MULTI-INTEREST COMMUNITIES AND COMMUNITY-BASED

RECOMMENDATION

Fang Wang

Pervasive ICT Research Centre, British Telecom Group, Orion 1/12, Adastral Park, Ipswich IP5 3RE, UK

Keywords: Community organisation, recommendation, association analysis.

Abstract: This paper introduces an approach to organising multi-interest communities in which a user may belong to

more than one community. The interests of a user are first identified from the resources he handled and then

refined through interest association analysis in order to remove false or redundant interests. To each

identified interest topic, users who have this topic are clustered together, so a series of multi-interest

communities are formed. Because members of a community may have interest in the topics of other

communities, the formed communities are also connected with each other, resulting in a kind of community

network indicating interest associations of groups of users. Based on formed multi-interest communities,

users will receive useful recommendations within their own communities and from other related

communities. This provides users opportunities to obtain information beyond their current interests so new

interests of the users may be discovered. The multi-interest communities approach has been examined on

the EachMovie data. The experimental results showed that the formed multi-interest communities were

more cohesive and condensed when users were clustered according to their refined interest topics. The users

also received much more recommendations based on multi-interest communities.

1 INTRODUCTION

The increasing use of the Internet brings out a new

era of “online communities” or “virtual

communities” that create a virtual space in which

members do not necessarily meet and communicate

with each other face-to-face. The formation of

online communities has provided the opportunity for

remote people to build social relationships and to

exchange or share information with each other.

Online communities have now been widely

discussed in education, business and peer-to-peer

computing. An important application of online

communities is to provide users useful

recommendations according to other like-minded

users’ experiences. It is obvious that, in order to

attain valuable recommendations, an essential factor

is to find members with similar characteristics and

construct proper online communities accordingly.

Many software platforms have perceived the

importance and benefits of online communities and

developed supportive services to encourage and

facilitate group activities, e.g., synchronous or

asynchronous communications between group

members (Cassiopeia, Vignette, Webfair, e-groups).

Example application areas of these platforms include

knowledge communities, business-to-business

communities and customer-related communities. A

few recent literatures in e-learning have investigated

methods to create online learning communities

because learning in collaboration, i.e., collaborative

learning, is often more effective when students learn

as a group (Seufert, et. al., 2002, Talavera and

Gaudioso, 2004). In addition to classifying students

according to their subjects or demographic

information derived from surveys, student clusters

were learned probabilistically (e.g., via the

Estimation Maximum algorithm) based on features

mined from student interaction with the e-learning

system. Within each formed community, useful

recommendations were obtained by inspecting the

reputation or content (e.g., meta-data) of each

community member (Nijholt, 2002).

Interest-based communities have been

intensively studied in peer-to-peer computing, as

grouping remote peers according to their similarity

will greatly improve the efficiency of resource

search and sharing in a distributed environment.

Generally, peers were given some attributes or

objectives and joined into the same group when they

Wang F. (2007).

MULTI-INTEREST COMMUNITIES AND COMMUNITY-BASED RECOMMENDATION.

In Proceedings of the Third International Conference on Web Information Systems and Technologies - Society, e-Business and e-Government /

e-Learning, pages 37-45

DOI: 10.5220/0001273800370045

 SciTePress

discovered other peers with the same defined

attributes or objectives (Khambatti, et. al., 2004,

Ogston, et. al., 2003). Peers were also linked

together when they had similar access patterns to the

same documents. This resulted in a kind of data-

sharing graph which was small-world. Wang

proposed a decentralised approach, self-organising

communities, to organise distributed peers into a

series of communities (Wang, 2002). Users with

matching behaviour, e.g., one answered a query of

another, were identified by a kind of middle agent

and then grouped together. Without sophisticated

user similarity calculation, self-organising

communities successfully clustered users with

similar interest or preferences. This work, however,

assumed users had only one category of interest and

belonged to one group each time. Self-organising

communities have been introduced in an e-learning

context to organise students into specific learning

groups (Yang, et. al., 2004).

Most existing work of online communities

attempt to allocate a user one most fitting

community whose members have the maximum

similarity. A user in such a community will receive

pertinent information most interested by others.

However, there is also a high probability that a user

will never be able to discuss certain interesting

information if this kind of information is widely

unconcerned by other members of the community.

