A SERIAL COMPUTING MODEL OF AGENT ENABLED MINING OF GLOBALLY STRONG ASSOCIATION RULES

International Journal on Computational Sciences & Applications (IJCSA) Vol.5, No.3, June 2015

DOI:10.5121/ijcsa.2015.5307 77

A SERIAL COMPUTING MODEL OF AGENT

ENABLED MINING OF GLOBALLY STRONG

ASSOCIATION RULES

G.S.Bhamra

1, A. K.Verma

2 and R.B.Patel

3

1M. M. University, Mullana, Haryana, 133207 - India 2Thapar University, Patiala, Punjab, 147004- India

3Chandigarh College of Engineering & Technology, Chandigarh- 160019- India

ABSTRACT The intelligent agent based model is a popular approach in constructing Distributed Data Mining (DDM)

systems to address scalable mining over large scale and ever increasing distributed data. In an agent based

distributed system, variety of agents coordinate and communicate with each other to perform the various

tasks of the Data Mining (DM) process. In this study a serial computing mode of a multi-agent system

(MAS) called Agent enabled Mining of Globally Strong Association Rules (AeMGSAR) is presented based

on the serial itinerary of the mobile agents. A Running environment is also designed for the implementation

and performance study of AeMGSAR system.

KEYWORDS Knowledge Discovery, Association Rules, Intelligent Agents, Multi-Agent System

1.INTRODUCTION

Data Mining (DM) technique is used to extract some interesting and valid data patterns implicitly

stored in large databases [1], [2]. Intelligent software agent technology is an interdisciplinary

technology dealing with the development and efficient utilization of autonomous software objects

called agents which have access to geographically distributed and heterogeneous resources. They

are autonomous, adaptive, reactive, pro-active, social, cooperative, collaborative and flexible.

They also support temporal continuity and mobility within the network. An intelligent agent with

mobility feature is known as Mobile Agent (MA). MA migrates from node to node in a

heterogeneous network without losing its operability. On reaching at a network node MA is

delivered to an Agent Execution Environment (AEE) where its executable parts are started

running. Upon completion of the desired task, it delivers the results to the home node. A Mobile

Agent Platform (MAP) or Agent Execution Environment (AEE), is a server application that

provides the appropriate functionality to MAs to authenticate, execute, communicate, migrate to

other platform, and use system resources in a secure way. A Multi Agent System (MAS) is

distributed application comprised of multiple interacting intelligent agent components [3].

Let { }, 1j

DB T j D= = K be a transactional dataset of size D where each transaction T is assigned

an identifier (TID ) and { }, i 1i

I d m= = K , total m data items in DB . A set of items in a particular

transaction T is called itemset or pattern. An itemset, { }, i 1i

P d k= = K , which is a set of k data


78

items in a particular transaction T and P I⊆ , is called k-itemset. Support of an itemset,

( )No_of_T_containing_P

%s PD

= is the frequency of occurrence of itemset P in DB , where

No_of_T_containing_P is the support count (sup_count) of itemset P . Frequent Itemsets (FIs)

are the itemset that appear in DB frequently, i.e., if ( ) min_th_sups P ≥ (given minimum

threshold support), then P is a frequent k-itemset. Finding such FIs plays an essential role in

miming the interesting relationships among itemsets. Frequent Itemset Mining (FIM) is the task

of finding the set of all the subsets of FIs in a transactional database [2].

Association Rules (ARs) are used to discover the associations among item in a database [4]. It is

an implication of the form [ ]support,confidenceP Q⇒ where, ,P I Q I⊂ ⊂ and P Q∩ = ∅ . An

AR is measured in terms of its support and confidence factor where support of the rule

( ( )s P Q⇒ ) is the probability of both P and Q appearing in T , i.e., ( )p P Q∪ and the

confidence of the rule ( ( )c P Q⇒ ) is the conditional probability of Q given P , i.e., ( )|p Q P .

An AR is said to be strong if ( ) min_th_sups P Q⇒ ≥ (given minimum threshold support) and

( ) min_th_confc P Q⇒ ≥ (given minimum threshold confidence). Association Rule Mining (ARM)

today is one of the most important aspects of DM tasks. In ARM all the strong ARs are generated

from the FIs. The ARM can be viewed as two step process [5], [6].

1. Find all the frequent k-itemsets (k

L )

2. Generate Strong ARs from k

L

a. For each frequent itemset, k

l L∈ , generate all non empty subsets of l .

b. For every non empty subset s of l , output the rule “ ( )s l s⇒ − ”, if

( )

( )

sup_countmin_th_conf

sup_count

l

s≥

Distributed Association Rule Mining (DARM) is the task of generating the globally strong

association rules from the global FIs in a distributed environment. Few preliminaries notations

and definitions required for defining DARM and to make this study self contained are as follows:

• { }, i 1i

S S n= = K , n distributed sites.

•

CENTRALS , Central Site.

• { }, 1i j i

DB T j D= = K , Horizontally partitioned data set of size i

D at the local sitei

S , where

each transaction j

T is assigned an identifier (TID).

•

1

n

iiDB DB

==U , the aggregated dataset of size

1

n

iiD D

==∑ ,

i jDB DB∩ = ∅

• { }, i 1i

I d m= = K , total m data items in each i

DB .

•

( )

FI

k iL , Local frequent k-itemsets at site

iS .

