A New Dynamic Community-Based Recommender System

Sabrine Ben Abdrabbah

, Nahla Ben Amor

and Raouia Ayachi

ole R&D, Audensiel, 93 rue Nationale, 92700 Boulogne-Billancourt, France

LARODEC, Institut Sup

erieur de Gestion de Tunis, Universit

e de Tunis, 41 Avenue de la Libert

e, Tunis 2000, Tunisia

LARODEC,

Ecole Sup

erieure des Sciences

Economiques et Commerciales de Tunis, Universit

e de Tunis,

41 Avenue de la Libert

e, Tunis 2000, Tunisia

Keywords:

Recommendation Systems, Collaborative Filtering, Dynamic Community Detection, Overlapping

Communities, Dynamic Networks.

Abstract:

Due to the rapid changes in users’ preferences over time, it becomes increasingly important to focus on the

temporal evolution of the users’ behavioral patterns to capture the most relevant items to the users. This

paper proposes a new framework for dynamic and overlapping community-based collaborative ﬁltering, which

models at ﬁrst the dynamic behavior of users’ interests into a temporal network of items. Then, we take

advantage of the dynamic and overlapping community detection techniques to ﬁnd the best partition of similar

items. The advantage of doing so is to (i) avoid processing the entire system to select similar items and

eventually overcome the scalability problem and (ii) provide users with a recommendation of items similar to

the latest appreciated items to match the current users’ taste and avoid consequently sparse cases. We conduct

experiments to study the sensitivity of some parameters (e.g., the datasets and the similarity measures) on the

recommendations’ quality. Experimental results show a considerable improvement in the proposed framework

recommendation’s accuracy compared to the state-of-the-art recommender systems.

1 INTRODUCTION

Due to the huge amount of resources accessible via

the internet, ﬁnding relevant information becomes a

challenging task. Recommender systems (RSs) were

introduced to achieve personalization and increase

users’ satisfaction by recommending items that ﬁt

their needs and tastes. Item recommendations can be

made using different approaches namely collabora-

tive ﬁltering, content-based and hybrid approaches.

Collaborative Filtering (CF) is the widely applied rec-

ommendation approach due to the high availability of

users’ preferences information (Qiang and Yan, 2012)

(i.e., the ratings assigned to items). Two main classes

of CF exist, namely, user-based and item-based. We

are particularly interested in the item-based approach

that consists in computing the preference prediction

based on the ratings given to similar items. Basi-

cally, the item-based CF analyzes the entire user-

item rating data to identify similar items. This can

present various challenges, including data sparsity,

scalability, and cold start problem. Community de-

tection was used in recommender systems to over-

come the recommendation issues (Sahebi and Cohen,

2011; Qin et al., 2010; Deng et al., 2014) and gen-

erate more accurate recommendations. The commu-

nity structure could reveal either the set of items or

users that share the same properties. Therefore, the

community-based recommendation has a low compu-

tational cost compared to the basic recommendation

approach since it focuses on pre-detected communi-

ties instead of the whole network. However, the exist-

ing community-based recommendation methods basi-

cally rely on static algorithms to detect communities

and fail to consider the temporal evolution of users’

interests.

Delving into the dynamic aspect of network be-

havior has become a necessity, especially for rapidly

growing complex networks. The customers’ pref-

erences and interests are changing over time. It is

crucial to track such temporal changes when deﬁn-

ing similar items. It is afﬁrmed that recommender

systems taking into account the dynamic behavior of

users’ interests can provide more relevant recommen-

dations of items (Yehuda, 2009). Based on this idea,

we aim to exploit the dynamics of users’ interests over

time to identify the appropriate communities corre-

sponding to each step of the network evolution. Incor-

Ben Abdrabbah, S., Ben Amor, N. and Ayachi, R.

A New Dynamic Community-Based Recommender System.

DOI: 10.5220/0011641300003393

In Proceedings of the 15th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2023) - Volume 2, pages 125-136

ISBN: 978-989-758-623-1; ISSN: 2184-433X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

125

porating such dynamic communities in recommenda-

tion tasks may provide the system with a good vision

of the most recent neighbors. Furthermore, in real-

world networks, the communities overlap such that

nodes may belong to several partitions at the same

time. For instance, an item naturally tends to belong

to several groups of different categories and popular-

ity. The overlapping aspect should also be taken into

account when detecting similar items to improve the

performance of the provided recommendations (Vla-

chos et al., 2014).