Therefore, single-interest communities in which a

user only belongs to one community tend to have a

preference to ‘popular’ topics in which most

community members are interested and exclude

unpopular subjects that attract few members’

attention.

This paper introduces a novel approach to

organise multi-interest online communities, in which

a user may belong to multiple communities. Multi-

interest communities take users’ manifold interests

into full consideration and form communities based

on user behaviour or activities, particularly the way

they handle resources, e.g., access to resources or

votes to resources. User interests are identified from

the attributes of the resources they have handled,

and further refined through interest association

analysis. Users who have the same refined interest

topic are grouped together, but have different

association strengths to the group, reflecting their

varied interest degrees. As a user may join into

several communities, the formed communities

develop into a community network at the same time.

The directed connections between communities

indicate how close one community is to another. As

a consequence, recommendations can be made

within the communities and across communities that

have close relationships.

The remainder of this paper is organised as

follows. The next section introduces the formation

of multi-interest communities. Recommendations

based on multi-interest communities are explained in

Section 3. Section 4 shows how the multi-interest

communities approach worked on EachMovie data

and how recommendations were accordingly made.

The last section concludes this paper.

2 MULTI-INTEREST

COMMUNITIES

The construction of multi-interest communities

involves three key steps: user interest identification,

community organisation and community network

formation. Based on formed multi-interest

communities, various kinds of recommendations can

be made.

2.1 User Interest Identification

A user’s interests are firstly identified from the

resources he has handled and then refined via

association analysis.

Identification of potential user interests

Resources or data in most practical applications such

as documents or movies are usually associated with

a set of well defined attributes or classes to

summarise their general characteristics. A movie, for

example, may have one or more genres including

Action, Animation, Art_Foreign, Classic, Comedy,

Drama, Family, Horror, Romance, and Thriller.

When a user accesses a resource R

, it usually

indicates that this user is interested in the attributes,

or at least some of the attributes of this resource.

Because it is difficult to judge from this single

access what exact attributes this user is interested in,

the attributes of this resource only suggest a

potential interesting topic to this user.

Suppose resource R

has attributes or classes {a

, …, a

}. A user who has handled this resource

may be interested in the combination of all these

attributes, marked as T

, or only part of the attributes

– this needs to be further investigated as shown

below. If all of the resources processed by a user are

, R

, …, R

}, the potential interest topics of this

user is the aggregation of those resources’ attributes,

noted as

T , i=1, 2, …, m.

WEBIST 2007 - International Conference on Web Information Systems and Technologies

Refining potential interests

In a potential interest set

T , a potential interest

topic is obviously true if this topic has only one item

(or attribute). A topic with more than one item may

not be a valid interest to a user because it is possible

that a resource R

of this topic is accessed only

because it involves a particular attribute that is of the

user’s real interest. However, if the user frequently

accesses resources of the same topic, it is possible

that this topic is a real one or at least near to a real

interest of the user. When the potential interest topic

is backed up by a good amount of resources of the

same kind, this proves that this interest topic

possesses enough Support and Confidence.

Obviously the more resources a user handles, the

more accurate would it be to verify the real interests

of the user.

The Support and Confidence of a topic are

defined similarly to those used in traditional

association analysis (Han and Kamber, 2000).

Suppose an interest topic T

is the conjunction

(combination) of attributes a

, a

, …, a

. Support of

is then the percentage of resources processed by a

user, which have topic T

processedresoucesofNumber

TtopichavethatresourcesprocessedofNumber

TSupport

=)(

(1)

Confidence of T

in attribute a

(j=1,2,…, n), is c% if

c% of resources that has attribute a

also has topic T

It is obvious that:

)(

)|(

aSupport

TSupport

aTConfidence =

(2)

A potential interest topic T

is said to be a valid

interest of a user if the confidences on all of the

attributes a

satisfy a confidence threshold γ.

If T

is proven to be invalid from its confidence

calculation, the combinations of part of the attributes

may still be valid interests if they have enough

confidences on their attributes. A potential interest

topic will be examined in this way until a valid topic

is obtained or there is only one item/attribute left.

The final interests of a user are composed of all of

the valid interest topics after examination on

including interests with either a single attribute or a

combination of multiple attributes.

