DISTRIBUTED COMMUNITY COOPERATION
IN MULTI AGENT FILTERING FRAMEWORK
Sahin Albayrak, Dragan Milosevic
DAI-Lab, Technical University Berlin, Germany, Salzufer 12, 10587 Berlin, Germany
Keywords: Cooperation in multi agent systems, Distributed inform
ation retrieval, Intelligent information agents, Coop-
erative filtering communities, Self improving cooperation, Filtering framework, Recommendation systems
Abstract: In nowadays easy to produce and publish information society, filtering services have to be able to simulta-
neously search in many potentially relevant distributed sources, and to autonomously combine only the best
found results. Ignoring a necessity to address information retrieval tasks in a distributed manner is a major
drawback for many existed search engines which try to survive the ongoing information explosion. The es-
sence of a proposed solution for performing distributed filtering is in both installing filtering communities
around information sources and setting a comprehensive cooperation mechanism, which both takes care
about how promising is each particular source and tries to improve itself during a runtime. The applicability
of the presented cooperation among communities is illustrated in a system serving as intelligent personal in-
formation assistant (PIA). Experimental results show that integrated cooperation mechanisms successfully
eliminate long lasting filtering jobs with duration over 1000 seconds, and they do that within an acceptable
decrease in feedback and precision values of only 3% and 6%, respectively.
1 INTRODUCTION
With an abundance of electronically available in-
formation, finding only a relevant one can amount to
a real challenge. Existed search and retrieval engines,
such as Google (Brin, 1998), Yahoo, AltaVista, and
many others (Saurabh, 2001) provide more capabili-
ties today then ever before, but the information that
is potentially available from World Wide Web con-
tinues to grow exponentially (Mohammadian, 2004).
There is unfortunately an open doubt that these cen-
tralised search engines will not be able to adequately
respond to this information explosion in the future.
A distributed knowledge discovery obviously be-
comes the only possible way to cope with these in-
formation overload problems.
While the needed information is usually scattered
i
n a vast number of distributed sources, a typical
user has no time to look all around, and its attention
has become a precious resource (Yang, 2004). Indi-
vidual inspection of each available source is imprac-
tical, and the tools to manage these sources effec-
tively will become critical (Blake, 2001). Without an
approach for identifying sources that potentially
have relevant documents, information rich sources
are almost useless. One obviously needs new tech-
nologies that will provide means for locating, re-
trieving and processing of data, being stored in nu-
merous distributed sources. Authors’ point of view
is that a comprehensive cooperation among sources
can be exactly one instance of a needed technology.
What is also very important, these cooperation
mech
anisms can make usable many filtering strate-
gies (Balabanovic, 1997)(Boley, 1998)(
Tauritz, 2002)
(Delgado, 2000) (
Michalewicz, 2000) (Zamir, 1997)
(Oard, 1996) and data mining methods (Han, 2001)
that are always more or less scalable. The design and
structure of many strategies may not be appropriate
at all for a very large when for instance the run times
behave exponentially to underlying collection size.
Two illustrative examples can be simultaneous clus-
tering with a dynamic keyword weighting (Frigui,
2004) and a self organising map (Kohonen, 2000).
The former gives remarkable good results, but un-
fortunately only on a collection with two thousand
documents. The later is applied on 7 million short
abstracts, but it takes 6 weeks on 6 processor com-
puter to train SOM. For these two and many other
strategies, it is crucial to have collections with de-
sired properties, where they can give excellent re-
sults. The problem that arises with many collections
is concerned with the means to organise them some-
how, and hopefully this challenge can be address
through the same cooperation mechanisms.
400
Albayrak S. and Milosevic D. (2005).
DISTRIBUTED COMMUNITY COOPERATION IN MULTI AGENT FILTERING FRAMEWORK.