In this paper, we propose a new framework for dy-

namic and overlapping community-based collabora-

tive ﬁltering (called for short DO2CF). The proposed

framework considers the evolution of users’ prefer-

ences over time and the overlapping aspect of items

when identifying the most similar items. In rough

outline, our framework consists of three main steps,

namely dynamic network construction, dynamic com-

munity detection and recommendation. The ﬁrst step

consists in modeling the dynamic aspect of users’

interests into a temporal network where nodes are

the items and edges between nodes are established

based on the co-rating relationships. In the second

step, we use dynamic and overlapping community de-

tection techniques that can track all the topological

changes of the temporal network to learn the evolu-

tion of users’ interests and detect a community struc-

ture accordingly. Indeed, each community reveals

a set of items having similar interests. Finally, we

exploit the network partition to (i) select the items’

neighbors more easily and quickly, (ii) identify the

candidate items that match the users’ current tastes,

and (iii) adapt the similarity measure to consider how

much the co-users like one item relative to the other

similar items instead of all the other items they rated.

This paper is structured as follows. Section 2

presents the basics of recommender systems, com-

munity detection, and community-based recommen-

dation. Section 3 further illustrates the different steps

of our framework. Finally, Section 4 presents the con-

ducted experiments and their corresponding results.

2 RELATED WORK ON

COMMUNITY-BASED

RECOMMENDATION

We recall in this section the basic concepts of recom-

mender systems and community detection separately.

Then, we expose community-based recommendation

and their related works.

2.1 Recommender Systems

The growth of web content and services has raised

the interest in recommender systems. RSs are basi-

cally classiﬁed into three major categories, namely

Collaborative Filtering (CF), content-based ﬁltering,

and hybrid approach. The CF exploits the user-item

rating data to suggest items to the active user. The

content-based ﬁltering focuses on the item features to

characterize the items similar to the ones that the ac-

tive user likes. The hybrid approach takes advantage

of the rating information of users as well as the con-

tents of items to produce recommendations. In this

work, we focus on collaborative ﬁltering as it is the

most popular and used recommendation technique.

There are two main CF approaches: item-based when

the recommendations are generated based on the ac-

tive user’s ratings for similar items, and user-based

when the recommendations are computed based on

the similar users’ ratings to the target item. We are

particularly interested in the item-based CF approach

as it provides a better recommendation quality than

the user-based algorithms (Sarwar et al., 2002). Due

to the nature of the data, CF could reveal some crit-

ical challenges such as data sparsity when there are

not enough users’ ratings, scalability when the num-

ber of ratings grows, and cold start problem when the

active user is a new coming in the system (i.e., he does

not have any rating data).

2.2 Community Detection

Complex networks are a natural way to represent so-

cial, biological, technological, and information sys-

tems. An interesting feature, that complex networks

present, is the community structure-property. Com-

munity detection was proposed as an efﬁcient tool

to capture the hidden structure of these networks by

partitioning the entities into a set of dense subgroups

commonly called communities. A community is gen-

erally deﬁned as a set of nodes strongly connected

among them and more weakly connected to the re-

maining of the network (Fortunato, 2010). Study-

ing the forming community structure in complex net-

works may help to learn the collective behavior of

their entities and help the understanding of the mod-

eled systems. The ﬁrst attempts (Palla et al., 2005;

Newman and Girvan, 2004) focused on detecting

communities in static networks that are constructed

by aggregating all observed data (also called static

graphs). However, the democratization of the web

2.0 has produced new challenges for community de-

tection, including the overlap of communities and the

temporal aspect of the data (i.e., dynamic networks)

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

126

(Aston et al., 2014). The static community detec-

tion algorithms cannot support such challenges be-

cause they overlook the time information that is cru-

cial to understand the phenomena taking place in the

dynamic networks. Recently, many algorithms (Caz-

abet and Amblard, 2011; Nguyen et al., 2011; Xie

et al., 2013) have been proposed to deal with dynamic

networks.

