Integrating an Association Rule Mining Agent in an ERP System:

A Proposal and a Computational Scalability Analysis

Rafael Marin Machado de Souza

, Fabrício Gomes Vilasbôas

, Pollyana Notargiacomo

and Leandro Nunes de Castro

Games, Learning, Simulation, Systems and Signals Laboratory (JAS3), Graduate Program in Electrical Engineering and

Computing, Mackenzie Presbyterian University, Rua da Consolação 930, São Paulo, Brazil

Natural Computing and Machine Learning Laboratory (LCoN), Graduate Program in Electrical Engineering and

Computing, Mackenzie Presbyterian University, Rua da Consolação 930, São Paulo, Brazil

Keywords: Software Agents, Data Mining, Association Rule Mining, Apriori, Erp Systems, Computational

Performance.

Abstract: Deployment flexibility, low development cost, and value-adding tools are some of the features that

developers are looking for in ERP systems. Modularization through software agents is one way of achieving

these objectives. In this sense, the present paper proposes the planning, implementation and integration of a

software agent for association rule mining into an ERP system. The development and use of tools for all

Knowledge Discovery in Databases (KDD) phases (pre-processing, data mining and post-processing), will

be presented. This includes input data, file loading for the agent processing, use of the Apriori association

rule mining algorithm, generation of output files with association rules, use of agent outputs for database

storage and use of the stored data by the item recommendation tool. Experiments were carried out focusing

the assessment of the running profile for databases of different sizes and using different computational

architectures.

1 INTRODUCTION

The use of agents, more specifically intelligent

agents, in the development of ERP systems has been

discussed in the literature and the benefits of

adopting this technology are demonstrated both in

the business and developer sides (Botta-Genoulaz et

al., 2005 Bih-Ru et al., 2006; Kishore et al., 2006).

Benefits include lowering development costs,

flexibility in business matching (made possible by

weak coupling), increased use of ERPs, and the

integration of intelligent decision-making processes

into the system itself. These benefits bring a high

success rate in deployments and make the systems

accessible to a wide range of businesses (Yi and Lai,

2008; Al-Mudimigh and Saleem, 2008; Al-

Mudimigh et al., 2009; Botta-Genoulaz et al., 2005).

As described by Russell and Norvig (2010),

software agents or software robots (softBots) are

programs capable of interacting with the proposed

environment using sensors to gather information and

return their processing by means of actuators,

according to previously specified performance

measures.

This paper proposes the development of a data

mining software agent (Lea, Gupta and Yu, 2005;

Papazoglou, 2001; Fox, Barbuceanu and Teigen,

2000), called RecBot, to perform association rule

mining using the Apriori algorithm (de Castro and

Ferrari, 2016, Kotsiantis and Kanellopoulos, 2006,

Kantardzic, 2003, Agrawal, Imielinski and Swami,

1993). Association rules allow the agents to create

intelligent recommendations for products and

services within the ERP itself, facilitating the

decision-making process of its users, increasing the

conversion rate and maximizing results. To illustrate

this development, the proposed agent will be

integrated into a real ERP and applied to a database

of business transactions in the health sector.

Furthermore, a study on the computational

performance of RecBot will be made to investigate

how this application behaves i) as a function of the

growth of the transactional database, and ii) when

run in different platforms.

778

Marin Machado de Souza, R., Vilasbôas, F., Notargiacomo, P. and Nunes de Castro, L.

Integrating an Association Rule Mining Agent in an ERP System: A Proposal and a Computational Scalability Analysis.

DOI: 10.5220/0007483307780786

In Proceedings of the 11th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2019), pages 778-786

ISBN: 978-989-758-350-6

The paper is organized as follows. In Section II a

review of association rule mining is made and the

proposed RecBot agent will be presented. Section III

presents the materials and methods for evaluating

the proposal and Section IV brings an analysis of the

execution profile. The work is concluded in Section

V with a general discussion on the proposal and

future work perspectives.

