta, A., Seidel, H. P.,

and Weinkauf, T. (2014). An edge-bundling layout

for interactive parallel coordinates. In Visualization

Symposium (PaciﬁcVis), 2014 IEEE Paciﬁc, pages 57–

64. IEEE.

Power, D. J. (2014). Using ‘Big Data’ for analytics

and decision support. Journal of Decision Systems,

23(2):222–228.

Rougeux, N., Pry, C., Eder, D., Baker, Paul, B. J., and Velez,

J. P. (2012). Crime in chicago: An interactive analysis

of crime in chicago’s 50 wards.

Zhou, H., Yuan, X., Qu, H., Cui, W., and Chen, B. (2008).

Visual clustering in parallel coordinates. In Computer

Graphics Forum, volume 27, pages 1047–1054. Wiley

Online Library.

ICEIS 2017 - 19th International Conference on Enterprise Information Systems

200