In Proceedings of the Seventh International Conference on Enterprise Information Systems, pages 400-407
DOI: 10.5220/0002538004000407
Copyright
c
SciTePress
To support information retrieval from the dis-
tributed sources, and to help to many filtering strate-
gies that are only small scale applicable, distributed
cooperation mechanisms are essential, as it is going
to be shown in the rest of this paper, being structured
as follows. The next section illustrates cooperation
problems through one scenario. The core of this pa-
per is then contained in the section, which gives
principles being in the basis of the used cooperation
approach that is naturally separated to estimating,
dispatching, composing and adapting steps. A paper
is finished with sections where implementation and
experimental details are shortly presented.
2 PROBLEM DESCRIPTION
The unavoidable consequence of the nowadays
overall information overload is that an open problem
becomes answering where the relevant information
is deployed. Internet users are forced to manually
make decisions about the most promising sources for
retrieving the desired information, being usually a
time consuming activity. Obviously, they are wast-
ing plenty of their valuable time when they are
searching for the needed information at the wrong
places. The following scenario, being separated into
estimating, dispatching, composing and adapting
steps, is going to illustrate challenges which any
comprehensive cooperation mechanisms have to
address in order to overcome the mentioned infor-
mation retrieval problems.
Figure 1 gives a playing ground for a scenario,
where this playing ground is composed of five dif-
ferent data sources
, , around which
distributed filtering communities
are installed.
Each and every community has one manager (
)
agent that is responsible for every single cooperation
activity. In order to enables cooperation, manager
is representing all other communities through
their descriptions
,
i
DB
}5,...,1{i
i
FC
i
M
i
M
j
CD
ji
, which illustrate their
underlying content. For example
has , ,
and as descriptions of corresponding
communities with whom cooperation is possible.
3
M
1
CD
2
CD
4
CD
5
CD
[Estimating] The scenario begins by estimating how
good each community can be for a particular request.
This estimation naturally assumes the additional
usage of a past experience, which shows how reli-
able each community has been in providing the rele-
vant results. The idea is to somehow make a balance
between what a particular community says about
itself and the lessons that have been learnt through
the previous cooperation with it. The dilemma,
which has to be addressed through the estimation
step, is concerned with a deciding either to use for a
particular job currently better suited communities or
to rely more on communities that have better satis-
fied user needs in the past.
Figure 1: Cooperation playing ground
[Dispatching] By using formed estimations about
how promising are communities, a decision to dis-
patch a job to some of them should be made. But,
not all of jobs are the same, and not for all of them
the same coverage concerning the found estimations
exists. For some jobs, many communities can pro-
vide quite good results. On the other hand, very spe-
cific jobs can be well processed only by specialised
communities, and asking others is the pure waste of
resources. A decision, how many communities
should contribute, is hopefully connected with how
good are estimations. It is maybe reasonable not to
dispatch a job to a community with a bad estimation,
but it is necessary to ensure that every job will be
processed at least by somebody.
[Composing] The last piece of a puzzle, known as
finding the needed information through a distributed
filtering, is concerned with putting all found results
together. In this composing activity, being per-
formed by the manager that was initially chosen by a
user, not only the quality of results but also the
community successfulness in the past should be
DISTRIBUTED COMMUNITY COOPERATION IN MULTI AGENT FILTERING FRAMEWORK
401
taken into consideration. The problem is to find out
how to compare the following two results. The one
was found by not so reliable community, but for that
result the responsible community says that it is ex-
actly what the user is expecting. The other is found
by the community that is known as a very successful
one, but at the same time that community is declar-
ing that this particular result has weak chances to
satisfy user’s needs. In other words, the question is
how to combine by community predicted result rele-
vance with a reliability of that particular community.
[Adapting] As soon as user feedback about the real
result relevance is received, the measure of commu-
nity successfulness in finding accurate results can be
adapted. The ultimate goal of these adaptation ac-
tivities, is establishing a more realistic picture about
the potentials of available communities. The final
effect is that a system is hopefully going to learn to
even better do cooperation among communities for
future filtering jobs.