•

( )

FISC

k iL , List of support count

( )

FI

k iItemset L∀ ∈ .

•

LSAR

iL , List of locally strong association rules at site

iS .

•

1

nTLSAR LSAR

iiL L

==U , List of total locally strong association rules.

•

( )1

nTFI FI

k k iiL L

==U , List of total frequent k-itemsets.


79

•

( )1

nGFI FI

k k iiL L

==I , List of global frequent k-itemsets.

•

GSAR

CENTRALL , List of Globally strong association rule.

Local Knowledge Base (LKB), at siteiS , comprises of

( )

FI

k iL ,

( )

FISC

k iL and

LSAR

iL which can provide

reference to the local supervisor for local decisions. Global Knowledge Base (GKB), atCENTRAL

S ,

comprises ofTLSAR

L ,TFI

kL ,

GFI

kL and

GSAR

CENTRALL for the global decision making [7]. Like ARM, DARM

task can also be viewed as two-step process [6]:

1. Find the global frequent k-itemset (GFI

kL ) from the distributed Local frequent k-itemsets

(( )

FI

k iL ) from the partitioned datasets.

2. Generate globally strong association rules (GSAR

CENTRALL ) from GFI

kL .

The existing agent based systems specifically dealing with DARM task are: Knowledge

Discovery Management System (KDMS) [8], Efficient Distributed Data Mining using Intelligent

Agents [9], Mobile Agent based Distributed Data Mining [10], An Agent based Framework for

Association Rule Mining of Distributed Data (AFARMDD) [11], [12], Multi-Agent Distributed

Association Rule Miner (MADARM) [13]. All these systems are academic research projects.

Qualitative comparison of these DARM frameworks is provided in [14]. Most of the existing

agent based frameworks for DARM task are only prototype model and lacks the appropriate

underlying AEE, scalability, privacy preserving techniques, global knowledge generation and

implementation using a real datasets.

The rest of the paper is organised as follows. Section 2 described the running environment for the

proposed system along with various algorithms involved. Serial computing model of AeMGSAR

is presented in Section 3. Algorithms for all the agents involved in this system are also discussed.

Section 4 describes the implementation and performance study of the system and finally the

article is concluded in Section 5.

2.ENVIRONMENT FOR THE PROPOSED SYSTEM

Every MAS needs an underlying AEE to provide a running infrastructure on which agents can be

deployed and tested. A running environment has been designed in Java. Various attributes of the

MA are encapsulated within a data structure known as AgentProfile . It contains the name of MA

( AgentName ), version number ( AgentVersion ), entire byte code ( BC ), list of nodes to be

visited by MA, i.e., itinerary plan ( NODESL ) , type of the itinerary ( ItinType ) which can be

serial or parallel, a reference of current execution state ( AObject ) and an additional data structure

known as Briefcase that acts as a result bag of MA to store final resultant knowledge ( iResult_S )

at a particular site. Computational time ( CPUTime ) taken by a MA at a particular site is also

stored in iResult_S . In addition to results, Briefcase also contains the system time for start of

agent journey ( startTripTime ), system time for end of journey ( endTripTime ) and total round trip

time of MA (TripTime ) calculated using end startTripTime TripTime TripTime← − . Stationary as well

as mobile agents involved in the models would be discussed later on. This environment consists

of the following three components:


80

• Data Mining Agent Execution Environment (DM_AEE): It is the key component that

acts as a Server. DM_AEE is deployed on any distributed sites iS and is responsible for

receiving, executing and migrating all the visiting DM agents. It receives the incoming

AgentProfile at site iS , retrieves the entire BC of agent and save it with

.AgentName class in the local file system of the site iS after that execution of the agent is

started using AObject . Steps are shown in Algorithm 1.

• Agent Launcher (AL): It acts a Client at agent launching station (CENTRAL

S ) and launches

the goal oriented DM agents on behalf of the user through a user interface to the

DM_AEE running at the distributed sites. Agent Pool (or Zone) at CENTRAL

S is a repository

of all mobile as well as stationary agents (SAs). AL first reads and stores AgentName

in AgentProfile . The entire BC of the AgentName is loaded from the Agent Pool and

stored in AgentProfile . NODESL and ItinType are retrieved and stored in AgentProfile .

startTripTime is maintained in Briefcase which is further added to AgentProfile . In case of

serial computing model, i.e., if ItinType Serial= , AL dispatches a specific single MA

along with NODESL , and it travels from node to node. AgentVersion is set as 1 for this

agent. AL also contacts the Result Manager (RM) for processing the Briefcase of an agent.

Detailed steps are given in Algorithm 2.

• Result Manager (RM): It manages and processes the Briefcase of all MAs. RM is either

contacted by a MA for submitting its results or by AL for processing the results of the

specific MA. On completion of itinerary, each DM agent submits its results to RM which

computes total round trip time ( TripTime ) of that MA and saves it in the Briefcase of that

agent. It ItinType Serial= then it saves the updated AgentProfile of an agent at CENTRALS .

When it is contacted by AL for processing the results of a specific agent it sends back the

AgentProfile of that agent. Steps are defined in Algorithm 3.