2.3 Community-Based

Recommendation

In recommender systems, there are two classes of en-

tities: items and users. Indeed, the community struc-

ture reveals either the set of items or users with the

same properties. Most of the existing works are inter-

ested in detecting communities of users since people

often tend to form communities in real life. For in-

stance, (Sahebi and Cohen, 2011) considered the dif-

ferent dimensions of social networks to extract latent

communities of users using the Principal Modular-

ity Maximization method. The resulting communities

help recommender systems to overcome the cold start

problem when the user has no rating history. (Qiang

and Yan, 2012) applied a Multi-label propagation al-

gorithm for static community detection in a bipartite

network composed of users and items. Based on the

active user’s communities, the authors recommended

items using the collaborative ﬁltering recommenda-

tion method. (Ying et al., 2013) proposed a Prefer-

ence Aware Community Detection method (PCD) to

extract communities based on both users’ rating his-

tory and social structure. Besides, they implemented

a Preference Aware Community-based Recommenda-

tion System (PCRS) that uses the discovered commu-

nities to recommend items to users. (Guo and Peng,

2014) chose at ﬁrst the spectrum analysis algorithm to

extract communities from a user-user network. Then,

they employed an evolutionary algorithm to identify

the best neighborhood size for each community. Fi-

nally, they integrated the best neighborhood size of

the active user’s community in collaborative ﬁltering

to generate recommendations. On the other hand,

some studies are interested in ﬁnding communities

of items and incorporating them into the recommen-

dation model. For instance, (Qin et al., 2010) con-

structed a YouTube Recommender Network (YRN)

based on reviews left as comments in the YouTube

videos. Then, they used CFinder algorithm to ﬁnd

communities in YRN. These communities are then

used to recommend, for the active user, diverse videos

that are not restricted to the same tag annotation. In

(Fatemi and Tokarchuk, 2013), the authors employed

the Louvain community detection method to extract

communities of movies from the social network of

movies that is constructed based on the Internet Movie

Database (IMDb). Based on these communities, ex-

tensive and diverse movies are recommended to both

individuals and groups. The studies presented above

overlooked the time dimension of users’ preferences

when computing recommendation. However, the data

collected from recommender systems is often times-

tamped. The time dimension plays a very impor-

tant role to properly capture the current need of users

based on the most recent data.

Some studies proposed to explore the dynamic as-

pect of users’ preferences in the recommendations

generation process. (Abrouk et al., 2010) proposed to

use fuzzy k-means clustering from time to time to dy-

namically capture the last users’ preferences when de-

tecting communities. These latter are then exploited

to predict the active user’s preferences for the unseen

items. In (Hamzaoui et al., 2012), the authors have

clustered items into groups using the K-mean algo-

rithm. Then, they identiﬁed the clustering center of

each cluster. The center is represented as the average

ratings over all the items in the cluster. Finally, they

computed the similarity between the target item and

all the identiﬁed centers to select its neighbors and

generate recommendations. The authors assigned a

time weight for each pair of <user, item> with the

aim to decay the old data when computing the item-

item similarity. In (Xin et al., 2014), the authors em-

ployed the Importance Greedy (IG) algorithm to de-

tect communities in the reader-reader similarity net-

work. The book recommendation list is generated

based on top ’k’ books which are recently borrowed

by the users belonging to the active user’s community.

The time of book borrowing is considered to highlight

the time factor when generating book recommenda-

tions. (Yin et al., 2014) employed a modiﬁed Proba-

bilistic latent semantic analysis (PLSA) model to dis-

cover communities. Then, they applied matrix factor-

ization on each community. To model the temporal

changes in users’ interests, they used a time decay

function which weighed the importance of the users

latest interests when generating recommendations.

The existing community based-recommendation

methods still rely on static community detection al-

gorithms that cannot deal with the dynamic aspect of

users’ interests. Static community detection cannot

support the network’s topological changes, and conse-

quently it misrepresents the true image of the network

partitions. Using these communities in the preference

prediction process may reduce the performance of the

generated recommendations.

A New Dynamic Community-Based Recommender System

127

3 A NEW FRAMEWORK FOR

DYNAMIC AND OVERLAPPING

COMMUNITY-BASED CF

We propose a new framework for dynamic and over-

lapping community-based CF (DO2CF) (shown in

Figure 1) comprising three main steps detailed below.

Step 1: Dynamic network construction

Step 2: Dynamic community detection

Network community structure

Step 3: Recommendation

Time-stamped

Data

Item-item dynamic network

Recommendation list

Add

Delete

Figure 1: A framework for dynamic and overlapping

community-based CF.