3 APPROACH
The authors’ point of view on satisfying information
needs is concerned with the fact that the asked in-
formation is usually scattered around many different
distributed sources, and that a real challenge be-
comes both finding which sources should be
searched for a particular request and putting together
found results. These challenges are addressed
through the installation of at least one so-called fil-
tering community around each and every informa-
tion source and by setting up sophisticated coopera-
tion mechanisms between communities.
System architecture is given on Figure 2. User
agent (U) is responsible for the creation of filtering
jobs by collecting the user preferences. It also knows
how a user feedback can be obtained and forwarded
to a manager agent. Manager agent (M) is the cor-
nerstone that fulfils all cooperation activities and
ensures the satisfied quality of filtering services. It
should be seen as the entity that first performs the
estimation of sources in order to be able to dispatch
the received filtering job to the right communities.
As soon as the activated communities have produced
results, manager will then compose the final result
set that will be returned back to the user agent. In the
case of receiving any feedback from the user agent
about the result relevance, manager agent will per-
form the adaptation of knowledge that it has about
the responsible communities.
Figure 2: Estimating, dispatching, composing and adapting
cooperation steps
The ultimate goal of the deployed cooperation
mechanisms is always to find which communities
are the most promising for providing results to the
received request. This can be achieved through esti-
mating communities, dispatching request, compos-
ing results and adapting reliability steps, as it is go-
ing to be shown in the following sub-sections.
In order to make the ongoing discussions more
understandable, terms filtering community, coordi-
nation, cooperation, filtering request and community
description will be defined as follows:
Def. 1. Filtering community
is a collection of
many different filtering agents that are tailored to
efficiently do searching on the underling collection
of objects. Instead of having only filtering agents,
each and every community has also one so-called
manager agent that is mainly responsible for per-
forming coordination and cooperation tasks.
FC
Def. 2. Coordination is a comprehensive activity,
being performed by a manager agent with an ulti-
mate goal to find which filtering agent from its
community is the most promising for doing filtering
in a current system runtime situation. The way how
one such coordination can be achieved is out of
scope of this paper, and it is addressed in (Albayrak,
2004b)(Albayrak, 2004c).
Def. 3. Cooperation is performed between manager
agents in order to find the most competent commu-
nities for processing the received request. This paper
is focused on showing how that can be achieved.
Def. 4. Filtering request
F
R
describes user prefer-
ences towards the documents that will satisfy the
imposed information needs in the best manner. Each
filtering request can be formally represented as the
collection of pairs
, where corresponds to
},{
)(r
ii
t
ω
i
t
ICEIS 2005 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS
402
a term in a request, and represents the impor-
tance of the corresponding term
. While larger
positive value of
means that it is more impor-
tant that
is found, larger negative value of
assumes that it is more important that
is not pre-
sent in the analysed document.
)(r
i
ω
i
t
)(r
i
ω
i
t
)(r
i
ω
i
t
Def. 5. Community description
represents cor-
responding community in concise and descriptive
way both to ensure an efficient exchange of descrip-
tions between communities and at the same time to
provide enough information to other communities
that want to cooperate with it. Formally, community
description
is , where CCD is a de-
scription of the underlying collection of documents
and
is community reliability in processing filter-
ing tasks in the past. Community content description
is a collection of given number of the most
important pairs
, where corresponds to a
particular term in a collection, and
is the num-
ber of documents from the underlying collection that
have term
.
CD
CD
},{
C
rCCD
C
r
CCD
},{
)(d
ii
t
ω
i
t
)(d
i
ω
i
t
3.1 Estimating Communities
The usage of the appropriate distance function is a
critical point in estimating how successful each fil-
tering community can be in processing the actual
request. Because both filtering request and commu-
nity content description are defined in highly sparse
space, the appropriate modification of the weighted
Jaccard index (
Michalewicz, 2000)(Han, 2001) is ex-
pected to give the distance function with a desired
behaviour. Formally, filtering request
and community description
can be
compared as:
}),({
)(r
ii
tFR
ω
}),({
)(d
ii
tCCD
ω
=
FRi
r
i
CCDandFRbothi
r
i
d
i
WJ
arctg
CCDFRd
)(
)()(
)(
2
),(
ω
ωβω
π
.
While weights
, being present in a commu-
nity description, always take only positive values,
weights
from a request can be negative. Such
negative
are bound to unwanted terms, and
always when unwanted term from a request is pre-
sent in a community description, that community
pays a penalty. The penalty is larger when unwanted
term is more important in a community description,
i.e.
will be reduced more when is larger
and at the same time
.
)(d
i
ω
)(r
i
ω
)(r
i
ω
WJ
d
)(d
i
ω
0
)(
<
r
i
ω
The introduced penalties should facilitate the se-
lection of more specialised communities, always
when such communities are available. In the case
where unwanted terms are present in a particular
community description, the probability that good
recommendations will be found by such community
is small. This holds because of the assumption that
community description contains the most representa-
tive terms from the underlying document collection.
The
arctg
function is used to bound values for
always to
)(d
i
ω
]
2
,0[
π
, where a tuning parameter
β
con-
trols the level of translation. Larger
β
means, that
value
2
π
is reached faster, and reverse. Such
bounding is necessary because of a
nature of
representing a number of documents that have corre-
sponding term
, that can be a ordinary big number.
)(d
i
ω
i
t
A denominator in
expression ensures that
WJ
d
]1,1[
WJ
d
, where larger means that particular
community has more promising content for process-
ing particular request. It is expected that communi-
ties, having in their content descriptions a lot of
wanted terms and omitting unwanted ones, will have
larger
values.
WJ
d
WJ
d
3.2 Dispatching Request
The main cooperation objective is to dispatch the
actual request only to the potentially good filtering
communities, being the ones that both have access to
the most relevant needed information and at the
same time have acceptable reliability in processing
past filtering tasks. While the idea about which
communities have the best underlying documents
can be assessed through the usage of in the previous
step found
values, each community description
has the reliability , showing how good the
corresponding community has performed in the past.
WJ
d
CD
C
r
The used dispatching principle is to send the cur-
rent request to
k communities, which have the best
combination of
and values. One simple so-
lution of combining
and values, is to define
so-called community promise-ness
as:
WJ
d
C
r
WJ
d
C
r
C
p
rd
CrWJd
C
rd
p
ββ
ββ
+
+
=
,
DISTRIBUTED COMMUNITY COOPERATION IN MULTI AGENT FILTERING FRAMEWORK
403
where tuning parameters
d
β
and
r
β
control the
influence of
and in making a final judge-
ment about how promising is a particular community.
According to the fact that
and take values
from
and , respectively, it follows that
, where bigger means that a particu-
lar community is more promising.
WJ
d
C
r
WJ
d
C
r
]1,1[ ]1,0[
]1,1[
C
p
C
p
The unnecessary loading of not promising
enough communities is avoided by additionally re-
questing that each community, being a candidate for
activation, has to have at least given minimal prom-
ise-ness. More precisely, the number of communities
that will receive a dispatched request is flexible, and
it is limited with
and , i.e.
min
k
max
k
maxmin
kkk
.
While in the case where many promising communi-
ties exist, at most
will be activated, in a oppo-
site case at least
communities will ensure that
at least somebody will try to find recommendations.
max
k
min
k
3.3 Composing Results
The fact, that multiple communities work together,
comes to the point when the results, being produced
by different communities, should be put together in
order to create the unique set of recommendations. A
community, having received the original request and
having initiated cooperation, will collect all the re-
sults, being found by different communities, and will
decide which results are good enough to be returned
as recommendations. In that way the performed co-
operation is completely transparent for the sender of
a filtering job, i.e. user agent does not think about
where the retrieved data were deployed.