2 ASSOCIATION RULE MINING

AND THE RECBOT AGENT

There are several practical applications in which the

objective is to find relationships among attributes (or

variables), not objects. The association analysis, also

known as association rule mining, corresponds to the

discovery of association rules that present attribute

values that occur concomitantly in a database

(Agrawal, Imielinski and Swami, 1993; de Castro

and Ferrari, 2016; Han, Kamber, and Pei, 2012).

This type of analysis is typically used in marketing

actions and for the study of transactional databases.

There are two central aspects in the mining of

association rules: the efficient construction of

association rules and the quantification of the

significance of the proposed rules. That is, a good

association rule mining algorithm needs to be able to

propose associations of items that are statistically

relevant to the universe represented by the database.

More formally, association rules have the form X →

and A

and ... and A

 B

and B

and ... and B

where A

, i = 1, ..., m, and B

, j = 1, ..., n, are pairs of

attribute values.

The X → Y association rules are interpreted as

follows: database records that satisfy the condition

in X also satisfy the condition in Y.

The significance of the proposed rules is

established on the basis of statistical arguments.

Rules that involve mutually exclusive items or that

cover a very small number of transactions are of

little relevance. Thus, it is possible to objectively

propose measures of interest that evaluate such

features of the rules, such as support and trust

(Agrawal, Imielinski and Swami, 1993).

The support, or coverage, of a rule is an

important measure, since rules with very low

support values occur only occasionally. Rules with

low support are also of little interest from the

business perspective, since it does not make much

sense to promote items that customers buy little

together. For this reason, support is typically used to

eliminate uninteresting rules.

The support of an association rule, A → C, indicates

the frequency of occurrence of the rule, that is, the

probability of this rule being found in the total set of

transactions of the base:

(1)

where (A∪C) is the rule support count, which

corresponds to the number of transactions that

contain a particular set of items, and n is the total

number of transactions in the base.

Mathematically the support count of a set of

items A is given by:

(A) = | {t

| A ⊆ t

, t

∈ T} |

The confidence, or accuracy, verifies the

occurrence of the consequent part of the rule in

relation to the antecedent:

(2)

where (A) is the support count of the antecedent.

While confidence is a measure of the rule’s

accuracy, support corresponds to its statistical

significance. Together, these are the most commonly

used measures of interest in the association rule

mining literature (Al-Mudimigh and Saleem, 2008;

Al-Mudimigh, Saleem and Ullah, 2009; Han,

Kamber, and Pei, 2012; de Castro and Ferrari,

2016). During the association rule mining process,

criteria based on minimum values of support and

confidence are established so that a rule is part of the

final set of rules. However, many potentially

interesting rules can be eliminated by a minimum

support criterion, just as confidence is a measure that

ignores the support of the set of items.

One way to reduce the computational cost of

association rule mining algorithms is to decouple the

support and confidence requirements from the rules.

Because rule support only depends on the item set,

infrequent item sets can be deleted early in the

process without having to calculate their confidence.

Thus, a common strategy adopted by association

rule mining algorithms is to decompose the problem

into two subtasks: generation of the frequent itemset;

and rule generation.

Integrating an Association Rule Mining Agent in an ERP System: A Proposal and a Computational Scalability Analysis

779

Table 1: PEAS (Performance, Environment, Actuators, Sensors) description of RecBot.

Agent Type

Predictive Performance

Computational

Performance

Environment

Actuators

Sensors

Association rule

mining

Know and attend the

customers needs,

increase sales and

maximize return

Computational

performance in

different platforms

Sales,

customers and

salesmen

Output files

Input files

The work of Agrawal (1993), and Agrawal and

Srikant (1994) were pioneers in proposing an

association rule mining algorithm, named Apriori,

whose objective is to discover product associations

in large transactional databases. The Apriori

algorithm is the best-known method for association

rule mining and employs depth-first search (Russell

and Norvig, 2010) to generate candidate itemsets of

k elements from sets of items with k1 elements.

The non-frequent candidate items are deleted, and

the entire database is tracked and the frequent

itemsets obtained from the candidate itemsets. This

paper uses the Apriori algorithm to generate the

RecBot rules.