2.2 Community Organisation

The organisation of communities is relatively

straightforward after obtaining refined interests of

all of the users: to each refined interest topic T

, a

community of this topic will be created to include all

users who possess this topic. By grouping users that

have the same valid interest topic together, a series

of communities are obtained. Because a user usually

has more than one interest topic after interest

refining, he will join in several communities at the

same time. Every user has an association strength

associated with each of his communities. The

association strength indicates the interest degree of a

user to an interest topic or a community. In real

applications, the strength may be defined as the

number of resources handled by this user within a

community or the confidence level of the user in a

community.

2.3 Community Network Formation

A community network illustrates the relationships

between communities. To be exact, it reflects the

correlations of the interests of groups of users. If

most members in a community A, for instance, are

also members of a community B, this suggests that

most members interested in the topic of community

A are also interested in the topic of community B.

As a result there is a strong connection from

community A to B and the connection strength

indicates the closeness of A to B.

However, even if most members of A belong to

B, it is possible that these members are only a small

portion of the members of B. So the connection from

B to A may be much weaker than that from A to B.

The resulting community network is hence a

directed and asymmetric graph.

By observing community correlations, we will

not only get knowledge of how individual users

share common interests together, but also how they

share the same resource access pattern and how

different kinds of interest topics are indirectly

related because of their groups of users. The latter

will provide valuable information for

recommendation across related communities.

3 RECOMMENDATION BASED

ON MULTI-INTEREST

COMMUNITIES

The formation of multi-interest communities can be

useful to a series of practical applications. For

instance, by clustering users into groups, it would be

much easier for them to share information and build

social relationships together. Personalised services

can also be provided to individual users or groups of

MULTI-INTEREST COMMUNITIES AND COMMUNITY-BASED RECOMMENDATION

users according to their identified interests. Another

typical application of multi-interest communities is

recommendation of valuable information to a user or

groups of users according to other similar or related

users’ experiences. Two kinds of recommendations

can be made based on multi-interest communities:

intra-community recommendation and inter-

community recommendation.

Intra-community recommendation

Intra-community recommendation takes place within

each community. As all of the members in a

community are interested in the same topic, it is

possible that a member will be interested in the

resources of the same kind and most accessed by

other members. In contrast to most collaborative

filtering techniques that do not discriminate

recommended data about their relative popularity,

this paper ranks resources of a community for

recommendation. This is similar to the work

reported in (Lawrence, et. al., 2001). However, in

addition to resource popularity, the recommendation

provided in this paper is also dependent on how

much information a user to be recommended

requires and how close the user is associated with

his community (i.e., how much the user is interested

in the community topic). Therefore, the

recommendation

),(

urecℜ

of a resource r

community C

to a user is a function of the user’s

requirement, the user’s relationship to the

community and the popularity or usefulness of the

resource deemed by the other users of the

community. This is illustrated by the following

equation:

),(),()(Re),(

kjkiiij

CrPopCuuqurec ⋅⋅=ℜ

(3)

where

]1,0[)(Re ∈

indicates the recommendation

requirement of user u

. It has a maximum value 1,

which means that the user welcomes all

recommendations.

0)(Re =

suggests that the

user does not want any recommendations.

),(

is the association strength of a user to his

community C

. It can be decided explicitly by the

user or implicitly from the user’s activities in the

community, e.g., how many resources the user has

accessed.

),(

CrPop

is the popularity or

usefulness of a resource in community C

. It can be

measured from the resource’s access frequency or

voted scores of the community members. Resource r

will be recommended to user u

if the final

),(

urecℜ

obtained from equation (1) is greater

than a threshold ε

Inter-community recommendation

In addition to recommendation among community

members, groups of users are able to receive useful

information recommended from other related

communities. Due to the introduction of inter-

community recommendation, users will not only

receive useful information from similar-minded

peers, but also information of different kinds but

most referred by other related peers. Inter-

community recommendation provides a user an

opportunity to obtain information beyond his current

communities and as a result, new interests of this

user may be identified.