Algortihm 1 DATA MINING AGENT EXECUTION ENVIRONMENT (DM_AEE)

1: procedure DM_AEE( )

2: while TRUE do

3: iAgentPofile listen and receive AgentProfile at S←

4: AgentName get AgentName from AgentProfile←

5: BC retrieve the BC of agent from AgentProfile←

6: isave the BC with AgentName.class in the local file system of S

7: AObject get AObject from AgentProfile← > current state

8: . ()AObject run > start executing mobile agent

9: end while

10: end procedure

Algortihm 2 AGENT LAUNCHER (AL)

1: procedure AL( )

2: option read option (dispatch / result)←

3: switch option do

4: case dispatch >dispatch the mobile agent to DM_AEE

5: AgentName read Mobile Agent's name←

6: add AgentName to AgentProfile


81

7: BC load entire byte code of AgentName from AgentPool←

8: add BC to AgentProfile

9: NODESL read Itinerary (IP addresses) of mobile agent←

10: ItinType read ItinType ( Serial / Parallel)←

11: add ItinType to AgentProfile

12: if " "ItinType Serial= then >Serial Itinerary

13: 1AgentVersion ←

14: add AgentVersion to AgentProfile

15: NODESadd L to AgentProfile

16: switch AgentName do

17: case LFIGA

18: minthrsup read minimum threshold support←

19: AObject new LFIGA(AgentProfile, minthrsup)←

20: end case

21: case LKGA

22: minthrconf read minimum threshold confidence←

23: AObject new LKGA(AgentProfile, minthrconf)←

24: end case

25: case TFICA

26: AObject new TFICA(AgentProfile)←

27: end case

28: case LKCA

29: (AObject new LKCA AgentProfile)←

30: end case

31: case GKDA

32: GSAR GSAR

CENTRAL CENTRAL CENTRALL load L generated by GKGA at S←

33: GSAR

CENTRALadd L to Briefcase

34: add updated Briefcase to AgentProfile

35: AObject new GKDA (AgentProfile)←

36: end case

37: end switch

38: add AObject to AgentProfile >current state

39: NODESTransfer AgentProfile to DM_AEE at first IP address in L

40: end if

41: end case

42: case result >process the result of mobile agent

43: AgentName read mobile agent's name←

44: ItinType read mobile agent's ItinType←

45: AgentInfoadd AgentName to L

46: AgentInfoadd ItinType to L

47: > Result processing for Serial Itinerary Agents

48: if " "ItinType Serial= then

49: AgentInfoAgentProfile contact RM for L←

50: Briefcase retrieve Briefcase from AgentProfile←

51: switch AgentName do

52: case LFIGA


82

53: process the Briefcase of LFIGA

54: end case

55: case LKGA

56: process the Briefcase of LKGA

57: end case

58: case TFICA

59: call GFIGA (Briefcase) > stationary agent

60: end case

61: case LKCA

62: call GKGA (Briefcase) > stationary agent

63: end case

64: case GKDA

65: process the Briefcase of GKDA

66: end case

67: end switch

68: end if

69: end case

70: end switch

71: end procedure

Algortihm 3 RESULT MANAGER (RM)

1: procedure RM( )

2: while TRUE do

3: listen and receive the incomming request

4: if icontacted by a mobile agent for submitting results from site S then

5: iAgentProfile receive the incomming AgentProfile from site S←

6: ItinType retrieve ItinType from AgentProfile←

7: Briefcase retrieve mobile agent's Briefcase from AgentProfile←

8: start startTripTime retrieve TripTime from Briefcase←

9: end endTripTime retrieve TripTime from Briefcase←

10: end startTripTime TripTime TripTime← −

11: add TripTime to Briefcase



14: CENTRALsave AgentProfile at S

15: end if

16: end if

17: if contacted by AL for processing the results then

18: AgentInfoAgentName retrieve AgentName from incomming L←

19: AgentInfoItinType retrieve ItinType from incomming L←


21: CENTRALAgentProfile load AgentProfile for AgentName from S←

22: dispatch AgentProfile to AL

23: end if

24: end if

25: end while

26: end procedure


83

The overall working of AeMGSAR system may be divided into following six stages:

1. Request Stage: Request for the DARM is initiated at CENTRALS by AL on behalf of the user

with necessary credentials.

2. Preparation Stage: AL through User Interface reads agent name; version number;

Itinerary for the MAs journey is obtained in terms of IP addresses of the distributed nodes

to be visited by a MA; any specific additional data for a specific MA is obtained; Agent

code for the specific MA is loaded from AgentPool; for serial itinerary a single specific

MA is dispatched by AL to travel and visit n distributed sites in parallel.

3. Local Mining Stage: ARM process is performed locally by specific DM agents on each

distributed site and results are kept as local knowledge base at that site.

4. Result Collection Stage: Collector agents visits each site and collect the results generated

by DM agents and submit the results back to RM at CENTRALS .

5. Knowledge Integration and Global Knowledge Generation Stage: Knowledge or result

integration is carried out by the RM with the help of stationary agent and Global Knowledge in the form of Globally Strong Association Rules may be generated with the

help of other stationary agents at CENTRALS .

6. Global Knowledge Dispatching Stage: Global knowledge is dispatched to the distributed

sites by a dispatching agent to compare it with the local knowledge at each site.