Table 1 presents the PEAS (Performance,

Environment, Actuators, Sensors) description of the

implemented RecBot agent.

Each KDD task, or even its sub-tasks, can be

performed by agents and in the most varied forms of

interaction, ranging from competition - where each

agent competes for information and the winner

forces the loser to become symbiotic - to mutual

cooperation, which leads to the evolution of all

agents of the system (Steels 1998, Eguchi, Hirasawa,

Hu and Ota, 2006).

The integration of the proposed model occurs in

a cooperative way, using sales data preprocessed by

the ERP system and in specific format as the agent

data entry. The output of the agent processing also

occurs in a specific format presented in the

following sections, and all post-processing (such as

persistence and use of association rules) will be

performed again by the ERP system (Fig. 1).

Figure 1: KDD responsibilities. Adapted from (Liu and

Motoda, 2002).

3 ERP RECBOT INTEGRATION

To illustrate the KDD and its use in business, as also

described by Bendoly (2003), a data preprocessing

interface was initially developed. This step can also

be performed by an agent, but here it was done by

the ERP system itself, which is responsible for

making the RecBot agent data file available and

running it. The agent developed here has the task of

receiving sales data previously processed by the

ERP system and mining the association rules,

allowing the recommendation of items.

RecBot was developed using the Object Pascal

language (Lazarus/Delphi), aiming to run on Linux

platforms. The development of this agent took

approximately 26 days, since it was necessary to

develop all the methods that involve data processing.

To reduce the development and total execution

time, a version of this agent was developed in C++

with support from the Data Analytics Acceleration

Library (DAAL), which offers computationally

optimized methods for all stages of the KDD

process. Its implementation, maintenance and

optimization is done by Intel® and can be purchased

together with Intel Parallel Studio or in a free

version available on Github. With the support of this

library it was possible to reduce the development

time, which previously was 26 days, to only 3 days.

The results generated by these implementations are

presented and discussed in Section IV. The input

files and outputs, such as item groups, association

rules, and runtimes, for parameterizations used in

both implementations of the same agent, are

available for online access in the RecBot Project

within the Mendeley Data platform, in directories

Inputs and Outputs, respectively (Souza, 2018).

As this is a software agent, the communication

pattern between agents has been defined through

execution parameters that must meet the following

sequence: Number of Load Threads, Confidence

Factor, Support Factor, File with input data and

folder for output files. An example of an agent call is

shown below:

./recbot 10 0.03 0.3 /data/inputs/db3k.csv/data/outputs/

ICAART 2019 - 11th International Conference on Agents and Artiﬁcial Intelligence

780

In this example, 10 threads are defined for input file

processing, a confidence factor of 3%, a support

factor of 30%, the input file is db3k.csv, which is in

the /data/inputs/directory, and the directory to store

the output files is /data/outputs/.

There is a peculiarity in the agent developed in

the C++ language. The DAAL library dynamically

defines the number of threads that will be used

during the agent processing, so the first parameter is

omitted for this version to run.

The algorithm performs a combinatorial analysis

of the items to construct the rules, counts the

transactions containing the items in order to

determine the support (frequency of occurrence) and

confidence (accuracy) of the rules. The aim is to find

rules that satisfy minimum values of support and

confidence, pruning all candidate rules that have not

reached the pre-defined minima (Agrawal,

Imielinski and Swami, 1993; Nath, Bhattacharyya1

and Gosh, 2012; de Castro and Ferrari, 2016).

An interface for transactional data processing

was developed in the ERP to select the filtered

transactions and allow the export of the file used for

communication with the agent. The data entry files

must be in the Comma-Separated Values (CSV)

standard and contain the preprocessed transactional

data in the "id, item" format, where id is a sequential

code used to represent the transaction and item is the

transaction, as illustrated in Figure 2.

Figure 2: Partial Input file structure.

The agent starts by loading the minimum support

and confidence values chosen for the execution and

loading the input data file. From this a matrix

composed of N transactions is fed, indicating 1 when

the item is present in the transaction, and 0

otherwise. Thus, the frequency of occurrence of each

item is calculated and those that meet the minimum

support are evaluated in pairs, until there are no

possibilities of combinations that meet the minimum

support (de Castro and Ferrari, 2016; Kantardzic,

2003). The support calculation is done by Eq. (1).

All subsets with support less than the minimum

support are excluded (pruned) from the next step,

and so the algorithm continues until there are no

more possible combinations.

With the subsets that meet the minimum support

properly selected, the next step is to use this result to

build the rules a → c (if a, then c), and these new

combinations must meet the minimum confidence

(Eq. (2)). Rules that do not meet the minimum

confidence will be pruned and the process is

repeated until all rules have been evaluated. The

complete process is shown in Figure 3.

Figure 3: How the agent works.

Thus, three outputs are created by the agent:

1) one with the subsets that meet the minimum

support and its respective support value

(outpuItemSet file), as shown in Figure 4; 2) another

file with the association rules that meet the

minimum confidence (outputRules file); and 3) one

with the running time of each agent process, as

described in Figure 4.

The association rules are loaded by the ERP

system and stored in a database for later use.

Determining the most flexible format for recording

these data in a relational database has been one of

the challenges of this work, because it was necessary

to consider various combinations of items made in

the sale (antecedents), and to use SQL queries to

find the items to recommend (consequents). As a

proposed solution, two tables were created, one to

record the antecedent items and their confidence

values and another to record the consequent items.

Each rule is assigned an ID.

Thus, to select the most relevant rules, a query is

made with the existing items of the current

transaction and all rules that have existing

combinations are selected. The consequent items of

these rules that do not exist in the current transaction

are selected and presented to the user as

recommendations.

Integrating an Association Rule Mining Agent in an ERP System: A Proposal and a Computational Scalability Analysis

781

(a) (b)

(c)

Figure 4: (a) Frequent itemset. (b) Association rules.

4 PERFORMANCE

EVALUATION

This section presents an analysis of the execution

profile of the Apriori algorithm implemented in two

versions: 1) one in Object Pascal (Lazarus/Delphi);

and 2) another in C++ with DAAL library support,

as described previously.

Two computer systems of shared memory

architecture and with the support of SIMD (Single

Instruction Multiple Data) type flows were used.

The first computer system consists of two 1.83 MHz

Dual Core Xeon processors and 12 GB of RAM, two

SATA 2.0TB and Linux Ubuntu Server 18.04 LTS.

The second computer system consists of two Intel

Xeon Platinum 8160 @ 2.10 GHz processors, each

with 24 physical cores (48 logical) and 33 MB cache

memory, 190 GB RAM, two Intel S3520 Series

1.2Gb and 240TB SSDs GB capacity and CentOS 7

operating system with kernel version 3.10.0-

693.21.1.3l7.x86_64. For reasons of ease of

identification, the first computer system was named

Woodcrest and the second one Skylake.

Therefore, to improve the use of the

computational systems available for our experiments

the algorithm was transcribed from the Object Pascal

language into the C++ language using the DAAL

library methods. The two computer systems have

microarchitecture processing units developed by

Intel that have made efforts to apply computational

optimization and parallelism techniques to both the

C++ language compiler and DAAL, both available

in the Intel Parallel Studio XE (Intel, 2018a). In the

shared memory systems, DAAL supports thread-

level parallelism using Threading Building Blocks

(Intel, 2018b), which is a runtime-based parallel

programming model for C/C++ code, and supports

vectorization, which is the programming model for

the SIMD architecture. Thus, there is the guarantee

of using all resources available by the system.

Table 2: Number of itemsets and rules in RecBot.

Dataset size

10

Support 2%

Confidence 20%

Support 3%

Confidence 30%

Item sets

Rules

Item sets

Rules

178

142

150

139

126

127

192

124

384

118

768

111

1536

121

3072

131

Samples of varying sizes from the same real-

world database were used for the experiments. The

initial sample has a size of 3,000 (3K) records, and it

doubles up to 3,072,000 records (Table 2). We also

tested the following minimum support (minsup) and

minimum confidence (minconf) values: minsup =

2%, minconf = 20%; and minsup = 3%, minconf =

30%. As previously explained, the higher the

support and confidence, the greater the correlation

between the items. Their configuration depends on

business demand of data, since the higher the value

of these parameters, the less itemsets will be

generated and, therefore, the less items will be

recommended to the user application.

Table 2 shows the number of items in the

frequent itemset and the number of rules generated

for each dataset size. It can be noted that increasing

the database does not imply increasing the number

of frequent items or the number of rules. While this

may seem counterintuitive, the explanation lies in

ICAART 2019 - 11th International Conference on Agents and Artiﬁcial Intelligence

782

Table 3: Performance evaluation of the RecBot agent implementations in the Woodcrest and Skylake architectures for

exponentially growing databases.

Dataset size

10

Server

C++

Object Pascal

minsup 2%

minconf 20%

minsup 3%

minconf 30%

minsup 2%

minconf 20%

minsup 3%

minconf 30%

Woodcrest

112

200

127

Skylake

Woodcrest

121

376

244

Skylake

114

Woodcrest

160

119

938

425

Skylake

242

133

Woodcrest

182

155

2040

957

Skylake

516

323

Woodcrest

244

226

4465

2319

Skylake

1140

762

Woodcrest

420

370

9670

4600

Skylake

3246

1889

192

Woodcrest

729

661

31615

13355

Skylake

100

8110

4835

384

Woodcrest

1345

1270

110160

70671

Skylake

150

141

17933

10721

768

Woodcrest

2540

2421

225943

154344

Skylake

254

236

36104

21856

1536

Woodcrest

5032

4768

459691

309819

Skylake

461

432

80123

47075

3072

Woodcrest

10122

9583

Skylake

826

815

237129

191673

the fact that the increase in the database also

increases the denominator of the coverage

calculation and the support of a rule, causing fewer

frequent items to meet the minimum support

criterion and fewer rules satisfying the minimum

confidence level.

As the Apriori algorithm is deterministic, a

single experiment was performed for each

configuration and the results presented in Table 3

are the total Execution Time (ET) of the two

implementations, for the different combinations of

minsup and minconf, and increasing dataset sizes.

Figure 5 shows that the C++ version provided a

98.46% reduction in execution time when compared

with the Object Pascal implementation. It can also

be observed that for the 3072K dataset the Object

Pascal version could not converge, because there

was not enough RAM space to store the data

structures and the program was shut down by the

operating system. By contrast, the C++ version

finished execution and the running time still

complies with the previously observed pattern.

Figure 5 plots the execution time (ET) of the

Woodcrest architecture for both implementations:

Object Pascal and C++. It is noted that the ET

increase is exponential, like the dataset increase, but

the increase for Object Pascal occurs at a

significantly higher rate than the increase for C++.

A similar behavior can be observed in Figure 6

for the Skylake architecture. The difference between

the execution time of the C++ version in relation to

the Object Pascal version comes to be 190858ms

faster for largest dataset (3072K objects), a 99.57%

reduction in execution time. The average growth rate

of the C++ runtime is approximately 1.37 and for the

Object Pascal version it is approximately 2.32. This

shows that by doubling the dataset size, the Object

Pascal time more than doubles, while the C++

version has an increase of approximately 44%.

Figure 7 summarizes the gain in performance when

comparing both implementations in both

architectures. It can be observed that the C++

version had a gain of 286.94 times to minsup = 2%

and minconf = 20% and a gain of 235.18 times to

minsup = 3% and minconf = 30% in the Skylake for

Integrating an Association Rule Mining Agent in an ERP System: A Proposal and a Computational Scalability Analysis

783

(a)

(b)

Figure 5: Growth rate comparison between each

implementation in the Woodcrest archtiecture. First

Column of dataset: C++; Second Column of dataset:

Object Pascal. (a) minsup = 2% and minconf = 20%; (b)

minsup = 3% and minconf = 30%.

the dataset with 3072K objects. The gain is obtained

from the ratio between the execution time of the

Object Pascal version and the execution time of the

C++ version. By analyzing the pattern of gain

growth, it can be observed that it is very close to

stabilizing in the Woodcrest environment, whilst in

the Skylake environment the gain still seems to be

growing. This reinforces the conclusion that the C++

version is best suited for database scalability. This

result also demonstrates the efficiency of the

Skylake architecture since the operations performed

by the algorithm are the same in all environments.

5 CONCLUSIONS

This paper introduced an association rule mining

agent to be integrated with an ERP system as an

alternative to implement a microservice structure.

For this, the agent was developed in two different

languages, Object Pascal (Delphi/Lazarus) and C++,

(a)

(b)

Figure 6: Growth rate comparison between each

implementation in the Skylake archtiecture. First Column

of dataset: C++; Second Column of dataset: Object Pascal.

(a) minsup = 2% and minconf = 20%; (b) minsup = 3%

and minconf = 30%.

making use of different programming techniques. In

the first implementation (Object Pascal) the Apriori

algorithm was fully coded using object orientation,

which led to a programming time 8 times greater

than the second one (C++), which was done using

third party libraries (DAAL Intel). The use of Intel

DAAL also provided considerable performance

gains.

The performance of these agent

implementations was compared in two distinct

computational architectures: Woodcrest and

Skylake. In the performance analysis presented, a

real transactional database was used with 3,072,000

objects, presented in increasing subsets from 3,000

objects, doubling at each experiment. Two distinct

configurations of minimum confidence and support

thresholds were also tested, allowing empirically

investigating the application scaling with the dataset

size.

In the time analysis, it was obtained a reduction

in the total execution time of 99.57% in the Skylake

and 98.46% in the Woodcrest for the database with

ICAART 2019 - 11th International Conference on Agents and Artiﬁcial Intelligence

784

(a)

(b)

Figure 7: Reduction in times of the execution time of the C++ version in relation to the Object Pascal in the (a) Skylake and

the (b) Woodcrest architecture.

3072K objects. It has also been observed that the

C++ version is more suited to dataset scalability,

presenting a runtime increase rate of approximately

1.37 times as opposed to 2.32 times of the Object

Pascal version.

With this, the efficiency of the presented

technique is determined, since its flexibility of

languages and environments demonstrated the

viability of the solution for different operating

platforms and computational architectures.

As further works it is proposed the use of the

implemented agent in real-world ERP systems and

the assessment of its business improvement, such as

the number of sold items and billing volume

(considering seasonalities and possible outliers).

ACKNOWLEDGEMENTS

The authors thank CAPES, CNPq, Fapesp, and

Mackpesquisa for the financial support. The authors

also acknowledge the support of Intel for the Natural

Computing and Machine Learning Laboratory as an

Intel Center of Excellence in Artificial Intelligence.

REFERENCES

Agrawal, R., Imielinski, T. and Swami, A. 1993. Mining

Association Rules between Sets of Items in Large

Databases. In: Proceedings of the 1993 ACM

SIGMOD international conference on Management of

data, Volume 22 Issue 2, pp 207-216.

Agrawal, R., Srikant, R. 1994. Fast Algorithms for Mining

Association Rules in Large Databases. Proceedings of

the 20th International Conference on Very Large Data

Bases (VLDB), p. 487-499.

Al-Mudimigh, A. S. and Saleem, F. 2008. A Framework

of an Automated Data Mining Systems Using ERP

Model. International Journal of Computer and

Electrical Engineering.

Al-Mudimigh, A. S., Saleem, F. and Ullah, Z. 2009. The

effects of data mining in ERP-CRM model: a case

study of MADAR. W. Trans. on Comp. v. 8, n. 5, p.

831-843.

B. Nath, D. K., Bhattacharyya1 and A. Ghosh. 2013.

Incremental association rule mining: a survey. WIREs

Integrating an Association Rule Mining Agent in an ERP System: A Proposal and a Computational Scalability Analysis

785

DataMining and knowledge Discovery, Volume 3,

Issue 3, pp 157–169.

Bendoly, E. 2003. Theory and support for process

frameworks of knowledge discovery and data mining

from ERP systems. Information & Management, v. 40,

pp. 639–647, 2003.