In inter-community recommendation, a

community A will receive recommendations from

community B only if there is a connection from A to

B, suggesting that some users interested in A’s topic

have also interest in the topic of B. Again the

resources of a community are ranked according to

their popularity. The actual recommendation

),,(

lkj

CCrecℜ

of a resource r

in community C

to community C

depends on the popularity of

resource r

in C

, the association strengths or

closeness from C

to C

and the member

requirements of community C

, as illustrated below:

),(),()(Re),,(

ljlkklkj

CrPopCCCqCCrec ⋅

⋅

ℜ

(4)

where

]1,0[)(Re ∈

indicates the requirement of

community C

),(

is the association strength

of the connection from community C

to community

and

),(

CrPop

is the popularity of resource r

community C

. Resource r

will be recommended to

community C

),,(

lkj

CCrecℜ

is greater than a

threshold ε

When community C

receives a recommended

resource r

, it will not disseminate this resource to

every user in the community. Resource r

will only

be forwarded to users that may have interest in it.

This will be verified by equation (3) as introduced

above, that is, recommendation of r

to a particular

user u

in community C

is also decided by the user’s

individual requirement and his closeness to the

community C

4 EXPERIMENTAL

SIMULATION

Multi-interest communities have been tested on the

EachMovie data provided by the DEC systems

WEBIST 2007 - International Conference on Web Information Systems and Technologies

research centre (EachMovie). EachMovie data were

collected from 72916 users who entered a total of

2811983 numeric ratings for 1628 different movies

(films and videos). Every movie falls into one or

more of 10 genres: Action, Animation, Art_Foreign,

Classic, Comedy, Drama, Family, Horror, Romance

and Thriller. User votes were recorded in the

following format:

Person_ID Movie_ID Score Weight Modified:

Date/Time

Here, Score and Weight are numerical numbers

between 0 and 1, i.e., 0 < Score, Weight <= 1. In

particular, the Score maps linearly to the zero-to-five

star rating, that is, it has a value of 0, 0.2, 0.4, 0.6,

0.8 or 1. Weight indicates whether the person rated a

movie as zero to five stars (weight = 1) or "sounds

awful" (weight < 1). (Most "sounds awful" weights

are 0.2, but for historical reasons about 10% are

0.5.)

The EachMovie data set was initially built for

examining collaborative filtering techniques. It was

used in this paper to particularly test the formation

of multi-interest communities and the

recommendation made on multi-interest

communities.

User Interest Identification

5000 random users were chosen from the

EachMovie data for the examination presented in

this paper. The interests of these users were first

identified from their voted movies as introduced in

Section 2.1.

Here we take user 10 as an example. This user

voted in total 88 movies, belonging to 27 kinds of

genres, as shown in Figure 1. This suggests that the

potential interests of user 10 included those 27

genres. For 5000 users, in total 76 kinds of potential

interest topics were identified from their votes.

In the potential interest topics of a user, some of

them may be false or subject to more general topics

so they should be removed from the user’s real

interests. In order to identify those invalid interest

topics, refining rules introduced in Section 2.2 were

employed to calculate the Support and Confidence

values for all potential interest topics. Figure 2

shows the Support values of user 10 on ten movie

genres. Figure 3 shows the confidences of user 10’s

27 potential interest topics on movie genres. In this

experiment, the confidence threshold γ was set as

0.1. Interest topics with a genre confidence lower

than 0.1 suggest that these topics are not of major

interest to the user. After interest refining, 7 valid

interests were obtained from user 10’s 27 potential

interests. The final valid interest topics include

(Comedy Romance), (Drama), (Art_Foreign),

(Comedy), (Animation Family), (Action Thriller)

and (Action), as indicated in Figure 1.

It is worth noting that (Action Thriller) is treated

as a separate valid interest though it is logically

subject to (Action). It is because this category

received many votes so the confidences on both

genres of Action and Thriller are high. This suggests

that user 10 has strong interest in this specific

category, though he has also a broad interest in

general Action movies. Another category

(Animation Family) only received 2 votes from user

10. However, in the EachMovie data, there were in

total only two movies of this category and user 10

voted for both of them. This category represents user

10’s particular interest in Animation and Family and

is hence recognised as a positive interest of this user.

Valid interest topic

(Art_Foreign Comedy)

(Comedy Family)

(Comedy Drama)

(Drama Romance)

(Art_Foreign Drama)

(Horror Thriller)

(Drama Family)

(Art_Foreign Comedy Romance)

(Action Classic Family)

(Comedy Drama Romance)

(Action) *

(Action Thriller) *

(Action Comedy)

(Action Comedy Thriller)

(Classic Drama Thriller)

(Comedy Drama Horror Thriller)

(Action Drama)

(Art_Foreign Drama Horror)

(Thriller)

(Animation Family) *

(Comedy) *

(Action Horror)

(Art_Foreign) *

(Drama Thriller)

(Comedy Romance) *

(Action Comedy Horror Thriller)

Votes

(Drama) *

Figure 1: Votes of User 10.

