Applying Ensemble-based Online Learning Techniques on Crime

Forecasting

Anderson José de Souza, André Pinz Borges, Heitor Murilo Gomes, Jean Paul Barddal

and Fabrício Enembreck

Programa de Pós-Graduação em Informática, Pontifícia Universidade Católica do Paraná,

Av. Imaculada Conceição, 1155, Curitiba, Paraná, Brazil

Keywords:

Data Stream Classiﬁcation, Crime Forecasting, Public Safety, Concept Drift.

Abstract:

Traditional prediction algorithms assume that the underlying concept is stationary, i.e., no changes are ex-

pected to happen during the deployment of an algorithm that would render it obsolete. Although, for many

real world scenarios changes in the data distribution, namely concept drifts, are expected to occur due to vari-

ations in the hidden context, e.g., new government regulations, climatic changes, or adversary adaptation. In

this paper, we analyze the problem of predicting the most susceptible types of victims of crimes occurred

in a large city of Brazil. It is expected that criminals change their victims’ types to counter police methods

and vice-versa. Therefore, the challenge is to obtain a model capable of adapting rapidly to the current pre-

ferred criminal victims, such that police resources can be allocated accordingly. In this type of problem the

most appropriate learning models are provided by data stream mining, since the learning algorithms from

this domain assume that concept drifts may occur over time, and are ready to adapt to them. In this paper

we apply ensemble-based data stream methods, since they provide good accuracy and the ability to adapt to

concept drifts. Results show that the application of these ensemble-based algorithms (Leveraging Bagging,

SFNClassiﬁer, ADWIN Bagging and Online Bagging) reach feasible accuracy for this task.

1 INTRODUCTION

Data Stream Mining is an active research area that

aims to extract knowledge from large amounts of con-

tinuously generated data, namely data streams. In ad-

dition to the conventional problems that impact con-

ventional machine learning, Data Stream Mining al-

gorithms must be able to perform both fast and in-

cremental processing of arriving instances, address-

ing time and memory limitations without jeopardizing

predictions’ accuracy. One of the major characteris-

tics of data streams is the assumption that data dis-

tribution changes over time, phenomenon known as

concept drifts (Schlimmer and Granger, 1986; Wid-

mer and Kubate, 1996). In order to deal with con-

cept drifts, ensemble-based methods are widely used

as they are able to achieve high accuracy, while adapt-

ing to both gradual and abrupt concept drifts (Barddal

et al., 2014; Bifet et al., 2009). In these methods, a set

of classiﬁers is trained, usually on different chunks of

instances, and their predictions are combined in order

to obtain a global prediction. Training classiﬁers in

different chunks of instances cause their models to be

different, i.e., a diverse set of classiﬁers is induced.

Intuitively, a diverse set is better than a homogeneous

set, since their combination may achieve higher ac-

curacy than any of the individual classiﬁers. Even

though this claim has not been theoretically proven

(Kuncheva et al., 2003), many works back this ar-

gument through empirical tests (Bifet et al., 2010b;

Gomes and Enembreck, 2013, 2014; Barddal et al.,

2014). Also, for data stream mining ensemble classi-

ﬁers provide another beneﬁt, which is gradual adap-

tations to the model by resetting individual classiﬁers

in response to changes in data, instead of the whole

model, which is often the case when a single classi-

ﬁer is used. This trait causes the ensemble to be less

susceptible to noise (false concept drifts) and to adapt

naturally to gradual concept drifts.

Data Stream Mining has been successfully applied

in many real world problems, such as ATM and credit

card operations fraud detection (Domingos and Hul-

ten, 2000), mass ﬂow (Pechenizkiy et al., 2010), wind

power prediction for smart grids (Bessa et al., 2009),

crime prediction by region (Gama et al., 2014) and

fraud crime detection (Zliobaite, 2010).

de Souza A., Pinz Borges A., Gomes H., Barddal J. and Enembreck F..

Applying Ensemble-based Online Learning Techniques on Crime Forecasting.

DOI: 10.5220/0005335700170024

In Proceedings of the 17th International Conference on Enterprise Information Systems (ICEIS-2015), pages 17-24

ISBN: 978-989-758-096-3

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

In this paper we present a real application of

data stream ensemble-based algorithms for crime

forecasting. We analyze a dataset obtained from

the emergency call center of Brazilian Military Po-

lice (PMSC), namely CRE190, which acts 24/7/365.

CRE190 receives approximately 2,000 calls a day, of

which 300 are converted to police occurrences. Police

occurrences are classiﬁed into 9 subgroups: Commu-

nity Aid, Crimes and Contraventions, Diverse Occur-

rences, Emergencies, Environmental Crimes, Main-

tenance Activities and Transit Occurrences. There-

fore, in order to provide more intelligent support for

PMSC, we study online learning techniques to de-

termine which type of target (private individual en-

tities, commercial stores or residencies) is more sus-

ceptible to crimes and contraventions in a short-term

future. We aim at pointing out the most appropri-

ate ensemble-learning algorithm for crime forecast-

ing to support PMSC to prevent crimes by apply-

ing statistical hypothesis testing to corroborate results

obtained. This paper is divided as follows. Sec-

tion 2 presents the problem deﬁnition and brieﬂy

described the dataset used while Section 3 presents

ensemble-based methods applied in this work. Sec-

tion 4 presents the experimental protocol and dis-

cusses about empirical results obtained. Finally, Sec-

tion 5 concludes this paper and presents future work.

2 THE BRAZILIAN CRIME

FORECASTING PROBLEM

There are few academic works relating the usage of

knowledge extraction from Brazilian Military Police

public safety databases. We believe that the main rea-

son for this fact is the lack of data mining experts

inside these entities. Also, we must emphasize the

difﬁculties on retrieving this kind of data, which ac-

cess’ is usually restricted to inside personnel, there-

fore, not much academic research is allowed. Most

part of the existing studies are limited to Brasil’s Civil

Police from many different parts of the country, which

do not abjure Military Police reality of the same re-

gions. This occurs since there is no integrated system

which merges data obtained from Federal, Civil and

Military polices. It happens that when a victim makes

an emergency call to CRE190 and opens a new oc-

currence, this data is not shared to Civil and Federal

polices, and vice-versa.

In Azevedo et al. (2011), authors emphasize the

lack of data acquisition and storage by Military Po-

lices, which do not allow planning using statistics or

more sophisticated techniques, such as forecasting.

Also, intelligent planning is accounted by authors in

Machado (2009), which enlightened the need of a his-

torical database of criminal actions to perform strate-

gical planning of police actions to prevent crimes.

Some geographic-based approaches are also rel-

evant to determine the most susceptible areas for

crimes in cities (Ferreira, 2012; Nath, 2006). Ap-

proaches such as Formal Concept Analysis (FCA)

have also been applied in Siklóssy and Ayel (1997)

to map and structure speciﬁc criminal organizations.

These organizations are deﬁned by its constituents,

modus operandi, resources and types of crime (e.g.

robbery, drug trade, homicide, art and luxury auto

theft). Also, FCA and rule-based classiﬁcation pre-

sented interesting results on domestic violence in Am-

sterdam, Netherlands (Poelmans et al., 2011).

Time series models were also applied for crime

forecasting in Wang et al. (2013) in order to determine

their order and whether they would be performed by

the same individuals.

Focusing on Brazilian’s Military Police, most part

of its operations are planned accordingly to empirical

knowledge which resides on public safety personnel.

Also, an interesting fact occurs: it is common that

a very low occurrence rate for a certain type of vic-

tim might occur during a period of time, followed by

a sudden increase in these rates and vice-versa. The

cause of these sudden increases and decreases cannot

be determined due to the limited aspect of the data

acquired and stored by the military police. Never-

theless, it is feasible to assume that these crime rates

change accordingly to the hidden context (Tsymbal,

2004), i.e., these changes may be due to attributes not

being monitored. Instead of focusing on ﬁnding why

these changes happens, it is easier to adapt the fore-

casting model to it. Concept drifts might occur due

to a variety of factors such as: period of the month,

holiday eves, payment days and economic inﬂations

which might characterize a higher cash ﬂow in certain

time span. In this sense, we hypothesize that concept

drift learning techniques are needed to forecast crimes

in online fashion, thus, presenting up to date predic-

tions for the most susceptible types of crimes in the

near future.

2.1 The Santa Catarina’s Military

Police Dataset

The dataset extracted from PMSC’s Crime and Con-

traventions Group contains three types of police oc-

currences bulletins regarding robbery and assaults

against private individual entities, residencies and

commercial stores. These three types were chosen

due to its numerous appearances all over the coun-

try; thus, diminishing these crimes are highly related

ICEIS2015-17thInternationalConferenceonEnterpriseInformationSystems

to public safety goals. All extracted data is origi-

nally generated at the emergency call center, namely

CRE190, where the call attendee ﬁlls a form regard-

ing the caller’s emergency, which is later used to no-

tify policemen. The next step occurs when policemen

insert extra data about the emergency, such as people

involved and procedure used. Until now, no analyti-

cal process has ever been made over this data in or-

der to provide PMSC extra information for decision-

making.

The data extracted contains 210 days (instances)

which regards occurrences from January 1

2010

to July 30

2010 and are divided in the follow-

ing attributes: day of the month, holiday eve, holi-

day, dawn, morning, afternoon, night, female amount,

male amount, part of the week and the most suscepti-

ble type of occurrence, i.e. private individual entities,

commercial stores or residencies (class). This dataset

accounts for an average of 2.36 occurrences per day.

Table 1 presents these attributes’ descriptions and do-

mains.

In this paper we focus on the task of predict-

ing the most susceptible type of victim: private in-

dividual entities, commercial stores or residences.

Crimes against private individual entities usually hap-

pen when people are arriving home, specially victim-

izing women while waiting to front gates open. Ad-

ditionally, this type of crime is also very common

against women that visit banks, since they usually do

not react to crimes, therefore being easy targets.

Crimes against commercial stores, in majority,

involve markets, bakeries, pharmacies and clothing

stores. In all of these situations, most part of cashiers

are women, therefore, are also very susceptible to

crimes. Thus, a possibility would be to exchange

cashiers during speciﬁc times of the day to reduce its

own chances to be a target.

Finally, crimes against residences occur during the

morning period or early evening. This type of crime

occur when a residence is left alone or as a contin-

uation of crimes against private individuals that are

arriving home as presented early.

Based on the predictions obtained, we expect to

allow policemen to determine efﬁcient routes accord-

ingly to the most susceptible types of victims, there-

fore patrolling residential, commercial or banking ar-

eas more efﬁciently.

3 ENSEMBLE-BASED

ALGORITHMS

Most of the existing work on ensemble classiﬁers re-

lies on developing algorithms to improve overall clas-

siﬁcation accuracy that copes with concept drift ex-

plicitly (Bifet et al., 2009) or implicitly (Kolter and

Maloof, 2007, 2005; Widmer and Kubate, 1996). An

ensemble classiﬁer can surpass an individual clas-

siﬁer’s accuracy if its component classiﬁers are di-

verse and achieve a classiﬁcation error below 50%

(Kuncheva, 2004). Thus, an ensemble is said diverse

if its members misclassify different instances. An-

other important trait of an ensemble refers to how

it combines individual decisions. If the combination

strategy fails to highlight correct and obfuscate incor-

rect decisions then the method menaced. In addition,

when dealing with data streams it is important to con-

sider concept drifts and hence include some kind of

adaptation strategy into the ensemble. Therefore, a

successful ensemble classiﬁer for data streams must

induce a diverse set of component classiﬁers, includ-

ing a combination strategy that ampliﬁes correct pre-

dictions and adapts to concept drifts.

In the next sections we survey the utilized

ensemble-based algorithms for this comparison.

3.1 Additive Expert Ensemble

The Additive Expert Ensemble (AddExp) (Kolter and

Maloof, 2005) is an extension to the DWM algorithm

(Kolter and Maloof, 2007), thus AddExp associates

each classiﬁer a weight ω which is decreased by an

user-given factor β every time it misclassiﬁes an in-

stance. The ensemble prediction corresponds to the

class value with the greatest weight after combining

classiﬁers’ weights. If the overall prediction is incor-

rect a new classiﬁer is added to the ensemble with

weight τ equal to the total weight of the ensemble

times some user-given constant γ. Since this strategy

can yield a large number of experts, authors consider

oldest ﬁrst and weakest ﬁrst pruning methods to ad-

dress efﬁciency issues.

3.2 Online Bagging

The Online Bagging algorithm is a variation for data

streams of the traditional ensemble classiﬁer Bagging

(Oza and Russell, 2001). Originally, a bagging en-

semble is composed of k classiﬁers, which are trained

with subsets (bootstraps) D

of the whole training set

D. However, sampling usually is not feasible in a data

stream conﬁguration, since that would require storing

all instances before creating subsets. Also, the proba-

bility of each instance to be selected for a given sub-

set is governed by a Binomial distribution and there-

fore can be approximate by evaluating every instance

one at a time and randomly choosing if the instance

will be included in a given subset (Oza and Russell,

ApplyingEnsemble-basedOnlineLearningTechniquesonCrimeForecasting

Table 1: Attributes extracted from the PMSC’s database.

Attribute Type Description

Day of the Month Numeric

Day of the month

represented by the instance

Holiday Eve Binary Determines whether this instance represents a holiday eve

Holiday Binary Determines whether this instance represents a holiday

Dawn Numeric

Amount of occurrences during the

00:01 until 05:59 period

Morning Numeric

Amount of occurrences during the

06:00 until 11:59 period

Afternoon Numeric

Amount of occurrences during the

12:00 until 17:59 period

Night Numeric

Amount of occurrences during the

18:00 until 00:00 period

Female Amount Numeric Amount of female victims

Male Amount Numeric Amount of male victims

Part of the Week

{Beginning,

Middle,

End}

Monday and Tuesday stand for Beginning,

Wednesday and Thursday stand for Middle

and Friday, Saturday and Sunday stand for End

Most Susceptible

Type of Occurrence (class)

{Private Individual Entity,

Commercial Store,

Residence}

Determines who is the most susceptible target for crimes

2001). Nevertheless, this solution is not applicable as

it is necessary to know beforehand the size N of the

training set for an accurate approximation of the orig-

inal sampling process. However, if we assume that

N → ∞, i.e. an inﬁnite stream, then the binomial dis-

tribution tends to a Poisson distribution with λ = 1,

thus it is feasible to approximate the sampling process

(Oza and Russell, 2001).

3.3 Adaptive Windowing Bagging

AdaptiveWindowing (ADWIN) Bagging extends On-

line Bagging, presenting ADWIN as a change detec-

tor (Bifet et al., 2009). ADWIN keeps a variable-

length window of recently seen instances, with the

property that the window has the maximal length sta-

tistically consistent with the hypothesis “there has

been no change in the average output value inside the

window". The ADWIN change detector is both pa-

rameter and assumption-free in sense that is automat-

ically detects and adapts to the current rate of change.

Its only parameter is a conﬁdence threshold δ that in-

dicates the conﬁdence level we want to be in the al-

gorithm’s output since base classiﬁers outputs are in-

herently random. Therefore, ADWIN Bagging is the

Online Bagging method presented in Oza and Russell

(2001) with the addition of the ADWIN algorithm as a

drift detector. When a drift is detected, the worst clas-

siﬁer of the ensemble of classiﬁers is removed and a

new classiﬁer is added trained with instances chosen

by the Online Bagging method.

3.4 Leveraging Bagging

Leveraging Bagging uses the ADWIN algorithm to

detect drifts, but enhances the training method by

proposing two improvements (Bifet et al., 2010b).

First, the bagging performance is leveraged by in-

creasing the resampling and using output detection

codes. The resampling process is done by using the

Online Bagging method using a Poisson distribution.

In this sense, the weights of the resamples are in-

creased by increasing the variable λ. Second, a ran-

domization is added at the output of the ensemble us-

ing output codes. To each class possible to predict, a

binary string of length n is assigned and then an en-

semble of n classiﬁers is created. Each of the clas-

siﬁers learns one bit for each position in this binary

string. When a new instance arrives, it is assigned x to

the class whose binary code is closest. Random out-

put codes are used instead of deterministic codes. In

standard ensemble methods, all classiﬁers try to pre-

dict the same function. Nevertheless, by using ran-

dom output codes, each classiﬁer will predict a differ-

ent function, therefore increasing the diversity of the

ensemble.

3.5 Online Updated Accuracy Ensemble

The Online Accuracy Updated Ensemble (OAUE)

maintains a weighted set of k component classiﬁers,

such that the weighting is given by an adaptation to

incremental learning of the weight function presented

in Brzezinski and Stefanowski (2013). Every w in-

ICEIS2015-17thInternationalConferenceonEnterpriseInformationSystems

stances, the least accurate classiﬁer is replaced by a

candidate classiﬁer, which has been trained only in

the last w instances. Since there is no drift detection

algorithm, OAUE relies on these periodic reconstructs

of the ensemble to adapt to concept drifts.

3.6 Social Adaptive Ensemble 2

SAE2 (Gomes and Enembreck, 2014) is an im-

proved version of the original Social Adaptive En-

semble algorithm (Gomes and Enembreck, 2013).

Both algorithms arrange the ensemble members in

a weighted graph structure, where nodes are classi-

ﬁers and connections between them are weighted ac-

cording to their “similarity”. Similarity is measured

as a function of their recent individual predictions

over the same instances. This graph structure is used

to combine similar classiﬁers into groups (Maximal

Cliques). These groupsare used during vote, such that

individual predictions are combined into each group

decision, which are then combined to obtain the over-

all prediction. The reason to combine predictions this

way is to prevent a large set of similar classiﬁers from

dominating the overall prediction. SAE2 uses a ﬁxed

window w to control how often updates are made to

the ensemble structure, such as classiﬁers removals

and additions accordingly to maximum and minimum

similarity parameters s

max

and s

min

; the maximum en-

semble size k

max

and the minimum accuracy for each

classiﬁer e

min

. Also, similarly to OAUE, SAE2 in-

cludes a background classiﬁer trained during the last

window, which allow the ensemble to quickly recover

from concept drifts.

3.7 Scale-free Network Classiﬁer

The Scale-free Network Classiﬁer (SFNC) (Barddal

et al., 2014) weights classiﬁers predictions based on

an adaptation of the scale-free network construction

model (Albert and Barabási, 2002). In SFNC, clas-

siﬁers are arranged in a graph structure, similarly to

SAE and SAE2, such that classiﬁers with higher ac-

curacy are more likely to receive connections. During

prediction, classiﬁer weighting is based on a central-

ity metric α, such as degree or eigenvector, instead of

solely the individual accuracy. Since, high accuracy

classiﬁers tend to receive more connections, depend-

ing on the centrality measure used, there is an indirect

relation between accuracy and the centrality measure.

A new classiﬁer is added every w instances if the en-

semble overall accuracy is below a given user thresh-

old θ and the ensemble size is below a parameter k

max

and it may receive connections from other classiﬁers

according to their individual accuracy. In addition to

Table 2: Algorithms evaluated and its parameters.

Algorithm Parameter Value

AddExp

β 0.5

γ 0.1

τ 0.05

Pruning Method Weakest First

Online Bagging k 10

ADWIN Bagging

k 10

δ 10%

Leveraging Bagging

k 10

δ 10%

k 6

OUAE

k 10

w {3, 7, 10}

SAE2

max

w {3, 7, 10}

min

66%

min

89%

max

99%

SFNC

max

w {3,7,10}

α Eigenvector

θ 95%

that, the classiﬁer with the lowest accuracyis removed

every w instances, and that triggers a rewiring process

to maintain the graph connected.

4 EMPIRICAL EVALUATION

There is a great monitoring necessity in regions that

try to maintain order and public safety. Nowadays, re-

ports are manually produced by PMSC every 6 hours

in order to provide both feedback for managers and to

inform policemen which are about to start patrolling.

In this section we evaluate a variety of ensemble-

based algorithms in order to determine which present

feasible accuracy to perform real-time crime forecast-

ing. Firstly we present the experimental protocol and

later discuss the results obtained.

4.1 Experimental Protocol

Since the dataset represents daily instances, we eval-

uated all algorithms using 3, 7 and 10-day window

sizes, although a 6 hour period would be essential.

The latter window size of 3 days allows us to analyze

police occurrencesthat happenedduring the weekend,

which are currently treated by the next Monday. The

window size of 7 days allows us to analyze whether

policemen work was effective when compared to an

early week. Finally, a 10-day window size is a special

case which also helps managers to verify the effec-

tiveness of police operations.

In order to compare algorithms, we used all of the

ApplyingEnsemble-basedOnlineLearningTechniquesonCrimeForecasting

algorithms presented in Section 3 with the default val-

ues presented at original papers, with the exception of

internal window sizes w (when existing), which were

set accordingly to the evaluation ones (3, 7 and 10). In

addition, we compare the results for a single Updat-

able Naïve Bayes classiﬁer and an incremental ZeroR

classiﬁer to determine the lowest bound. All algo-

rithms were implemented under the Massive Online

Analysis framework (Bifet et al., 2010a). Also, all en-

sembles use an Updatable Naïve Bayes base learner

(John and Langley, 1995). Table 2 ensemble-based

algorithms’ parameter values used in this evaluation.

To obtain accuracy for algorithms, we adopted the

Prequential test-then-train procedure (Gama and Ro-

drigues, 2009) due the monitoring of the evolution of

performance of models over time although it may be

pessimistic in comparison to the holdout estimative.

Nevertheless, authors in Gama and Rodrigues (2009)

observe that the Prequential error converges to an pe-

riodic holdout estimative (Bifet et al., 2010a) when

estimated over a sliding window.

The Prequential error of a classiﬁer is computed,

at a timestamp i, over a sliding window of size w

′

ac-

cordingly to Equation 1 where L(·, ·) is a loss function

for the obtained class value y

and the expected ˆy

′

(i) =

′

∑

k=i−w

′

L(y

, ˆy

) (1)

Finally, the accuracy of each classiﬁer on a times-

tamp i is computed as 1− P

′

(i).

In this comparison we determined evaluations

over windows of 3, 7 and 10 days (instances) for our

experiments since these values are used to develop

structural police operations as discussed earlier.

4.2 Results Obtained

In Table 3 we present the accuracy results obtained,

where one can see that Leveraging Bagging outper-

forms all other algorithms in all different evaluation

window sizes. This occurs mainly due the lever-

aged resampling process, which is responsible for in-

directly boosting accuracy (Bifet et al., 2009). Ad-

ditionally, results show that both OUAE and SAE2

do not perform well when submitted to the crime-

forecasting problem while algorithms ADWIN Bag-

ging, Leveraging Bagging and SFNC presented the

best overall average results.

Between the best ranked algorithms, one can see

that Leveraging Bagging presents the best results for

all window sizes (3, 7 and 10). In order to deter-

mine whether there is signiﬁcant statistical difference

between Leveraging Bagging and other algorithms,

we used a combination of Friedman and Bonferroni-

Dunn’s non-parametric tests. Firstly, Friedman test

pointed out that there is statistical difference be-

tween algorithms using α = 0.05. Therefore, we

used Bonferroni-Dunn’s 1× N test pivoting Leverag-

ing Bagging and determined that it is signiﬁcantly su-

perior to other algorithms with the exception of AD-

WIN Bagging, Online Bagging and SFNC when com-

pared with a 95% conﬁdence level.

In Figure 1 one can see the accuracy of algorithms

during the whole stream, where Leveraging Bagging

is able to surpass all others in a 3-day window. Also,

one can see that algorithms do not perform well dur-

ing the ﬁrst 50 instances. This occurs mainly due the

lack of instances seen by the classiﬁers and the fact

that the analyzed data refers to a coastal region, there-

fore, from January until March, most part of the pop-

ulation is absent and victims realize that the crime oc-

curred only after this period.

We also emphasize the good adaptation of both

SFNC and Online Bagging algorithms, outperforming

others during the stream. In Figure 1 we present re-

sults for a single Naïve Bayes and an Incremental Ze-

roR classiﬁer, in order to determine the lower bound

for accuracy, therefore, conﬁrming that the Bagging

variations and SFNC yield remarking results.

Although most part of algorithms presents a 60%

accuracy during the stream, we believe these results

are interesting for a generic and naïve approach based

on a limited amount of features and instances.

4.3 Discussion

Generally, the algorithms used in this benchmark pre-

sented a reasonable behavior when submitted to the

crime forecasting problem. All the analytics here pre-

sented was performed aiming at reproducing the Mil-

itary Police procedures for strategical planning. On-

line learning algorithms with the capability of detect-

ing and adapting to concept drifts are essential for

the crime forecasting task since it empowers decision-

making in order to prevent crimes with diminished

resource usage and best manpower allocation. Re-

sults here presented contribute positively for Public

Safety, which so far did not explicitly used online

learning techniques to prevent crimes in speciﬁc city

regions. We must emphasize that online learning al-

gorithms can be used by any Public Safety related

agency that pursuits intelligent automatic support for

decision making.

Another important trait is the lack of profession-

als with adequate knowledge to apply the techniques

here presented in Military Police. In order to take

advantage of the techniques here presented and other

ICEIS2015-17thInternationalConferenceonEnterpriseInformationSystems

Table 3: Accuracy Obtained for the Tested Algorithms.

Average Accuracy (%)

Window Size Naïve Bayes Incremental ZeroR AddExp Online Bagging ADWIN Bagging Leveraging Bagging OUAE SAE2 SFNC

3 50.86 39.81 61.08 64.60 64.60 67.21 57.36 54.82 65.12

7 50.33 39.54 60.49 63.85 61.10 64.29 62.93 54.25 63.85

10 49.77 38.89 60.26 63.52 59.95 63.83 61.28 43.02 63.72

Average Ranking 7.67 9.00 5.67 3.50 3.17 1.00 5.33 7.33 2.33

0 20 40

80 100 120 140

160

180 200

100

Instances Evaluated (Days)

Accuracy (%)

AddExp Online Bagging ADWIN Bagging Leveraging Bagging OUAE

SAE2 SFNC

Naïve Bayes

Incremental ZeroR

Figure 1: Accuracy obtained using a 3-day window evaluation.

data mining algorithms, a great effort must be put on

to break bureaucratic paradigms so that Public Safety

endorses and applies machine learning in day to day

occasions, therefore enabling future research and the

development or optimization of algorithms for this

speciﬁc task.

5 CONCLUSION

In this paper we analyze and compare seven

ensemble-based algorithms for the crime-forecasting

problem. This approach was chosen due the possibil-

ity to perform forecasting incrementally accordingly

to the arrival of occurrences, i.e. CRE190. Speciﬁ-

cally at CRE190, estimates show that 93,000 police

occurrences annually at 24/7/365.

Assuming the great need of Public Safety organi-

zations regarding tools for crime forecasting and pre-

vention, we believe this kind of technique is helpful

and helps these organizations on establishing speciﬁc

plans to act against crime. We also assume that an ini-

tial 60+% accuracy is feasible and tends to improve

on its effective implementation, where police occur-

rences (instances) would arrive intermittently, there-

fore, allowing intelligent actions by crime prevention

teams. So far we were unable to apply our approachin

a real-time application, therefore we were incapable

of measuring the effect size on crime prevention.

One of the major limitations of this work is the

dataset size. Therefore, in future work, we look for-

ward to continue this research by analyzing larger

datasets with an extended set of features, other crime

types and criminal data from other regions.

REFERENCES

Albert, R. and Barabási, A. L. (2002). Statistical mechanics

of complex networks. In Reviews of Modern Physics,

pages 139–148. The American Physical Society.

Azevedo, A. L. V. d., Riccio, V., and Ruediger, M. A.

(2011). A utilização das estatísticas criminais no

planejamento da ação policial: Cultura e contexto or-

ganizacional como elementos centrais. 40:9 – 21.

Barddal, J. P., Gomes, H. M., and Enembreck, F. (2014).

Sfnclassiﬁer: A scale-free social network method to

handle concept drift. In Proceedings of the 29th An-

nual ACM Symposium on Applied Computing (SAC),

SAC 2014. ACM.

Bessa, R., Miranda, V., and Gama, J. (2009). Entropy and

correntropy against minimum square error in ofﬂine

ApplyingEnsemble-basedOnlineLearningTechniquesonCrimeForecasting

and online three-day ahead wind power forecasting.

Power Systems, IEEE Transactions on, 24(4):1657–

1666.

Bifet, A., Holmes, G., Kirkby, R., and Pfahringer, B.

(2010a). Moa: Massive online analysis. The Journal

of Machine Learning Research, 11:1601–1604.

Bifet, A., Holmes, G., and Pfahringer, B. (2010b). Leverag-

ing bagging for evolving data streams. In Machine

Learning and Knowledge Discovery in Databases,

pages 135–150. ECML PKDD.

Bifet, A., Holmes, G., Pfahringer, B., Kirkby, R., and

Gavaldà, R. (2009). New ensemble methods for evolv-

ing data streams. In Proc. of the 15th ACM SIGKDD

international conference on Knowledge discovery and

data mining, pages 139–148. ACM SIGKDD.

Brzezinski, D. and Stefanowski, J. (2013). Reacting to dif-

ferent types of concept drift: The accuracy updated

ensemble algorithm. IEEE Transactions on Neural

Networks and Learning Systems, 25(1):81–94.

Domingos, P. and Hulten, G. (2000). Mining high-speed

data streams. In Proc. of the sixth ACM SIGKDD in-

ternational conference on Knowledge discovery and

data mining, pages 71–80. ACM SIGKDD.

Ferreira, J., J. P. M. J. (2012). Gis for crime analysis: Ge-

ography for predictive models. 15:36–49.

Gama, J. and Rodrigues, P. (2009). Issues in evaluation

of stream learning algorithms. In Proc. of the 15th

ACM SIGKDD international conference on Knowl-

edge discovery and data mining, pages 329–338.

ACM SIGKDD.

Gama, J. a., Žliobait

e, I., Bifet, A., Pechenizkiy, M., and

Bouchachia, A. (2014). A survey on concept drift

adaptation. ACM Comput. Surv., 46(4):44:1–44:37.

Gomes, H. M. and Enembreck, F. (2013). Sae: Social adap-

tive ensemble classiﬁer for data streams. In Compu-

tational Intelligence and Data Mining (CIDM), 2013

IEEE Symposium on, pages 199–206.

Gomes, H. M. and Enembreck, F. (2014). Sae2: Ad-

vances on the social adaptive ensemble classiﬁer for

data streams. In Proceedings of the 29th Annual ACM

Symposium on Applied Computing (SAC), SAC 2014.

ACM.

John, G. H. and Langley, P. (1995). Estimating continuous

distributions in bayesian classiﬁers. In Eleventh Con-

ference on Uncertainty in Artiﬁcial Intelligence, pages

338–345, San Mateo. Morgan Kaufmann.

Kolter, J. Z. and Maloof, M. A. (2005). Using additive ex-

pert ensembles to cope with concept drift. In ICML

’05 Proc. of the 22nd international conference on Ma-

chine learning, pages 449–456. ACM.

Kolter, J. Z. and Maloof, M. A. (2007). Dynamic weighted

majority: An ensemble method for drifting concepts.

In The Journal of Machine Learning Research, pages

123–130. JMLR.

Kuncheva, L. I. (2004). Combining Pattern Classiﬁers:

Methods and Algorithms. John Wiley and Sons, New

Jersey.

Kuncheva, L. I., Whitaker, C. J., Shipp, C. A., and Duin,

R. P. (2003). Limits on the majority vote accuracy

in classiﬁer fusion. Pattern Analysis & Applications,

6(1):22–31.

Machado, D. M. S. (2009). O uso da informação na gestão

inteligente da segurança pública. In A força policial:

órgão de informação e doutrina da instituição policial

militar, pages 77–85.

Nath, S. V. (2006). Crime pattern detection using data min-

ing. In Web Intelligence and Intelligent Agent Tech-

nology Workshops, 2006. WI-IAT 2006 Workshops.

2006 IEEE/WIC/ACM International Conference on,

pages 41–44.

Oza, N. C. and Russell, S. (2001). Online bagging and

boosting. In Artiﬁcial Intelligence and Statistics,

pages 105–112. Society for Artiﬁcial Intelligence and

Statistics.

Pechenizkiy, M., Bakker, J., Žliobait

e, I., Ivannikov, A., and

Kärkkäinen, T. (2010). Online mass ﬂow prediction in

cfb boilers with explicit detection of sudden concept

drift. SIGKDD Explor. Newsl., 11(2):109–116.

Poelmans, J., Elzinga, P., Viaene, S., and Dedene, G.

(2011). Formally analysing the concepts of do-

mestic violence. Expert Systems with Applications,

38(4):3116 – 3130.

Schlimmer, J. C. and Granger, R. H. (1986). Incremen-

tal learning from noisy data. Machine Learning,

1(3):317–354.

Siklóssy, L. and Ayel, M. (1997). Datum discovery. In Ad-

vances in Intelligent Data Analysis, Reasoning about

Data, Second International Symposium, IDA-97, Lon-

don, UK, August 4-6, 1997, Proceedings, pages 459–

463.

Tsymbal, A. (2004). The problem of concept drift: deﬁ-

nitions and related work. Technical Report TCD-CS-

2004-15, The University of Dublin, Trinity College,

Department of Computer Science, Dublin, Ireland.

Wang, T., Rudin, C., Wagner, D., and Sevieri, R. (2013).

Learning to detect patterns of crime. In Machine

Learning and Knowledge Discovery in Databases,

volume 8190 of Lecture Notes in Computer Science,

pages 515–530. Springer Berlin Heidelberg.

Widmer, G. and Kubate, M. (1996). Learning in the pres-

ence of concept drift and hidden contexts. Machine

Learning, 23(1):69–101.

Zliobaite, I. (2010). Learning under concept drift: an

overview. CoRR, abs/1010.4784.

ICEIS2015-17thInternationalConferenceonEnterpriseInformationSystems