Figure 1. AeMGSAR Serial Computing Model


84

3.SERIAL COMPUTING MODEL OF AEMGSAR

Serial computing model of AeMGSAR system is shown in Figure 1. It consists of total seven

agents, five of these are MAs dispatched from CENTRALS with serial itinerary multi-hop migration

and other two are intelligent SAs running at CENTRALS to perform different tasks. The CPU time

taken by a MA while processing on each site along with some other specific information is

carried back in the result bag at CENTRALS . Agents in serial number 1-5 visit n sites serially other

parameters are collected from different resources. Detailed relationship among these agents and

working behaviour of each agent is as follows:

1. Local Frequent Itemset Generater Agent (LFIGA): This is a MA that carries the

AgentProfile & min_th_sup . LFIGA generates and stores ( )

FI

k iL and ( )

FISC

k iL at site iS by

scanning the local iDB at that site with the constraint of min_th_sup . It carries back the

computational time ( CPUTime ) at each site iS and

endTripTime . This agent is embedded

with Apriori algorithm [15] for generating all the frequent k-itemset lists. It may be

equipped with decision making capability to select other FIM algorithms based on the

density of the dataset at a particular site. More details are available in Algorithm 4.

2. Local Knowledge Generater Agent (LKGA): This is a MA that carries the

AgentProfile & min_th_conf . LKGA applies the constraint of min_th_conf to generate and

store LSAR

iL by using the ( )

FI

k iL and ( )

FISC

k iL lists already generated by LFIGA agent at site iS .

LSAR

iL list also support and confidence for a particular association rule along with the site

name. It carries back the computational time ( CPUTime ) at each site iS and endTripTime .

Detailed steps are given in Algorithm 7.

3. Total Frequent Itemset Collector Agent (TFICA): This is a MA that carries the

AgentProfile . TFICA collects list of local frequent k-itemset ( ( )

FI

k iL ) generated by LFIGA

agent and carries back the list of total frequent k-itemset TFI

kL in the result bag to RM at

CENTRALS . In addition to this resultant knowledge, it also carries back the computational

time ( CPUTime ) at each site iS and endTripTime . It executes Algorithm 8.

4. Local Knowledge Collctor Agent (LKCA): This is a MA that carries the AgentProfile .

LKCA collects the list of locally strong association rules ( LSAR

iL ) generated by LKGA

agent and carries back the list of total locally strong association rules ( TLSARL ) in the result

bag to RM at CENTRALS . In addition to this resultant knowledge, it also carries back the

computational time ( CPUTime ) at each site iS and endTripTime . Steps are shown in

Algprithm 9.

5. Global Knowledge Dispatcher Agent (GKDA): This is a MA that carries the

AgentProfile containing global knowledge ( GSAR

CENTRALL ). It dispatches global knowledge at

every site for further decision making and comparing with the local knowledge at that

site. It executes Algorithm 12.

6. Global Frequent Itemset Generater Agent (GFIGA): It is a stationary agent at CENTRALS ,

mainly used for processing the result bag of TFICA, i.e., total frequent k-itemset list

( TFI

kL ) generated y TIFCA to generate the global frequent itemset list, GFI

kL . More details

are available in Algorithm 10.

7. Global Knowledge Generater Agent (GKGA): It is also a stationary agent at CENTRALS ,

mainly used for processing the GFI

kL list and TLSARL list to compile the global knowledge,


85

i.e., the list of globally strong association rules, GSAR

CENTRALL . Detailed steps are shown in

Algorithm 11.

Algortihm 4 LOCAL FREQUENT ITEMSET GENERATER AGENT (LFIGA)

Input:

• AgentProfile,A collection of agent attributes set by the AL

• min_th_sup, the given minimum threshold support

Output: FI &SCL , the list of frequent itemsets and their support counts

1: procedure LFIGA( AgentProfile,min_th_sup )

2: startCPUTime get system time←

3: Briefcase get Briefcase from AgentProfile←

4: i i iDB load DB from local file system of site S←

5: . (0)iT DB get← >No. of records

6: . (1)iI DB get← >No. of items

7: . (3)iDB[T][I] DB get← > itemset data bank

8: minsupcount (T × min_th_sup) / 100←

9: >generate frequent-1 itemset list (1FIL ) and support count list (

1FISC )

10: 1CFIL {1,2,3...I}← > candidate frequent-1 itemset

11: for i 1,I← do > initialize the support count array 1SCFIL to zero

12: 01SCFIL [i] ←

13: end for

14: 1k ←

15: for all 1candidate c CFIL∈ do > find support count for every candidate

16: for all transaction t DB∈ do

17: if c t⊂ then

18: 1 1[ ] [ ] 1SCFIL k SCFIL k← +

19: end if

20: end for

21: 1k k← +

22: end for

23: >prune 1 1 1CFIL to generate FIL and FISC

24: for 1,k I← do

25: if 1[ ]SCFIL k minsupcount≥ then

26: k 1 1add c CFIL to FIL∈

27: 1 1add SCFIL [k] to FISC

28: end if

29: end for

30: if 1FIL ≠ ∅ then

31: FI

1add FIL to L

32: FISC

1add FISC to L

33: end if

34: 2k ←

35: while 1kFIL − ≠ ∅ do

36: k k -1CFIL Call GenerateCFIL(FIL )← > see Algorithm 5

37: for 1, .ki CFIL length← do > initialize the array kSCFIL to zero

38: [ ] 0kSCFIL i ←


86

39: end for

40: 1i ←

41: for all kcandidate c CFIL∈ do > find support count for every candidate

42: for all transaction t DB∈ do > scan DB

43: if c t⊂ then

44: 1 1[ ] [ ] 1SCFIL k SCFIL k← +

45: end if

46: end for

47: 1i i← +

48: end for

49: >prune kCFIL to generate

kFIL and kFISC

50: for 1, .ki SCFIL length← do

51: if [i]kSCFIL minsupcount≥ then

52: i k kadd c CFIL to FIL∈

53: k kadd SCFIL [i] to FISC

54: end if

55: end for

56: if kFIL ≠ ∅ then

57: FI

kadd FIL to L

58: FISC

kadd FISC to L

59: end if

60: 1k k← +

61: end while

62: FI &SCadd T to L

63: FI FI &SCadd L to L

64: FISC FI &SCadd L to L

65: FI &SC

isave L in the local file system of this site S

66: endCPUTime get system time←

67: end startCPUTime CPUTime CPUTime← −

68: iadd CPUTime to Result_S

69: iadd Result_S to Briefcase


71: NODESL get itinerary list from AgentProfile←

72: NODES NODESL remove first IP address from L← >visited site

73: NODESadd updated L to AgentProfile

74: if NODESL ≠ ∅ then > itinerary not empty

75: AObject new LGFIGA(AgentProfile, min_th_sup)←

76: add AObject to AgentProfile

77: NODEStransfer AgentProfile to DM_AEE at first IP address in L

78: else

79: endTripTime get system time for end of agent journey←

80: endadd TripTime to Briefcase


82: CENTRALtransfer AgentProfile to RM at S

83: end if

84: end procedure


87

Algortihm 5 GENERATECFIL

Input: 1,kL Frequent k -1 itemsets−

Output: kC , Candidate Frequent k itemsets

1: procedure GENERATECFIL (1kL −)

2: for all 1 k -1itemset l L∈ do

3: for all 2 k -1itemset l L∈ do

4: if 1 2 1 2 1 2(l [1] = l [1]) (l [2] = l [2]) (l [k - 1] = l [k - 1])∧ ∧ ∧L then

5: 1 2c l l← ⊗ > join step: generate candidates

6: end if

7: if HASINFREQUENTSUBSET (1, kc L −

) then > see Algorithm 6

8: delete c

9: else

10: kadd c to C

11: end if

12: end for

13: end for

14: return kC

15: end procedure

Algortihm 6 HASINFREQUENTSUBSET

Input: ,c Candidate k itemsets

Output: 1 1kL , Frequent k itemsets− −

1: procedure HASINFREQUENTSUBSET (1, kc L −

)

2: for all (k - 1) subset s c∈ do

3: if 1ks L −∉ then

4: return TRUE

5: else

6: return FALSE

7: end if

8: end for

9: end procedure

Algortihm 7 LOCAL KNOWLEDGE GENERATER AGENT (LKGA)

Input:

• AgentProfile,A collection of agent attributes set by the AL

• min_th_conf, the given minimum threshold confidence

Output: LSARL , the list of locally strong association rules

1: procedure LKGA( AgentProfile,min_th_conf )



4: FI &SC FI &SC

iL load L from local file system of this site S←

5: & . (0)FI SCT L get← >No. of records

6: & . (1)FI FI SCL L get← > frequent k-itemset list


88

7: & . (2)FISC FI SCL L get← > support count list

8: for 2, .FIk L size← do

9: . ( )FI

kL L get k← >get frequent k-itemset list

10: for all kl L∈ do

11: subsetsl generate all non - empty subsets of l←

12: FISC

spcountl get support count of l from L←

13: (l / T) 100support spcountAR ← × > support of the association rule

14: for all subsetsnon - empty subset s l∈ do

15: FISC

spcounts get support count of s from L←

16: conf spcount spcountAR (l / s )×100← > confidence of the association rule

17: if confAR min_th_conf≥ then

18: strong support confAR "s l - s[AR %,AR %]"← ⇒

19: print strongAR

20: strongadd l to AR

21: IP

i iS get IP address of this site S←

22: IP

i strongadd S to AR

23: LSAR

strongadd AR to L

24: end if

25: end for

26: end for

27: end for

28: LSAR

isave L in the local file system of this site S










38: AObject new LKGA(AgentProfile, min_th_conf)←



41: else





46: end if

47: end procedure


89

Algortihm 8 TOTAL FREQUENT ITEMSET COLLECTOR AGENT (TFICA)

Input: AgentProfile,A collection of agent attributes set by the AL

Output: FIL , the list of locally frequent itemsets

1: procedure TFICA( AgentProfile,min_th_conf )



4: FI &SC FI &SC


5: & . (1)FI FI SCL L get← > frequent k-itemset list

6: FI

iadd L to Result_S










16: AObject new TFICA(AgentProfile)←



19: else





24: end if

25: end procedure

Algortihm 9 LOCAL KNOWLEDGE COLLECTOR AGENT (LKCA)


Output: LSARL , the list of locally strong association rules

1: procedure LKCA( AgentProfile )



4: LSAR LSAR


5: LSAR

iadd L to Result_S







90





15: AObject new LKCA(AgentProfile)←



18: else





23: end if

24: end procedure

Algortihm 10 GLOBAL FREQUENT ITEMSET GENERATER AGENT (GFIGA)

Input: Briefcase, Result bag of TFICA agent

Output: GFIL , the list of global frequent itemsets

1: procedure GFIGA( Briefcase )