It is assumed that each result comes together
with the predicted quality
, showing how good is
a particular result, and being a number from
. These qualities are set by communities
that have found corresponding results, and conse-
quently the community that is putting results to-
gether does not have any influence on them. In order
to protect from the malicious communities, saying
that their results are always the perfect ones, the
community reliability
is also taken into account
when composing results. Instead of ranking results
based only on their quality
, a product is
used to better assess the real quality, and the asked
number of results with the largest
values will
be chosen. Results, being found by communities
with low reliability, will actually pay penalties and
reduce their chances to be included in a final rec-
ommendation set.
p
q
]%100,0[
C
r
p
q
pC
qr
pC
qr
3.4 Reliability Adaptation
After completing a recommendation set and receiv-
ing a user feedback about the actual relevance of
found results, learning through reliability value ad-
aptation takes place in order to ensure that the as-
signed reliability value reflects as accurate as possi-
ble corresponding filtering community ability to
satisfy the imposed information needs. The adapta-
tion of
is based on the comparison between by
filtering community predicted result relevance
and the actual relevance
, being generated from a
user feedback (Figure 3). The used adaptation rule
can be expressed as
C
r
p
q
a
q
12
))((
+
=
k
paCC
qqtlr
εγ
In the last expression
ε
is a tolerance, which de-
fines how close the predicted relevance of results
should be to the actual relevance in order to reward
the responsible filtering community,
k in-
creases the influence of large
deviations from
,
)0( >k
a
q
p
q
C
γ
is a tuning parameter, and
l
is a
decreasing learning rate which insures that already
learnt reliability value will not be easily destroyed.
The filtering communities with the solid history will,
according to such learning rate, change their reliabil-
ity values in smaller extent than the novel ones.
t
t
elt
γ
=
0
)(
Figure 3: Adaptation of community reliability
C
r
While a reward is limited to , a penalty
for really bad estimations of
, being quite differ-
ent from
, is theoretically unlimited. In reality,
penalties are also limited because
and take
values from
, which gives
12
0
+k
C
l
εγ
p
q
a
q
a
q
p
q
]%100,0[
%100
pa
qq
.
4 IMPLEMENTATION
Filtering communities are implemented in JIAC IV
(Java Intelligent Agent Component-ware, Version
IV) being a comprehensive service-ware framework
for developing and deploying agent systems cover-
ing design methodology and tools, agent languages
and architecture, a FIPA compliant infrastructure,
ICEIS 2005 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS
404
management and security functionality, and the ge-
neric scheme for the user access (Fricke, 2001). The
agent system JIAC IV is based on CASA (Compo-
nent Architecture for Service Agents) having a
modular internal structure consisting of an open set
of components that can be adjusted to different re-
quirements at the design-time as well at the run-time
(Sesseler, 2002). This flexibility facilitates not only
the easy integration of new agents inside filtering
communities, but also enables experiments with dif-
ferent cooperation mechanisms.
In order to estimate in a real time conditions both
the flexibility of JIAC IV service framework and the
applicability of a self adapting cooperation among
filtering communities, comprehensive personal in-
formation assistant (PIA) has been developed. A
typical user is guided by PIA starting from a collec-
tion of both information interests and delivery pref-
erences to a relevant article presentation and feed-
back gathering. The authors’ believe is that PIA is
going to demonstrate in real conditions the applica-
bility of both agent-based filtering systems and co-
operation mechanisms being situation aware.
5 EXPERIMENTAL RESULTS
The expected benefit of the presented cooperation
approach should be found first in making possible to
install cooperative communities around many small
sub-domains and then in using many wonderful fil-
tering strategies, which are only small-scale applica-
ble. Such a distributed system should manage to
retrieve data almost as good as the centralised one,
however together with the great reduction of filter-
ing time. While PIA system is going to be used for
making comparisons in response time and user feed-
back domains, one small simulated environment will
give the comparisons of precision and recall values.
5.1 Long Lasting Job Elimination
A trial to escape long lasting filtering jobs, being
usually a consequence of the small scale strategy
applicability inside a single community, was a main
motivation for the realisation of multiple communi-
ties that are mutually organised through the pre-
sented self adapting cooperation. Even though these
long lasting jobs will probably produce perfect re-
sults in next few hours, to obtain nearly perfect re-
sults within few minutes is usually much more ap-
propriate. Because a final judgement, concerning
such trade-off statements, is always given by the
user, this section gives comparisons between PIA,
based on a centralised single community, and being
based on multiple cooperative communities, in both
user feedback and response time domains. As a per-
fect test environment, PIA system has been chosen
mostly because it currently supports more than 120
different web sources, grabs daily around 3 thousand
new semi-structured and unstructured documents,
has almost 500 thousand already pre-processed arti-
cles, and actively helps to 26 workers from authors’
laboratory in their information retrieval activities.
Before the 18
th
of July 2004, PIA was working
without cooperative communities, and the 37 last
received feedback and corresponding response time
values are given on Figure 4 and Figure 5.
Figure 4: Feedback without community cooperation
Figure 5: Response time without community cooperation
After the 18
th
of July 2004, PIA is working with
five cooperative communities and in the first ten
days feedback values were received for 37 jobs.
Figure 6 and Figure 7 respectively present these
feedback and corresponding response time values.
Figure 6: Feedback with community cooperation
Figure 7: Response time with community cooperation
The given figures clearly show that while the in-
tegration of multiple cooperative communities does
not significantly affect user feedback (Figure 4 and
Figure 6 show only a slight feedback value decrease
that is hopefully within 3%), it successfully elimi-
nates long lasting jobs (7 problematic long lasting
jobs marked with circles on Figure 5, having a re-
sponse time that is longer than 1000s, do not occur
anymore on Figure 7). While an even bigger fluctua-
tion in quality was not detected by users probably
DISTRIBUTED COMMUNITY COOPERATION IN MULTI AGENT FILTERING FRAMEWORK
405
because they were satisfied with a shorter waiting
time, what is more important, by integrating coop-
erative multiple communities, PIA can provide fil-
tering services on significantly larger collections.
5.2 Precision Versus Recall
The ability of the presented cooperation mechanisms
to successfully find communities, which will provide
the most relevant results for the actual request, can
be best assessed by using precision (
), recall (
p
)
and fallout (
) values (Han, 2001), being broadly
accepted measures for the comparison of different
algorithms in the area of information retrieval. On
the one hand, both for precision, being the propor-
tion of retrieved documents that are relevant, and for
recall, representing the proportion of relevant docu-
ments that are retrieved, greater values correspond to
a system with better properties. On the other hand,
fallout relates to the proportion of irrelevant docu-
ments that are retrieved, and consequently smaller
values are preferred. Formally, precision, recall and
fallout measures are defined as
,
, , where
, , and correspond to either re-
trieved
or not retrieved documents that are
either relevant
or irrelevant as in Table 1.
f
)/(
)()()( r
ir
r
r
r
r
nnnp +=
)/(
)()()( nr
r
r
r
r
r
nnnr += )/(
)()()( nr
ir
r
ir
r
ir
nnnf +=
)(r
r
n
)(r
ir
n
)(nr
r
n
)(nr
ir
n
)(r
n
)(nr
n
r
n
ir
n
Table 1: , , and definitions
)(r
r
n
)(r
ir
n
)(nr
r
n
)(nr
ir
n
relevant irrelevant
retrieved
)(r
r
n
)(r
ir
n
not retrieved
)(nr
r
n
)(nr
ir
n
To compare the information retrieval system
based on a centralised single community with the
one that has multiple cooperative communities, a
small controlled domain
with only 500
documents is formed. For each of 10 testing filtering
requests
at least 10 and at most 20 docu-
ments are manually selected as the relevant ones. In
the first scenario, requests
will be resolved
by using a system that has only one centralised
community. That centralised community will use as
the underlying document collection
and will
return 10 results for each request from
. The
precision, recall and fallout values, corresponding to
such centralised system, are given in Table 2.
)500(
all
D
10
1
}{
=ii
FR
10
1
}{
=ii
FR
)500(
all
D
10
1
}{
=ii
FR
Table 2: Precision, recall and fallout for the system with
single centralised community for each of 10 request from
, where the last row gives the average values
10
1
}{
=ii
FR
)(r
r
n
)(nr
r
n
)(r
ir
n
)(nr
ir
n
p
r
f
8 12 2 478 80 40 0.42
7 11 3 479 70 38.89 0.62
7 5 3 485 70 58.33 0.61
8 7 2 483 80 53.33 0.41
5 11 5 479 50 31.25 1.03
7 5 3 485 70 58.33 0.61
9 8 1 482 90 52.94 0.20
6 4 4 486 60 60 0.82
8 6 2 484 80 57.14 0.41
7 6 3 484 70 53.84 0.62
7.2 7.5 2.8 482.5 72 50.41 0.58
In the second scenario, domain
D
is manu-
ally split on agent technology
, telecommuni-
cation
, renewable energy , sport news
and political news domains, where the
index in exponent says how many documents belong
to the particular sub-domain. Around every sub-
domain a separate community is installed, and the
same 10 filtering request
from the first sce-
nario are again resolved by using such a system with
5 distributed communities. The obtained precision,
recall and fallout values are given in Table 3.
)500(
all
)108(
at
D
)83(
t
D
)147(
re
D
)87(
sn
D
)75(
pn
D
10
1
}{
=ii
FR
Table 3: Precision, recall and fallout for the system with 5
distributed communities for each of 10 request from
, where the last row gives the average values
10
1
}{
=ii
FR
)(r
r
n
)(nr
r
n
)(r
ir
n
)(nr
ir
n
p
r
f
7 13 3 477 70 35 0.63
7 11 3 479 70 38.89 0.62
6 6 4 484 60 50 0.82
6 9 4 481 60 40 0.82
5 11 5 479 50 31.25 1.03
7 5 3 485 70 58.33 0.61
8 9 2 481 80 47.06 0.41
6 4 4 486 60 60 0.82
7 7 3 483 70 50 0.62
7 6 3 484 70 53.85 0.62
6.6 8.1 3.4 481.9 66 46.44 0.70
As it has been expected, the centralised system
has slightly better precision, recall and fallout values
than the distributed one. While the average precision
and recall values have decreased for 6% and 4%,
respectively, the average fallout value has increased
for 0.13%. This is unfortunately the price that has to
ICEIS 2005 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS
406
be paid for the great reduction of a response time,
which has been reported in the previous sub-section.
This much shorted duration of filtering can be also
the only possible explanation that the received feed-
back value has been reduced for only 3%, which is
two times less that is the reduction of a precision.
6 CONCLUSION
The goal of this paper was to provide solutions to
the challenges in filtering community cooperation,
being unavoidable in nowadays rich information
society. This was achieved through methods being
able both to determine how promising is each avail-
able community for a particular request and to com-
pose the final recommendation set by choosing the
best from the found results.
Even though the first solutions for describing in-
formation being stored at each community are given,
a future work will be focused on the usage of shared
ontologies for taking care about very diverse seman-
tics that the same word has in different communities.
The price, being paid in a user feedback, precision
and recall domains, will be tried to be reduced also
through the application of specialised strategies for
keeping updated the content descriptions of distrib-
uted, heterogeneous and dynamic filtering communi-
ties. As soon as it becomes unfeasible that each and
every community has descriptions of all others, the
Time-To-Live parameter will be assigned to every
filtering job, which will enable their further propaga-
tion, being the basic idea behind all P2P data sharing
systems. Even though given results are just the ini-
tial step towards intelligent cooperation in a multi
agent framework, authors’ hope is that the deployed
cooperation lays the foundation for the provision of
sophisticated filtering services.
REFERENCES
Albayrak, S., Milosevic, D., 2004a. Generic Intelligent
Personal Information Agent. IPSI-2004S. Serbia.