Support

Thrille

Romance

Horror

Family

Drama

Comedy

Classic

Art_foreign

Animation

Action

Figure 2: User 10’s Support on movie genres.

(Action Horror)

(Comedy Family)

(Comedy Drama)

(Drama Romance)

(Art_Foreign Drama)

(Horror Thriller)

(Drama Family)

(Art_Foreign Comedy Romance)

(Action Classic Family)

(Comedy Drama Romance)

* (Action)

* (Action Thriller)

(Action Comedy)

(Action Comedy Thriller)

(Classic Drama Thriller)

(Comedy Drama Horror Thriller)

(Action Drama)

(Art_Foreign Drama Horror)

(Thriller)

* (Comedy)

* (Drama)

(Drama Thriller)

Thrille

Family

Drama

Comedy

Classic

Animat

ion

Actio

0.2

0.4

0.6

0.8

(Action Comedy Horror Thriller)

* (Animation Family)

(Art_Foreign)

Art_foreig

Horror

Romance

Confidence

* (Comedy Romance)

(Art_Foreign Comedy)

Figure 3: Confidences of user 10’s potential interest topics

on movie genres.

MULTI-INTEREST COMMUNITIES AND COMMUNITY-BASED RECOMMENDATION

Through the interest refining process, the total

number of interest topics of 5000 users was reduced

from 76 to 58. On average a user has 9 interests.

Some users showed broad interests with as many as

21 interest topics, whereas some others had only 1

interest topic. Figure 4 shows the distribution of the

number of user interests. This result suggests that

the number of people that own k different interests

does not vary when the number of interests is

smaller than 9. That is, the probability that someone

has k<9 interest topics is independent of k. Beyond

this point, for intermediate values of k (9<k<16), it

is likely that someone chosen at random belongs to

exactly k communities. Finally, as k goes beyond

16, the probability that a user belongs to k>16

communities decays exponentially fast in k.

Number of interests

200

400

600

800

1000

1200

1 10 10

Number of users

Figure 4: Distribution of number of user interests.

Formation of multi-interest communities

For each refined interest topic, all of the users that

showed this interest were suggested to join into the

same group of this topic. This finally resulted in 58

communities. The number of members of these

communities ranged from 1 to 8449. The top ten

communities that have the most members are

(Drama), (Action), (Comedy), (Action, Thrill),

(Animation Family), (Comedy Romance), (Drama

Thriller), (Action Drama), (Thriller) and (Horror

Thriller). Actually these communities were also the

most popular groups before user interest refining but

had more members.

For a comparison of the communities made

before and after interest refining, the community

cohesiveness and the average user strength of each

community were calculated for both cases. Here the

weighted similarity (Steinbach, et. al., 2002), which

is the squared length of the community centroid, was

used to test the internal cluster similarity, as shown

in equation (5).

∑

∈

Cuu

uuine

CCentroid

),(cos

)(

(5)

where users

u and

u are members of community

and

are voting vectors of users

u and

respectively. User strength was measured as the

number of votes voted for the movies in each

community, normalised by the total number of

movies in the community.

Table 1 lists the average community

cohesiveness and average user strength in

communities before and after user interest refining.

It proves that the community cohesiveness, or intra-

community user similarity, was improved from 4.11

to 7.32 when users are clustered according to their

refined interests. At the same time, the average user

strength which indicates user closeness to their

communities was increased from 0.35 to 0.45. These

results suggest that more condensed communities

were obtained when interest refining was employed

to purify user interest topics. By removing relatively

irrelevant community members and loose

communities, the resulting communities had

improved closeness among community members

and between community and their members.

Table 1: Community cohesiveness and average user

strength before and after interest refining.

Community

cohesiveness

Average user

strength

Before interest

refining

4.11 0.35

After interest refining 7.32 0.45

Community Network

Depending on the relationships of community

topics, 58 communities formed a complex network.

For any two communities that have connections,

which means those communities have common

users, the strength of connection from community C

to C

was calculated as follows:

∑

∈

∩∈

CCu

),(

(6)

where the denominator is the sum of the strengths of

all users in community C

and the numerator is the

strength sum of the common users of community C

and C

. This community connection strength shows

how much the users in community C

are interested

in the topic of community C

. As shown by equation

(6), the connections between C

and C

are

directional, depending on the percentage of the

common users in each community.

The community network of EachMovie

communities is illustrated as a directed graph as

shown in Figure 5. Only 2600 relationships were

presented among 3306 possible connections between

WEBIST 2007 - International Conference on Web Information Systems and Technologies

58 communities. Some communities had no direct

connection to each other at all, such as communities

(Animation Comedy) and (Classic), and (Romance)

and (Drama Horror), which means those

communities had no common users. Some

communities showed a kind of hierarchy, such as the

connection from (Animation Comedy) to (Comedy),

as all of the users of the former were also members

of the latter. Hierarchical connections usually have

high connection strengths.

Close relationships have been found between

popular communities, such as connections from

(Art_Foreign Classic) to (Action) and from (Action

Horror Thriller) to (Drama), though those

communities are semantically non-related. Actually

the most popular 10 communities were nearly fully

connected with each other and these communities

shared at least half of their members. The

connections between popular communities are also

asymmetric. For example, the relationships from

(Action) to (Art_Foreign Classic) and from (Drama)

to (Action Horror Thriller) were much weaker than

those of the other direction.

Figure 5: Community network.

Community-based Recommendation

Intra-community Recommendation

Intra-community and inter-community

recommendations can now be made based on the

organised communities and community network.

User 10 was again taken as an example to show how

those two kinds of recommendations could be made

for him. It is assumed that this user welcomed all

recommendations, that is,

1)(Re

. As

introduced in Section 4.1, user 10 belonged to seven

communities, including (Comedy Romance),

(Drama), (Art_Foreign), (Comedy), (Animation

Family), (Action Thriller) and (Action). The top 10

movies in each community were selected for

recommendation, as listed in Appendix A, except

those movies marked with *, which were already

viewed and voted by this user. As a result, there

were in total 97 movies sent to user 10 for

consideration.

Inter-community recommendation

The seven communities where user 10 belonged to

connected many other communities. This was

because the other members of those seven

communities showed strong or weak interests in

other movie genres. Some communities were linked

by more than one of those seven communities in the

formed community network with varied connection

strengths. Figure 6 shows the average connection

strengths of the connected communities. As shown

in Figure 6, there were around 10 other movie

communities most linked by user 10’s communities.

Those ten communities were hence selected for

inter-community recommendation to user 10. The

top 10 movies in these strongly related communities

are listed in Appendix B. Most of the movies were

not viewed or voted by user 10. Actually, among

those selected 10 other communities, 4 of them were

new to user 10 as this user had no votes to the

movies of these communities. User 10 had only one

or two votes to the other six connected communities.

Depending on how many recommendations user

10 would accept, we could further inspect this user’s

real interests or even find out his new interests that

had not been shown in this user’s votes before. This

work, however, is impossible to do due to the

anonymity of EachMovie users. We hope to do more

tests on real users to complete this work in our near

future.

Connected communities

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0 5 10 15 20 25 30 35 40 45 5

Connection strength

Figure 6: Average connection strengths of related

communities of user 10.

5 CONCLUSTIONS AND FUTURE

WORK

This paper introduces an approach to organising

multi-interest communities in which a user may

belong to more than one community. The interests

of a user are first identified from the resources he

MULTI-INTEREST COMMUNITIES AND COMMUNITY-BASED RECOMMENDATION

handled and then refined through interest association

analysis in order to remove false or redundant

interests. To each identified interest topic, users who

have this topic are clustered together, so a series of

multi-interest communities are obtained. Because

members of a community may have interest in the

topics of other communities, the formed

communities are also connected with each other,

resulting in a kind of community network indicating

interest associations of groups of users. Experiments

on the collected EachMovie data showed that the

formed multi-interest communities were more

cohesive and condensed when users were clustered

based on their refined interest topics.

Intra-community and inter-community

recommendations can be made based on formed

multi-interest communities. The former recommends

to a user popular resources deemed by other

community members within a community, whereas

the latter suggests resources of other categories but

most interesting to the other members of a

community. Consequently a user will receive

information within and beyond his identified

interests. From his responses to the

recommendations, e.g., accept or reject some of

them, the user’s real interests will be further

identified and even new interests of the user may be

discovered.

The recommendation approach presented in this

paper is more suitable to users that welcome all

information recommendations. Some users in reality

may only wish to receive carefully selected

information, which means their recommendation

requirements

1)(Re

. How to tailor

recommendations according a user’s requirements

and preferences will be studied as our future work.

In addition, a user’s interest to a certain topic or

resource will be divided into positive and negative,

instead of all positive interests as shown in this

paper. Because the users of the EachMovie data are

anonymous, it is difficult to judge whether they

would like the groups allocated to them and the

recommendations suggested by our approach. The

proposed multi-interest communities and

community-based recommendation will be further

tested on real users so that the users’ feedback will

be used to examine and improve the approach

presented in this paper.

ACKNOWLEDGEMENTS

The author is grateful to the support from BT Long

Term Research Venturing, colleagues in Nanjing

University of China and the National Natural

Science Foundation of China Grant No.60402027.

RERERENCES

Cassiopeia http://www.cassiopeia.com/

EachMovie http://research.compaq.com/SRC/eachmovie/

e-groups http://www.egroups.com/

Han, J. and Kamber, M., 2000. Data Mining: Concepts

and Techniques, Morgan Kaufmann Publishers.

Iamnitchi, A., Ripeanu, M. and Foster, I., 2004. Small-

World File-Sharing Communities, The 23

Conference of the IEEE Communications Society.

Hong Kong.

Khambatti, M., Ryu, K.D. and Dasgupta, P., 2003,

Structuring Peer-to-Peer Networks using Interest-

Based Communities. 1st International Workshop on

Databases, Information Systems, and Peer-to-Peer

Computing. Berlin, Germany, pp. 48-63.

Lawrence, R.D., Almasi, G.S., Kotlyar, V., et. al. 2001.

Personalization of Supermarket Product

Recommendations. Data Mining and Knowledge

Discovery 5(1-2):11-32.

Nijholt, A., 2002. Computer-facilitated community

building for E-learning, Proc. IEEE Inter. Conf. on

Advanced Learning Technologies. Kazan, Russia,

pp.541-543.

Ogston, E., Overeinder, B., van Steen, M., and Brazier, B.,

2003. Group Formation Among Peer-to-Peer Agents:

Learning Group Characteristics. Inter. Workshop on

Agents and Peer-to-Peer Computing, pp. 59-70.

Seufert, S., Lechner, U. and Stanoevska, K., 2002. A

reference model for online learning communities.

Inter. J. on E-Learning, Jan-Mar, pp.43-55.

Steinbach, M., Karypis, G. and Kumar, V., 2000. A

comparison of document clustering techniques, KDD

Workshop on Text Mining.

Talavera, L. and Gaudioso, E., 2004, Mining student data

to characterize similar behaviour groups in

unstructured collaboration spaces. Workshop on

Artificial Intelligence in CSCL, 16

European

Conference on Artificial Intelligence. pp.17-23.

Vignette http://vignette.com/

Webfair http://www.webfair.com/

Wang, F., 2002. Self-organising communities formed by

middle agents. Proc. of the 1st Inter. Conference on

Autonomous Agents and Multi-Agent Systems,

Bologna, Italy, pp 1333-1339.

Yang, F., Shen, R. and Han, P., 2004. A novel self-

organizing e-learner community model with award

and exchange mechanisms, Journal of Zhejiang

University Science, 5(11): 1343-1351.

WEBIST 2007 - International Conference on Web Information Systems and Technologies

Appendix A: Intra-community recommendations for user 10.

Movies marked with * were already viewed and voted by this user.