3: ( )nTFI FI

ii=1L retrieve total frequent itemsets L from Briefcase← U

4: ( )1

nGFI FI

iiL retrieve global frequent itemsets L from Briefcase

=← I

5: print GFIL

6: GFI

CENTRALsave L in the local file system of site S



9: print CPUTime

10: return GFIL

11: end procedure

Algortihm 11 GLOBAL KNOWLEDGE GENERATER AGENT (GKGA)

Input: Briefcase, Result bag of LKCA agent

Output: GSAR

CENTRALL , the list of globally strong association rules

1: procedure GKGA( Briefcase )


3: ( )nTLSAR LSAR

ii=1L retrieve total strong rules L from Briefcase← U

4: ( )GFI GFI

CENTRALL load global frequent itemsets L from S←

5: for all TLSAR

strongAR L∈ do

6: strongL get frequent itemset from AR←

7: if GFIL L∈ then


91

8: print IP

strong iAR along with the site address (S )

9: GSAR

strong CENTRALadd AR to L

10: end if

11: end for

12: GSAR

CENTRAL CENTRALsave L in the local file system of site S



15: print CPUTime

16: return GSAR

CENTRALL

17: end procedure

Algortihm 12 GLOBAL KNOWLEDGE DISPATCHER AGENT (GKDA)


Output: GSAR

CENTRAL iDispatch L at each distributed site S

1: procedure GKDA( AgentProfile )



4: GSAR GSAR

CANTRAL CENTRALL get L from Briefcase←

5: GSAR

CENTRAL isave L in the local file system of site S










15: AObject new GKDA(AgentProfile)←



18: else





23: end if

24: end procedure


92

Figure 2. Control Panel of AeMGSAR

4.IMPLEMENTATION AND PERFORMANCE STUDY

All the agents as well as control panel as shown in Figure 2 are designed in Java. Synthetic

dataset (iDB ) is stored across three distributed sites

1S , 2S and

3S , with 3500, 3850 and 3900

transactions and 10 items in each respectively using Transactional Data Set Generator (TDSG)

tool [16]. Binary and transactional versions of these datasets are shown in Appendix A. The

required configuration of the system is shown in Table 1 with additional deployment of DM_AEE

at each distributed site and AL and RM at CENTRALS . Round Trip time taken by various MAs is

shown in Figure 3. CPU time consumed by various MAs at site 1S , 2S and 3S is shown in Figure

4, Figure 5 and Figure 6, respectively. CPU time for GFIGA and GKGA is 101357102 nano

seconds and 33317458 nano seconds, respectively. ( )

FI

k iL and ( )

FISC

k iL at distributed sites generated by

LFIGA agent with 20% min_th_sup are shown in Appendix B.1, B.2 and B.3. LSAR

iL at distributed

sites generated by LKGA agent with 50% min_th_conf are shown in Appendix B.4, B.5 and B.6.

Globally frequent itemsets generated by GFIGA at CENTRALS is shown in Figure 7. Fifteen numbers

of 2-itemsets and eight number of 3-itemsets are globally frequent in TFI

kL list and 4, 5 and 6-

itemsets, which are locally frequent, are not globally frequent. Globally strong association rules

( GSAR

CENTRALL ) generated by GKGA at CENTRALS for globally frequent 3-itemsets are shown in Figure 8

and GSAR

CENTRALL for 2-itemsets are shown in Appendix B.7.

On comparing this system with the traditional central data warehouse (DW) based approach for

ARM where entire data from the distributed sites is centrally collected in a DW [17], it is found

that the storage cost is reduced as data is mined locally and only the resultant knowledge is

carried at the central site by mobile agents. As size of the resultant data carried across by mobile


93

agents is small so network communication cost is also reduced in this case. Data mining is

performed locally by agents, so computational cost at central site is also minimised. AeMGSAR

reflects the global knowledge because all the strong association rules generated are also strong at

each distributed site. The system relies upon the Java's in-built security system. As MAs are scalable in nature so performance would not be affected by adding more sites.

Table 1. Network Configuration

Site Name Processor OS LAN Configuration

IP a Network

SCENTRAL Intel b MS

c 192.168.46.5 NW

d

S1 Intel b MS c 192.168.46.212 NW d

S2 Intel b MS

c 192.168.46.189 NW

d

S3 Intel b MS

c 192.168.46.213 NW

d

a. IP address with Mask: 255.255.255.0 and Gateway 192.168.46.1

b. Intel Pentium Dual Core(3.40 GHz, 3.40 GHz) with 512 MB RAM

c. Microsoft Windows XP Professional ver. 2002

d. Network Speed: 100 Mbps and Network Adaptor: 82566DM-2 Gigabit NIC

Figure 3. Round Trip time taken by various MAs

Figure 4. CPU Time taken by various MAs at site 1S


94



Figure 7. Lists of global frequent k-itemsets at CENTRALS


95

Figure 8. Globally strong association rules for globally frequent 3-itemsets

5.CONCLUSION

Mobile agents strongly qualify for designing distributed applications and the amalgamation of

DDM and agent technology gives favourable results. Most of the existing agent based

frameworks for DARM task are only prototype model and lacks the appropriate underlying

execution environment, scalability, privacy preserving techniques, global knowledge generation

and implementation using a real datasets. In this study, a scalable MAS, called Agent enabled

Mining of Globally Strong Association Rules (AeMGSAR), is presented based on the serial

itinerary of the mobile agents. In this system the overall task of mining the globally strong

association rules is divided into subtasks which are handled by various mobile as well as

stationary agents. An AEE is also designed for the implementation and performance study of

AeMGSAR system. Serial itinerary used for mobile agent migration increases the overall cost of

DARM task so a parallel computing model could be designed where clones of each mobile agent

is dispatched in parallel to all distributed sites.