Bih-Ru, L., Gupta, M. C., Wen-Bin, Y. 2005. A prototype

multi-agent ERP system: an integrated architecture

and a conceptual framework. Technovation, 25(4), pp.

433-441.

Botta-Genoulaz V., Millet, P.-A. and Grabot, B. 2005. A

survey on the recent research literature on ERP

systems. Computers in Industry, v. 56, n. 6, pp. 510-

522, 2005. ISSN 0166-3615,

https://doi.org/10.1016/j.compind.2005.02.004

Castro, L. N. 2016. Introdução a Mineração de dados:

Conceitos, Algoritmos e aplicações. São Paulo:

Editora Saraiva.

Eguchi, T., Hirasawa, K., Hu, J. and Ota, N. 2006. A

Study of Evolutionary Multiagent Models Based on

Symbiosis. IEEE TRANSACTIONS ON SYSTEMS,

MAN, AND CYBERNETICS, Vol. 36, n. 1.

Flynn, M. J. 1972. Some computer organizations and their

effectiveness. IEEE transactions on computers, IEEE,

v. 100, n. 9, p. 948–960.

Fox, M., Barbuceanu, M. and Teigen, R. 2000. Agent-

oriented supply-chain management. International

Journal of Flexible Manufacturing Systems,

v. 12, p. 165–188.

Han, J., Kamber, M. and Pei, J. 2012. Data mining:

concepts and techniques. Amsterdam: Morgan

Kaufmann.

Intel Corp. 2017. DAAL - Data Analytics Acceleration

Library. Available: https://software.intel.com/pt-

br/intel-daal

Intel Corp. 2018a. Intel® Parallel Studio XEAvailable:

https://software.intel.com/en-us/parallel-studio-xe

Intel Corp. 2018b. Intel® Threading Building Blocks

(Intel® TBB) Available: https://software.intel.com/en-

us/intel-tbb

J. Yi, J. and Lai, C. 2008. Research on Reengineering of

ERP System Based on Data Mining and MAS.In:

International Symposium on Knowledge Acquisition

and Modeling, IEEE Computer Society, pp. 180–184.

Kantardzic, M. 2003. Data Mining: Concepts, models,

methods and algorithms. Piscataway: IEEE Press.

Kishore, R., Zhang, H., Ramesh, R. 2006, Enterprise

integration using the agent paradigm: foundations of

multi-agent-based integrative business information

systems. Decision Support Systems, 42(1), pp. 48-78.

Kotsiantis, S. and Kanellopoulos, D. 2006. Association

Rules Mining: A Recent Overview. GESTS

International Transactions on Computer Science

and Engineering, v.32, n. 1, p. 71-82.

Liu, H. and Motoda, H. 2000. Feature Selection for

Knowledge Discovery and Data Mining, Kluwer

Academic Publishers.

Lea, B-R. and Gupta, M. C. and Yu, W-B. 2005. A

prototype multi-agent ERP system: an integrated

architecture and a conceptual framework.

Technovation. v. 25, n. 4, p. 433-441.

Papazoglou, M. 2001. Agent-oriented technology in

support of e-business. Communications of the ACM,

v.44, n. 4, p. 71–77.

Rajaraman, V. and Siva, R. M. C. 2016. Parallel

Computers Architecture and Programming. PHI

Learning Pvt. Ltd.

Russell, S. and Norvig, P. 2010. Artificial Intelligence A

Modern Approach, 3nd ed. New York, NY, USA:

Prentice Hall.

Steels, L. 1998. The Origins of Ontologies and

Communication Conventions in Multi-Agent Systems.

In: Autonomous Agents and Multi-Agent Systems, Vol.

1, Issue 2, pp 169–194.

Souza, Rafael Marin Machado de. 2018. RecBot Inputs

and Outputs. Mendeley Data.

Tanenbaum, A. S. 2007. Organização e Estrutura de

Computadores. 5. ed. [S.l.]: Pearson Prentice Hall.

ICAART 2019 - 11th International Conference on Agents and Artiﬁcial Intelligence

786