(Comedy Romance) (Drama) (Art_Foreign) (Comedy) (Animation Family) (Action Thriller) (Action)

1 Late Bloomers Hard Eight Identification of a

Woman

The Full Monty * Toy Story Aliens (1986) Air Force One

2 Love and Other

Catastrophes

Rosewood The Eighth Day Raising Arizona

(1987)

The Lion King In the Line of Fire The Terminator

(1984)

3 Groundhog Day In the Company of

Men

Jean de Florette (1986) A Fish Called

Wanda (1988)

Winnie the Pooh

and the Blustery

Day

Breakdown Die Hard (1988)

4 When Harry Met

Sally... (1989)

Schindler's List Manon of the Spring

(1986)

This Is Spinal

Tap (1984)

Snow White and the

Seven Dwarfs

Operation Condor

(Feiying Gaiwak)

Terminator 2:

Judgment Day

5 Strictly Ballroom The Shawshank

Redemption

The City of Lost

Children

Local Hero

(1983)

* Beauty and the

Beast

Heat * The Rock

6 Sleepless in Seattle Sling Blade Le Colonel Chabert Monty Python's

Life of Brian

The Fox and the

Hound (1981)

Star Trek II: The

Wrath of Khan

(1982)

Speed

7 My Best Friend's

Wedding

* Lone Star Paris Was a Woman Back to the

Future (1985)

Robin Hood (1984) Clear and Present

Danger

Blade Runner

(1982)

8 The American

President

Ran (1985) Madame Butterfly My Favorite

Year (1982)

Tim Burton's The

Nightmare Before

Christmas

Face/Off Hoodlum

9 * The Truth about

Cats and Dogs

Hamlet (1996) Le Confessionnal Living in

Oblivion

The Hunchback of

Notre Dame

* Independence Day

(ID4)

Full Metal

Jacket (1987)

10 * While You Were

Sleeping

Traveller Delicatessen Grosse Pointe

Blank

Casper Highlander (1986) The Big Blue

(1988)

Appendix B: Intre-community recommendations for user 10.

Movies marked with * were already viewed and voted by this user.

(Thriller) (Horror) (Horror

Thrill)

(Family) (Drama

Thriller)

(Comedy

Family)

(Action

Drama)

(Romance) (Animation) (Drama

Romance)

1 The Game

(1997)

Paradise Lost:

The Child

Murders at

Robin Hood

Hills (1996)

* The

Silence of

the

Lambs

Daniel

Defoe's

Robinson

Crusoe

Mother

Night

Babe Contact William

Shakespeare's

Romeo and

Juliet (1996)

The Wrong

Trousers

Love Jones

2 Nightwatch The Shining

(1980)

Jaws

(1975)

Mary Poppins

(1964)

* The Usual

Suspects

Liar, Liar Miller's

Crossing

Before

Sunrise

A Close

Shave

The Whole

Wide World

3 Reservoir

Dogs

An American

Werewolf in

London (1981)

Scream Pinocchio

(1940)

Once Upon

a Time in

America

(1984)

Mrs.

Doubtfire

Braveheart One Fine

Day

A Grand Day

Out

Somewhere

in Time

(1980)

4 Bound Freeway The

Abyss

(1989)

Willy Wonka

and the

Chocolate

Factory

(1971)

Murder in

the First

Home

Alone

GoodFellas Benny &

Joon

Wallace &

Gromit: The

Best of

Aardman

Animations

Infinity

5 Unforgiven Bram Stoker's

Dracula

Copycat Fly Away

Home

The

Frighteners

The Santa

Clause

Fire Down

Below

Pretty

Woman

Ghost in the

Shell

(Kokaku

Kidotai)

Blue Sky

6 Crimson

Tide

Interview with

the Vampire

Cape Fear

(1991)

A Little

Princess

Sleepers Matilda G.I. Jane It Could

Happen to

You

Heavy Metal

(1981)

Chasing

Amy

7 Ransom The Howling

(1981)

Mary

Shelley's

Frankenst

ein

The Jungle

Book

Tombstone

Muppet

Treasure

Island

The Long

Kiss

Goodnight

For the

Moment

The

Transformers:

The Movie

(1986)

Jerry

Maguire

8 Red Rock

West

The Prophecy

(God's Army)

Mute

Witness

Jumanji The Client The Stupids * Seven Bed of Roses n/a Some Kind

Wonderful

(1987)

9 The Grifters A Nightmare

on Elm Street

(1984)

Body

Snatchers

Cool

Runnings

Outbreak The Big

Green

Donnie

Brasco

Pie in the

Sky

n/a Ghost

10 Dolores

Claiborne

Cat People

(1982)

In the

Mouth of

Madness

Andre The Firm Jack True

Romance

Dirty

Dancing

n/a Up Close

and Personal

MULTI-INTEREST COMMUNITIES AND COMMUNITY-BASED RECOMMENDATION