REFERENCES [1] U. M. Fayyad, G. Piatetsky-Shapiro, P. Smyth & R. Uthurusamy, (1996) Advances in Knowledge

Discovery and Data Mining, AAAI/MIT Press.

[2] J. Han & M. Kamber, (2006) Data Mining: Concepts and Techniques, 2nd ed. Morgan Kaufmann.

[3] G. S. Bhamra, R. B. Patel & A. K. Verma, (2014) “Intelligent Software Agent Technology: An

Overview”, International Journal of Computer Applications (IJCA), vol. 89, no. 2, pp. 19–31.

[4] R. Agrawal, T. Imielinski & A. Swami, (1993) “Mining association rules between sets of items in large

databases”, in Proceedings of the ACM-SIGMOD International Conference of Management of Data,

pp. 207–216.

[5] R. Agrawal & J. C. Shafer, (1996) “Parallel mining of association rules”, IEEE Transaction on

Knowledge and Data Engineering, vol. 8, no. 6, pp. 962–969.

[6] M. J. Zaki, (1999) “Parallel and distributed association mining: a survey”, IEEE Concurrency, vol. 7,

no. 4, pp. 14–25.

[7] X. Wu & S. Zhang, (2003) “Synthesizing high-frequency rules from different data sources”, IEEE

Transactions on Knowledge and Data Engineering, vol. 15, no. 2, pp. 353–367.


96

[8] Y.-L. Wang, Z.-Z. Li & H.-P. Zhu, (2003) “Mobile agent based distributed and incremental techniques

for association rules”, in Proceedings of the International Conference on Machine Learning and

Cybernetics(ICMLC 2003), vol. 1, pp. 266–271.

[9] C. Aflori & F. Leon, (2004) “Efficient Distributed Data Mining using Intelligent Agents”, in

Proceedings of the 8th International Symposium on Automatic Control and Computer Science, pp. 1–

6.

[10] U. P. Kulkarni, P. D. Desai, T. Ahmed, J. V. Vadavi & A. R. Yardi, (2007) “Mobile Agent Based

Distributed Data Mining”, in Proceedings of the International Conference on Computational

Intelligence and Multimedia Applications (ICCIMA 2007), IEEE Computer Society, pp. 18–24.

[11] G. Hu & S. Ding, (2009a) “An Agent-Based Framework for Association Rules Mining of Distributed

Data”, in Software Engineering Research, Management and Applications 2009, ser. Studies in

Computational Intelligence, R. Lee and N. Ishii, Eds. Springer Berlin - Heidelberg, vol. 253, pp. 13–

26.

[12] G. Hu & S. Ding, (2009b) “Mining of Association Rules from Distributed Data using Mobile

Agents,” in Proceedings of the International Conference on e-Business(ICE-B 2009), pp. 21–26.

[13] A. O. Ogunde, O. Folorunso, A. S. Sodiya, J. A. Oguntuase & G. O. Ogunleye, (2011) “Improved

cost models for agent based association rule mining in distributed databases”, Anale SEria

Informatica, vol. 9, no. 1, pp. 231–250, Available: http://anale-

informatica.tibiscus.ro/download/lucrari/9-1-20-Ogunde.pdf

[14] G. S. Bhamra, A. K. Verma, & R. B. Patel, (2015) “Agent Based Frameworks for Distributed

Association Rule Mining: An Analysis”, International Journal in Foundations of Computer Science &

Technology (IJFCST), vol. 5, no. 1, pp. 11-22.

[15] R. Agrawal & R. Srikant, (1994) “Fast Algorithms for Mining Association Rules in Large Databases”,

in Proceedings of the 20th International Conference on Very Large Data Bases (VLDB’94). Morgan

Kaufmann Publishers Inc., pp. 487–499.

[16] G. S. Bhamra, A. K. Verma, & R. B. Patel, (2011) “TDSGenerator: A Tool for generating synthetic

Transactional Datasets for Association Rules Mining”, International Journal of Computer Science

Issues (IJCSI), vol. 8, no. 2, pp. 184-188.

[17] G. S. Bhamra, A. K. Verma, & R. B. Patel, (2014) “An Investigation into the Central Data Warehouse

based Association Rule Mining”, International Journal of Computer Applications (IJCA), vol. 96, no.

10, pp. 1-12.

AUTHORS

Gurpreet Singh Bhamra is currently working as Assistant Professor at

Department of Computer Science and Engineering, M. M. University, Mullana,

Haryana. He received his B.Sc. (Computer Sc.) and MCA from Kurukshetra

University, Kurukshetra in 1995 and 1998, respectively. He is pursuing Ph.D.

from Department of Computer Science and Engineering, Thapar University,

Patiala, Punjab. He is in teaching since 1998. He h as published 13 research

papers in International/National Journals and International Conferences. He has

received Best Paper Award for “An Agent enriched Distributed Data Mining on

Heterogeneous Networks”, in “Challenges & Opportunities in Information

Technology” (COIT-2008). He is a Life Member of Computer Society of India. His research interests are in

Distributed Computing, Distributed Data Mining, Mobile Agents and Bio-informatics.