3.1 Step 1: Dynamic Network

Construction

The users’ interests in items are rapidly changing over

time. This step consists in modeling the dynamic as-

pect of users’ interests into a temporal network as

it is the most appropriate representation when data

evolves rapidly. Indeed, the snapshot-based repre-

sentation cannot track all the interactions that appear

over the multiple stages of the network life. Such

representation is more adapted to time windows data

composed of long-lasting relations (for instance, a

weekly/monthly/yearly complete crawl of a system)

and cannot ﬁt the evolving nature of users’ interac-

tions in recommender systems. To build the tempo-

ral network, we start by modeling items as nodes and

items relationships as edges. Nodes and edges can be

added to (+) or removed from (-) the graph. We con-

sider that a link exists between two items as long as

these items interact frequently enough during a given

Table 1: Example of rating database.

UserID ItemID Rating(1 −

Timestamp

Alex Titanic 3 12/03/2014

Alex Troy 3 12/03/2014

Amel Troy 3 16/03/2014

Doudi Titanic 3 16/03/2014

Kim Batman 5 17/03/2014

Sam Titanic 4 14/03/2014

Sam Troy 4 14/03/2014

Kim Titanic 5 17/03/2014

Doudi Batman 3 16/03/2014

period. Items interactions are deﬁned based on the

co-ratings relationship delivered from users’ proﬁles.

An item-item interaction is created if one user gives

the same rating to both items in the same period. To

model the network topology changes, we deﬁne two

parameters N as the number of interactions and P as

the number of days.

• An edge is inserted if there are at least N interac-

tions between two items over a period of P days.

• An edge is removed if, over a period of P days,

they occur less than N interactions between two

items.

Example 1. Let us consider a movie rating database

that contains three movies (Titanic, Troy, Batman)

and ﬁve users (Alex, Amel, Doudi, Kim, Sam). The

users’ ratings data are presented in Table 1.

In order to construct the temporal network of the

period T from 12/03 to 17/03, we ﬁrst model the

movies as nodes and we consider that N = 2 and

P = 3. Thus, to create an edge between two movies,

we need at least two interactions between these ones

over a period of three days. To this end, we will check

if Titanic and Troy can be linked. So, we start by ex-

tracting all Titanic-Troy interactions:

• Alex rates both Titanic and Troy with the same

rating (i.e., 3) in the same date (March the 12

so the ﬁrst interaction between Titanic and Troy is

created on 12/03/2014.

• Sam gives the same rating (i.e., 4) to both Titanic

and Troy in the same date (March the 14

), so

the second interaction between Titanic and Troy

is created on 14/03/2014.

Over the ﬁrst 3 days (i.e., from (12/03) to (14/03)),

there are two interactions between Titanic and Troy.

Then, the edge between these nodes is created. Over

the 3 following days (i.e., from (14/03) to (17/03)),

the edge between Titanic and Troy is removed since

there are less than two interactions. Moreover, an

edge is created between Titanic and Batman since

there are more than two interactions between them

over the period from (14/03) to (17/03). The net-

work is ﬁnally constructed since there are no more

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

128

rating data after (17/03). The different steps of the

temporal network construction are presented in Fig-

ure 2.

Iteration 1: from 12/03/2014 to 14/03/2013

[2, 3]

# 12/03/2014

Iteration 2: from 15/03/2014 to 17/03/2013

[2, 3]

# 15/03/2014

- Titanic Troy

+ Titanic Batman

+ Titanic Troy

Figure 2: An example of the temporal network construction.

The parameters P and N are considered key to cre-

ating more semantic links between nodes. The es-

timation of these parameters highly depends on the

datasets. Recommendation datasets are often char-

acterized by three features, including the number of

users, the number of items, and the number of ratings

(i.e., numerical evaluations given by users to items).

At ﬁrst, we process all the data to ﬁnd out the num-

ber of interactions N corresponding to each pair of

items (i.e., how many users give the same rating to

both of these items) and the period P in which these

N interactions occurred. Then, we select many signif-

icant values for N and P: N

min

(resp. N

med

, N

and

max

) which indicates the minimal (resp. median, 3rd

quartile and maximal) number of interactions and P

min

(resp. P

med

, P

and P

max

) which indicates the mini-

mal (resp. median, 3rd quartile and maximal) number

of days. Knowing that the 3rd quartile indicates the

values below which 75% of observations in a set of

data fall. A temporal network is constructed based on

each possible combination between N and P.

Example 2. Let us re-consider the same rating

database of Table 1. We compute the parameters

N and P corresponding to each pair of items in the

dataset. The number of interactions N for [Titanic,

Troy] (resp. [Titanic, Batman] and [Troy, Batman]) is

equal to 2 (resp. 2 and 1). The period P in which these

interactions occurred between [Titanic, Troy] (resp.

[Titanic, Batman] and [Troy, