Behavioral Analysis for Child Protection in Social Network through

Data Mining and Multiagent Systems

Mário Sérgio Rodrigues Falcão Jr., Enyo José Tavares Gonçalves, Tciciana Linhares Coelho da Silva

and

Marcos de Oliveira

Universidade Federal do Ceará (UFC), Quixadá, Brazil

Keywords: Child Sex Grooming, Multiagent Systems, Data Mining, Cognitive Analysis.

Abstract: The Internet connects millions of people worldwide, enabling diverse ways of interaction and social

organization. Online Social Networks such as Facebook, MySpace, and Twitter, have created a new form of

socialization that can provide good experiences for users. However, such systems, as well as connecting

people, expose their users lives to others, making them subject of exploitation in many ways. This work

explores specifically children and teenagers degree of exposition on Facebook. Due to the risk offered in

distinct layers of the Internet, the aim of this work is to develop a smart tool that helps to avoid the action of

individuals that are possibly a risky for children and teenagers, users of the social network Facebook,

applying Data Mining techniques in a Multiagent System.

1 INTRODUCTION

The Internet is a global network that connects

millions of computers around the world, serving as a

major factor of communication and social

integration (Belloni, 2001). A group of applications

for internet are built based on the ideological and

technological foundations of Web 2.0, which allow

the creation and exchange of User Generated

Content (UGC) (Kietzmann et al., 2011), i.e. blogs,

social networking pages, chats, professional

networks (LinkedIn), community networks (social

networks in neighbourhoods or cities), political

networks and especially electronic social networks

such as Facebook, Twitter, Google+, MySpace, etc.

They compose the large set of social media

(Lemieux et al., 2008).

It is relevant that in our modern society people

are closer to technology than before, especially

children. A child from the 80's was more used to

handle toys, and nowadays children have better

skills on handling information technology, due to the

routine contact with it (Bombonatto, 2012).

Online Social Networks are virtual organizations

composed by people, and give the opportunity for

them to interact with distinct types of individuals

(Mazman and Usluel, 2009). That allows for the

possibility of development of certain relationships

that match certain characteristics indicating that the

children and teenagers are being sexually groomed.

Children with emotional distress, emotionally

dysfunctional, and with low self-esteem are more

likely to be tricked and open to feelings of

defencelessness, a factor that some child sex

offenders use as advantage in its investees. Dialogs

involving subjects such as family problems, secrets

and participation in controversial issues, are some of

the varied resources adopted by these offenders to

gradually seduce their victims. After that follows

events such as sending photos with sexual

connotations, improper conversations on

pornography are dominant in other dialogues. Thus,

the responsible by the child can be assured that a

seduction process was started, and protection

measures should be taken.

There is a noted lack of inspection by most of the

parents related to the use of the Internet by their

children. Many children are exposed to a world

before unknown, and that put together different

types of personalities (Pereira, 2009). A survey,

conducted by Minor Monitor Company, says that

about 38% of users in the social network Facebook

do not have the allowed age for their use, and 30%

of parents allow unsupervised use (Silvio, 2012).

One possible reason for this lack of inspection may

be associated with the shortage of computational

tools to help them.

306

Jr., M., Gonçalves, E., Silva, T. and Oliveira, M.

Behavioral Analysis for Child Protection in Social Network through Data Mining and Multiagent Systems.

In Proceedings of the 18th International Conference on Enterprise Information Systems (ICEIS 2016) - Volume 2, pages 306-313

ISBN: 978-989-758-187-8

To reduce this deficit of inspection, it is

interesting to have available automated tools that can

identify irregular behaviour patterns or potential

suspects in the social network. Therefore, the group

of users, composed by children and teenagers using

these networks, would be less vulnerable.

Given this gap, it is necessary to propose

mechanisms that can be easily used by parents, to

ensure the dignity of children and teenagers that use

social networks, as they are still in psychological

growth, and are subject to individuals with bad

intentions on the Internet (Buckingham, 2000).