Albayrak, S., Milosevic, D., 2004b. Self Improving Coor-
dination in Multi Agent Filtering Framework. IEEE,
WIC and ACM Conference on IAT and WI. China.
Albayrak, S., Milosevic D., 2004c. Situation-aware Coor-
dination in Multi Agent Filtering Framework. The
19th Symposium on Computer and Information Sci-
ences (ISCIS 04). Antalya. Turkey. pp 480-493.
Balabanovic, M., Shoham Y., 1997. Fab: Content-Based
Collaborative Recommendation. Communication of
the ACM. 40(3). pp. 66-72.
Blake, C., Pratt W., 2001. Better Rules, Fewer Features: A
Semantic Approach to Selecting Features from Text.
Proceedings of the IEEE International Conference on
Data Mining. San Jose. California. pp. 59-66.
Boley, D., 1998. Principal Direction Divisive Partitioning.
Data Mining and Knowledge Discovery. 2(4).
Brin, S., Page L., 1998. The anatomy of a large-scale hy-
per textual (Web) search engine. 7
th
WWW Confer-
ence on Computer Networks. 30(1-7). pp. 107-117.
Delgado, J., 2000. Agent based information filtering and
recommender systems. Ph.D. Thesis, Nagoya Institute.
Faloutsos, C., Oard, D., 1995. A Survey of Information
Retrieval and Filtering Methods. Technical Report
CS-TR-3514. University of Maryland.
Fricke, S., Bsufka, K., Keiser, J., Schmidt, T., Sesseler, R.,
Albayrak, S., 2001. Agent-based Telematic Services
and Telecom Applications. Comm. of ACM. pp. 43-48.
Frigui, H., Nasraoui, O., 2004. Simultaneous clustering
and dynamic keyword weighting for text documents.
Survey of Text Mining Clustering, Classification, and
Retrieval. Springer-Verlag. pp. 45-72.
Han, J., Kamber, M., 2001. Data Mining: Concepts and
Techniques. Morgan Kaufmann Publishers.
Kohonen, T., Kaski, S., Lagus, K., Salojärvi, J., Honkela,
J., Paatero, V., Saarela, A., 2000. Self organization of
massive document collection. IEEE Transactions on
Neural Networks. 11(3). pp. 574-585.
Michalewicz, Z., Fogel, D., 2000. How to Solve It: Mod-
ern Heuristics, Springer-Verlag New York.
Mohammadian, M., Jentzsch, R., 2004. Computational
Intelligence Techniques Driven Intelligent Agents for
Web Data Mining and Information Retrieval. Intelli-
gent Agents for Data Mining and Information Re-
trieval. Idea Group Publishing. pp. 15-29.
Oard, D., Marchionini, G., 1996. A Conceptual Frame-
work for Text Filtering. MD 20742. Maryland.
Saurabh, S., Eui-Hong, H., Vipin, K., 2001. Web Search
Engine Technologies: A Survey. Minnesota. USA.
Sesseler, R., Albayrak, S., 2002. JIAC IV – An Open,
Scalable Agent Architecture for Telecommunications
Applications, 1
st
International NAISO Congress on
Autonomous Intelligent Systems, Geelong. Australia.
Sheth, B. D,, 1994. A Learning Approach to Personalized
Information Filtering. M.S. Thesis. MIT- EECS. USA.
Tauritz, D., 2002. Adaptive Information Filtering: con-
cepts and algorithms. Ph.D. Thesis. Leiden University.
Yang, H., Zhang, M., 2004. Potential Cases, Database
Types, and Selection Methodologies for Searching
Distributed Text Databases. Intelligent Agents for
Data Mining and Information Retrieval. IG. pp. 1-14.
Zamir, O., Etzioni, O., Madani, O., Karp, R., 1997. Fast
and Intuitive Clustering of Web Documents. Knowl-
edge Discovery and Data Mining (KDD 1997). pp.
287-290.
DISTRIBUTED COMMUNITY COOPERATION IN MULTI AGENT FILTERING FRAMEWORK
407