Dr. Anil Kumar Verma is currently working as Associate Professor at

Department of Computer Science & Engineering, Thapar University, Patiala. He

received his B.S., M.S. and Ph.D. in 1991, 2001 and 2008 respectively, majoring in

Computer science and engineering. He has worked as Lecturer at M.M.M.

Engineering College, Gorakhpur from 1991 to 1996. He joined Thapar Institute of

Engineering & Technology in 1996 as a Systems Analyst in the Computer Centre

and is presently associated with the same Institute. He has been a visiting faculty to

many institutions. He has published over 100 papers in referred journals and

conferences (India and Abroad). He is a MISCI (Turkey), LMCSI (Mumbai),

GMAIMA (New Delhi). He is a certified software quality auditor by MoCIT,


97

Govt. of India. His research interests include wireless networks, routing algorithms and securing ad hoc

networks and data mining.

Dr. Ram Bahadur Patel is currently working as Professor and Head at Department

of Computer Science & Engineering, Chandigarh College of Engineering &

Technology, Chandigarh. He received PhD from IIT Roorkee in Computer Science &

Engineering, PDF from Highest Institute of Education, Science & Technology

(HIEST), Athens, Greece, MS (Software Systems) from BITS Pilani and B. E. in

Computer Engineering from M. M. M. Engineering College, Gorakhpur, UP. Dr.

Patel is in teaching and research since 1991. He has supervised 36 M. Tech, 7 M.

Phil. and 8 PhD Thesis. He is currently supervising 6 PhD students. He has published

130 research papers in International/National Journals and Refereed International

Conferences. He has written 7 text books for engineering courses. He is member of

ISTE (New Delhi), IEEE (USA). He is a member of various International Technical Committees and

participating frequently in International Technical Committees in India and abroad. His current research

interests are in Mobile & Distributed Computing, Mobile Agent Security and Fault Tolerance and Sensor

Network.

APPENDIX A – SYNTHETIC DATASETS

A.1 BDS3500T10I.txt and corresponding TDS3500T10I.txt(

1DB ) at site 1S

These synthetic binary and transactional datasets of 3500 records are created by TDSG tool at

site1S . In the binary version each column head represents the item number and each row

represents a transaction where integer ‘1’ is used for a purchased item and ‘0’ is used if it is nor

purchased. The corresponding transactional version has a Transaction It (TID) for each

transaction and Itemset is the set of all the purchased items for that particular transaction.


98

A.2 BDS3850T10I.txt and corresponding TDS3850T10I.txt(2DB ) at site

2S

These synthetic binary and transactional datasets of 3850 records are created by TDSG tool at site

2S .

A.3 BDS3900T10I.txt and corresponding TDS3900T10I.txt(3DB ) at site

3S

These synthetic binary and transactional datasets of 3900 records are created by TDSG tool at site

3S .


99

APPENDIX B–RESULTANT KNOWLEDGE OF AEMGSAR

SYSTEM

B.1 (1)

FI

kL and (1)

FISC

kL at site 1S

List of frequent k-itemset, i.e., (1)

FI

kL is represented by column L and column SC shows the support

count of the corresponding frequent k-itemset, i.e., (1)

FISC

kL at site 1S . These frequent itemsets and

their support counts are obtained by processing the synthetic dataset (1DB ) as shown in Appendix

A.1.

B.2 (2)

FI

kL and (2)

FISC

kL at site 2S

These frequent itemsets and their support counts are obtained by processing the synthetic dataset

(2DB ) as shown in Appendix A.2.


100

B.3 (3)

FI

kL and (3)

FISC

kL at site 3S

These frequent itemsets and their support counts are obtained by processing the synthetic dataset

( 3DB ) as shown in Appendix A.3.


101

B.4 1

LSARL at site

1S

Column L represents frequent k-itemset and column AR(support, confidence) shows the list of

locally strong association rules, i.e., 1

LSARL at site

1S . Each strong rule has its associated support

and confidence factor. The minimum threshold is taken as 20% and minimum threshold

confidence as 50% for generating the strong rules by making use of the data as shown in

Appendix B.1.


102

B.5 2

LSARL at site

2S



LSARL at site 2S . Each strong rule has its associated support



Appendix B.2.


103

B.6 3

LSARL at site

3S



LSARL at site 3S . Each strong rule has its associated support



Appendix B.3.


104

B.7 GSAR

CENTRALL at site CENTRALS

Column L represents globally frequent k-itemset, i.e., itemsets which are locally strong at all the

distributed sites and column AR(support, confidence) shows the list of globally strong

association rules, i.e., GSAR

CENTRALL for such itemsets. Each globally strong rule has its associated

support and confidence factor. The minimum threshold is taken as 20% and minimum threshold

confidence as 50%. Site represents the IP address of the site where the rule is locally strong. IP

address 192.168.46.212 is used for site 1S , 192.168.46.189 for site

2S and address

192.168.46.213 is used for site 3S .

Date post:	06-Aug-2015
Category:	Engineering
Upload:	ijcsa
View:	25 times
Download:	2 times