This article proposes an alternative to parents

before the scenario presented. We implemented a

Multiagent System that applies Data Mining to

automatically classify the exposition of children in

the Facebook. That is made based on the interactions

developed in it. This paper is organized as follows:

Section 2 presents the theoretical foundation, the

Section 3 shows some related works, Section 4

explores the architecture of the systems developed,

observing the Data Mining modules that were

adopted in this work, section 5 describes the results,

and Section 6 presents the conclusion and lists

possible future work.

2 RELATED WORK

2.1 Automatic Text Analysis in Instant

Messages to Detect Child Sexual

Grooming

In a study by (Santin, Freitas and Paraiso, 2011), a

software service was developed to classify

conversations stages in chat rooms, through a set of

pre-selected words and rules. The algorithm SVM

(support vector machine) was used to classify the

stages of interaction between entities, children and

possible suspects. They used the database found in

the website “www.perverted-justice.com”, which

has real talks between pedophiles and children. This

database was not used directly in this work, but

served as examples of how sex offenders interact

with their victims and persuade them.

Inconveniently, the words of the chat should be

exactly the same to those found in the set of words

in the database, which greatly limits the accuracy to

identify the stages, since words can vary from region

to region or from person to person.

The main difference between this work and

(Santin, Freitas and Paraiso, 2011) is the analysis of

events that happen in the context of the social

network Facebook, and do not depend as much from

accuracy as analysing a given string, not just

focusing on a literal analysis as the automatic text

analysis that they do.

2.2 Fake Profile Identification in

Online Social Networks

The Social Privacy Protector Software tool for

Facebook (SPP), aims to identify "fake" profiles on

the social network Facebook, and improve the

privacy and security settings of users (Fire, Kagan,

Elyashar and Elovici, 2014).

The SPP has three layers of protection that

enhance user privacy through the implementation of

different methods. The system first displays a

possible profile that can pose as threat, and then

immediately provides the means to restrict personal

information to the suspect profile. Then the second

layer allows the owner of the profile on the social

network to adjust your privacy settings according to

your personality type.

The third layer of the system alerts the user about

the number of third-party applications installed in

his profile that have access to his private data

Similarly to the work presented in this paper, the

social network used by the SPP is the Facebook, and

it makes use of data mining techniques for profile

classification. However, it should be noted that the

focus of the SPP is to identify “fake” profiles, that

is, profiles that are note true in relation to the

identity of the their owners regardless of their goals.

Similarly, the approach described here seeks to

identify people who threaten children. This children

group refers to people under 12 years of age

incomplete, according to Brazilian law.

2.3 A Learning Model for Intelligent

Agents Applied to Poultry Farming

The work in question relates to a system responsible

for analysing data on poultry farming and to provide

results that meet the basic needs for birds like

amount of food being consumed, temperature,

amount of water and energy to be consumed among

other factors (Ribeiro et al, 2015).

The system run practical examples, and achieved

good results, consistent with reality, and equivalent

to data that poultry professionals would consider

correct and appropriate for a given situation. The

experiments were conducted in two parts: the first in

which the agents worked according to the average

weight of chickens and the second according with

the weekly food. The agents with the best accuracy

Behavioral Analysis for Child Protection in Social Network through Data Mining and Multiagent Systems

307

classified the chickens that obtained the best weight

and disease resistance them took the indicated

supplements and ate correctly taking into

consideration temperature, humidity, time, light, etc.

Similarly, the work presented in this paper also

used MAS and Data Mining, the logic of both works

is somewhat similar in the sense that agents form

part of the system responsible for making decisions,

with predefined heuristics, and through data mining,

data classification becomes possible.

2.4 A Problem-solving Agent to Test

Rational Agents a Case Study with

Reactive Agents

In this work described by (Silveira, Campos and

Cortés, 2014), a set of tests was conducted with a

rational agent, owner of knowledge and judgment

skills in certain subjects. Several tests were

conducted with the objective of obtaining the results

and thereby try to improve agent performance on

their weaknesses.

The agent was composed of behaviours

previously defined by the project designer and

sensors for detecting aspects in accordance with the

environment. The tests are designed to identify

weaknesses that the agent would probably have

when certain events happen.

The results were satisfactory for research, but not

for the agent tested, some vulnerabilities have been

found and immediately reported to the project

designer.

Similar to this work some tests were performed

with our agents, the black-box type, different input

types and amount of data to see the reaction of the

agents. During the process, some inconsistencies

were found, such as misclassification of information,

and then treated so that the agents could work

cohesively and efficiently.

3 MULTI-AGENT SYSTEM AND

DATA MINING FOR

CHILDREN PROTECTION ON

FACEBOOK

The system consists of Autonomous Agents

developed in the Jade framework (Java Agent

Development Framework) (Bellifemine et al., 2007),

and a data-mining module that uses the WEKA’s

classification algorithms (Hall et al, 2009) over the

data collected from Facebook using the Facebook4j

framework (Yamashita, 2009).

3.1 System Architecture

Multi-agent systems were used to make decisions

with precision, working together on data from

Facebook. The system has more than one agent in

order to modularize the functions of analyzers and

distribute responsibilities, so that each agent has a

specific function, for example, the agent analyzer of

posts will check the times, the kind of privacy, the

content present in the text and among other

heuristics defined in its scope. The agents have a

structure with predefined heuristics by professionals

of psychology aimed at detecting irregularities on

data found on Facebook of children. Formulas that

calculate the level of privacy in accordance with

what the child posts, likes or shares are defined,

moreover, these agents have access to dictionaries

full of indecent language to compare with the words

that child uses.

The autonomous agents interact among them

collecting data from the Facebook server via the

Facebook4j framework (Yamashita, 2009). Files in

JSON format are requested and read by autonomous

agents. These filter the information and work

together to create an object that is capable of being

classified by the mining module, such an object

generated has references to relevant information for

the data mining module, then the latter is in charge

of receiving this object by means of a method call,

and use of the classification algorithm to classify

this instance according with the model defined in the

system. After that, the vulnerability level at which

the child is found is returned to the central module

of the server, and this elaborates a report to the end

user by explaining the reasons why the child is or

not classified as at risk.

After all processing of information the central

server prepares the data considered suspect and

returns it in an HTML page.

The information provided in the final report are:

friends considered suspicious, posts made by the

child, or for someone where she was marked, with

descriptions or comments considered not advisable

for his/her age, amount of pictures in high exposure,

books , videos, TV channels, groups and music not

recommended, registered family in the social

network or not, amount of pokes received after

midnight, shared albums with third parties and that

have a subject not suitable for children and a privacy

note stating a degree for the child overall exposure

within the social network, i.e., indicating how

exposed is his/hers data to friends, friends of friends,

ICEIS 2016 - 18th International Conference on Enterprise Information Systems

308

strangers and games or applications that obtain data

in a way not so perceptive to human eyes.

The architecture used did not cause problems

unless on the synchronization from the central server

with the MAS. In that case, implementation of

thread control was necessary so that the server can

wait for all the autonomous agents response and then

immediately proceed to code execution.

3.2 Facebook Social Network Data

Collection

So, to care on with the investigation it is necessary

the existence of content. Given that the social

network Facebook has a large volume of data and

different types of user profiles, some attributes play

an important role in the interpretation of sentiment,

habits, tastes, etc. As a result some features are

relevant for analysis and decision of the potential

risks posed to the child's account. These attributes

can include: news feed, interests, music, videos,

pages, liked or shared links, date of birth,

inspirations, type of education, conversations, posts,

improper dates and times for registered activities in

social networking, games, events, TV channels,

family, shared albums and photos in excess, bad

comments on photos and posts, private messages

from friends or others containing sexual connotation,

privacy notice that tells you how your personal

information is exposed and friends who might pose a

threat. All of this information can be obtained

through the Facebook Graph API, the primary way

for apps to read and write to the Facebook social

graph. The target audience are children under 12

years. All available information for Facebook will

be obtained in a matter of a minute and following

will be analyzed by the System.

3.3 Technologies Used

The developed system consists of autonomous

agents developed in the JADE framework, and uses

the standard protocols from FIPA (Foundation for

Intelligent Physical Agents) for communication

among agents.

The Weka data mining tool was chosen because

it is known as one of the "top 10 free" tools of the

current business intelligence market, according to

the website "predicts analytics today" (Nyce and

CPCU, 2007). The classification algorithm used in

this work was decision tree (Quinlan, 1986)

implemented in Weka as j48. The basic idea of the

decision tree algorithm is recursively choose the best

attribute to divide the nodes of the tree. After

selecting an attribute, the data is divided into

multiple partitions according to the value of the

chosen attribute. For each partition recursively

computed the best attribute is chosen to split the data

in the current node of the tree. The decision rules are

stored and new rules are generated.

3.4 Data Used and Validated

Classification Model

The data to generate the classification model are

obtained from Facebook and organized into tuples

that have the following columns: posts, (if presents

only private albums), private albums, preferences on

books, music, videos, if updates the news feed,

family registered in the account or not, games,

videos with no children connotation in your

comments or descriptions, pokes after midnight and

conversations with adult content, and the last

column responsible for classifying the tuple as "risk"

or "no risk”. The data that will fill the columns are

analyzed by intelligent agents, which assign

predefined values according to the discovered

content, which will be used for the classification

model. The values chosen by the agents will be

explained further at the end of this section. The

Decision Tree model is shown in Figure 1, so new

instances can be classified as "no"(is not at risk), or

"yes"(is at risk) warning the child to a possible

danger based on the paths of the Tree.

Initially, the pattern obtained had a high rate of

accuracy and was named template best case

scenario, i.e., had weak presence of events

uncommon as false positives and false negatives. At

first it had a high correction rate to classify instances

not to be correct but for being an addict model, that

is, only one column with a risk offerer value the

generated model already rate this instance as "risk"

and do not bother to go further in the other columns

aimed at generating a reasoned result in the

maximum possible columns, as you can see at the

Figure 1.

Figure 1: Vicious Model generated by the J48 algorithm.

However after a test sequence file composed of

data responsible for generating the model yielded a

Behavioral Analysis for Child Protection in Social Network through Data Mining and Multiagent Systems

309

pattern suitable to receive different types of bodies,

i.e., no longer grounded in only one column but in

much as possible according to the child data be

received, as you can see at the Figure 2.

Figure 2: Model generated by the J48 algorithm.

A random algorithm created five thousand lines,

with certain defined heuristics, to generate the

classification model. The training of the model was

as follows: instances are generated according to the

values assigned to the tree fields in Figure 2, and

according to these values at the end, it is classified

as "no"(is not at risk) or "yes"(is at risk).

In order to refine the model and increase its

accuracy some tests were run to have a satisfactory

model. Among the tests used and provided by Weka,

we can mention the Use Training Set, Supplied Test

Set, Cross-Validation and the Percentage Split. The

tests run served to refine the model and make it

more able to categorize each instance of less

incorrectly possible. As a result the inevitability of

adding outliers, identification data that should follow

an expected pattern but do not (Campos, 2014),

became important to improve the model to less

trivial cases.

The database used to form the classification

model has 5270.00 lines (30% were used for model

testing and 70% for training) composed of false

positives, false negatives, true positives and true

negatives. The tests conducted with J48 algorithm

presented precision of 94.53% with 4,982 instances

as correct and 288 classified as incorrect. The SMO

algorithm achieved an accuracy of 90.13% with

4750 instances as correct and 520 classified as

incorrect. The IBK algorithm had an accuracy of

94.03% with 4,954 instances as correct and 316

classified as incorrect. Finally, the NaiveBayes

algorithm had an accuracy of 82.86% with 4,367

instances as correct and 903 classified as incorrect.

Even the result with lowest accuracy (generated

by Naive Bayes) presents a quite satisfactory

approach. We believe it is possible to use any of the

tested approaches, however we kept in this

experiments the one with greater accuracy, i.e.

decision tree (J48).

The tree fields are filled with values defined by

autonomous agents, as mentioned previously,

according to their analysis; the field posts informs

the safety of posts made by the child and the ones

where he/she was marked. Any value above 3.0 is

already considered as unsafe, the value is obtained

by using the following formula F:

F = ((quantity of dangerous dates) * 3+ (privacy

note * 1) + (quantity of dangerous messages * 2) +

(amount of suspicious friends * 2) + (number of

suspicious descriptions and comments in the post *

2)) / 10.

This formula came after a series of analyzes and

tests with different Facebook accounts. For this

reason, a weighted average was made in accordance

with the information provided by the Facebook. In

all cases in which this value was greater than 3.0

actually exist irregularities in the data.

The games field is a numeric value that tells you

how much inadvisable games or apps are often used

by children, for example, Tinder, Cupid, Interesting,

Skout, Let's date, Meetmoi and Catrachos. Any

amount in excess of five is already considered as

unsafe. The videos field is a numeric value that tells

you how much the child saw not advisable videos.

The videos deemed not advisable are identified by

name or description, if any keyword contains

pejorative terms or sexual context video is classified

as dangerous.

The field conversations are a Boolean that

receives "true" or "false" indicating whether the

private conversations of the child have any

disagreeable talk or inappropriate content. These

talks are extracted from Facebook and analysed by

agents that are referred to in a database with pre-

defined words that might pose a risk to the child.

4 VALIDATION

Two children's profiles were selected on Facebook

and, with the permission of their parents, their data

were used for analysis. The tool carried out the

classification in the two instances of the case study

fully automatically.

The two children have 12 years old and are

female, the first child was born in Floriano City - PI

and the second in Quixadá City - CE. Between these

ICEIS 2016 - 18th International Conference on Enterprise Information Systems

310

Table 1: Model columns for classification.

Album Preferences

Feed

News

Pokes Family Games Videos Conversations

True|False Green|Yellow|Red

True|

False

Numeric

True|

False

Numeric Numeric True|False

Table 2: Results.

Child

Times

at dawn

Not

advisab

le TV

show

Not

advisa-

ble

Singers

Family

Presence

Inbox

Conversa

-tions

Pokes

after 12

Privac

y Rate

Posts

Child

from

Floriano

Logged

224 times

after 12

Malha-

ção

6 different

singers.

No family

member is

registered

in your

account.

Fuck,

motherfuck

e, dirty and

the

expression:

I want to

see you

naked told

by a friend.

22 12,85

"And what if

it is to be

alone and

happy

without him,

my life

continued; D

#Happy

#Funny #Girl

#Top (08 /

April / 2014

12:31:57).”

Child

from

Quixadá

Nothing Nothing

3 different

singers.

No family

member is

registered

in your

account.

Nothing Nothing 1,86 Nothing

two selected children, the child from Floriano did

not know she was being monitored; on the other

hand, the child from Quixadá knew about the

experiment and was present during the work's

validation next to his mother.

The Parents of the children in question provided

the authorization for the analysis already knowing

that there would not be disclosure of personal

information from children, as the name, Facebook

profile or any data that could expose their personal

information would not be revealed. However, the

results are listed observing the content found in

children profiles, as shown in Table 2.

4.1 Considerations about Child from

Floriano City

All the words in the column "Private Conversations"

were recorded in mostly during the night, between

07:00 pm until 02:00 am the next day. Of this same

child in 2014 Carnival period some friends asked if

itself would have gone to two bands carnival

celebrations. In addition a friend in particular called

for a party at his home at 08:24 pm on 23 August

2013.

If you notice that the posts column there are

certain reserved words that pose no danger, but

when together the context of the sentence has

subtended goal, in this case, it comes to an end a

dating. At first few sentences were analysed

manually to form the classification model.

It is noted that although Facebook account

analysed in question had been inactive from the

beginning of 2014 to June 2015, still some data

could be retrieved and identified as unsafe for a

child 12 years. The data-mining model classified the

child as subject to risks.

4.2 Considerations about Child from

Quixadá City

There were no suspicious events after midnight

(00:00 to 06:00) in the second child account, the

mother just made sure to point out that she does not

Behavioral Analysis for Child Protection in Social Network through Data Mining and Multiagent Systems

311

allow the use of social networking during late

periods of the night.

No family member is registered in her account,

which may represent a kind of lack of protection

thus unknown people may feel more free to

approach the child.

However, the privacy score for the child in

question is low, 1.86, meaning that the child's data is

"protected" and less accessible to applications or

people.

In relation to music, three singers were identified

as with song lyrics not advisable for children, which

led the tool to classify them as irregular to the child.

The following sentences were posted in the three

singer's pages:

 First Singer: Come sing with me my people

Ouro Branco-MG ... today has a lot of music

and joy in the heating Hot Party Country! I

want to see everyone there huh! From 22h.

#ourobranco #escarpasfolia.

 Second Singer: Left Ponta Grossa-PR! Today

the party is right there! Let's go! #goWithGod #

Tour2014.

 Second Singer: And the party starts… I was

homesick! Today the party will be in

Porangatu-GO city.

#goWithGod#carnival2014#hotParty#causingE

ffects#HotSituation#caseUndefined.

 Third Singer: It is coming! With full house,

let's go with Thaeme and Thiago Victor Hugo

and all rich guys!! # # ticketsAreOver

#hotParty.

 Third Singer: And the party continues, today

is the day to sing and cheer a lot with the guys

in Atlanta! Who goes there? # TourUSA2015.

The data-mining model classified the child as not

subject to risks.

After the results, those responsible for the child

considered no problem in relation with her musical

taste. This is one among the advantages that the tool

offers, identify the suspicious events and providing

to the responsible for the child the opportunity to

consider stopping the access to information

identified in the mining process, such as music,

book, video, and other possible improper things

through the Facebook.

The results were shown to the children's parents

and served to improve the approach to people.

Through interaction between parents, children and

our system, we can see that curiosity by both parties

was notorious, especially the children's parents about

specific words, pictures posted and friends of their

children. So we conclude that the approach should

be taken with great care not to hurt third of human

rights or harm the image of someone because of a

spoken word, picture posted or specific action taken

within the social network.

5 CONCLUSIONS AND FUTURE

WORKS

This work proposes an intelligent tool capable of

analyse data from a child’s Facebook page, and

inform his/hers possible level of vulnerability.

Child sex grooming besides being present on the

Internet also takes up space within social networks,

this is because of large amount of information and

resources offered by these social interaction

services, and parents often fail to monitor or track

the risks present in this virtual world.

The software developed to combat these risks

uses the Jade framework and Weka data-mining

tool. In addition to a previous version, three new

agents were added (private conversations scanner,

external conversations checker and news feed tester)

to provide a broader classification model to accept

different data from the child's account.

The agents are the brain of the system and make

most of the work, by modelling the child’s instance

profile data so that he/she can be classified as

vulnerable or not. All agents follow a

communication protocol and the heuristics

previously defined in order to find irregularities in

the children's data.

The algorithm used for classification and applied

in the tests, the J48, was elected to represent the data

in a binary tree, thus the visualization and

interpretation of how the model works became more

noticeable to the project's authors.

Initially, the model used for the instances

classification caused differences in the results due to

conflicts of information, i.e., instances with similar

data were classified as different classes. After a

filtering sequence and correct choice of the types of

data to be used in the model, the algorithm

classification began to return consistent results.

A plausible option for future work would be the

test in new children's profiles, with preference for

children who have frequent activity on Facebook or

contain relevant amount of data for analysis.

Moreover, always check for new data and aim to

improve the classification process to dirty language

within the texts analysed, thus the accuracy of the

classification model can be more accurate. The way

the data is presented to the responsible for the child

could be improved as well to something more

ICEIS 2016 - 18th International Conference on Enterprise Information Systems

312

automated, for example, generating a daily report in

the form of a pdf file that could be sent daily to

them.

For the classification model a possible future

work would be adding a sentiment analysis, a model

that will be trained with data from Facebook of

children (especially messages from chats and posts),

for texts made by the child, thus, the verifier would

not be so manual and an expansion of the sample

space for detection of irregularities would be

completed. Based on the training set, the model is

able to identify words in new texts related

connotations classes for which he was trained, for

example, if classes for sexual harassment

classification are positive, negative, ambiguous and

neutral, the task to classify a word in particular

would be unnecessary and thus the context would

bring us a more precise idea about the risk.

The use of heat maps or graphics could possibly

improve the data presentation in the final report,

easing understanding and providing more utility to

the analysed data. In conjunction to this, psychology

advice on how parents should approach and advise

their children would be a way to improve the work.

REFERENCES

Belloni, M.L., 2001. O que é mídia-educação (Vol. 78).

Autores Associados.

Bombonatto, Q. 2012. Associação Brasileira de

Psicopedagogia. XVI Encontro de Psicopedagogia do

Ceará, na UNICHRISTUS. Fortaleza, Brasil.

Buckingham, D., 2000. After the Death of Childhood:

Growing Up in the Age of Electronic Media.

Bellifemine, F.L., Caire, G. and Greenwood, D., 2007.

Developing multi-agent systems with JADE (Vol. 7).

John Wiley & Sons.

Fire, M., Kagan, D., Elyashar, A. and Elovici, Y., 2014.

Friend or foe? Fake profile identification in online

social networks. Social Network Analysis and Mining,

4(1), pp.1-23.

Kietzmann, J.H., Hermkens, K., McCarthy, I.P. and

Silvestre, B.S., 2011. Social media? Get serious!

Understanding the functional building blocks of social

media. Business horizons, 54(3), pp.241-251.

Campos, L.M.L., 2014. Mineração de Dados com

Detecção de Outliers em Tarefas de Predição de Séries

Temporais, XI Simpósio de excelência em gestão e

tecnologia. SEGET.

Mazman, S.G. and Usluel, Y.K., 2009. The usage of social

networks in educational context. World Academy of

Science, Engineering and Technology, 49(1), pp.404-

408.

Nyce, C. and CPCU, A., 2007. Predictive analytics white

paper. American Institute for CPCU. Insurance

Institute of America, pp.9-10.

Pereira, S.E.F.N., 2009. Redes sociais de adolescentes em

contexto de vulnerabilidade social e sua relação com

os riscos de envolvimento com o tráfico de drogas.

Quinlan, J.R., 1986. Induction of decision trees. Machine

learning, 1(1), pp.81-106.

Yamashita, R., 2009. Facebook4J A most easily usable

Facebook API wrapper in Java. Retrieved from

http://facebook4j.org/en/index.html.

Hall, M., Frank, E., Holmes, G., Pfahringer, B.,

Reutemann, P. and Witten, I.H., 2009. The WEKA

data mining software: an update. ACM SIGKDD

explorations newsletter, 11(1), pp.10-18.

Ribeiro, R., Teixeira, M., Wirth, A., Borges, A. and

Enembreck, F. A., 2015. A learning model for

intelligent agents applied to poultry farming. In:

International Conference on Enterprise Information

Systems (ICEIS). Proceedings of the 17th

International Conference on Enterprise Information

Systems (ICEIS), Barcelona, Spain.

Silveira V., Campos, G., Cortés, M., 2014. A Problem-

solving Agent to Test Rational Agents - A Case Study

with Reactive Agents. In: International Conference on

Enterprise Information Systems (ICEIS). Lisbon,

Portugal.

Santin, P. L. L., Freitas, C. O. A., Paraiso, E. C., 2011.

Análise automática de textos de mensagens

instantâneas para detecção de aliciamento sexual de

crianças e adolescentes. V. 2, n. 2, p. 43-59, PUC -

Paraná,.

Silvio, C., 2012. 38% das crianças no Facebook têm idade

abaixo do permitido, LeiaJá, v. 21, n.7, edição 344,

p.18-22, São Paulo.

Lemieux, V., Ouimet, M. and Pereira, S., 2008. Análise

estrutural das redes sociais.

Behavioral Analysis for Child Protection in Social Network through Data Mining and Multiagent Systems

313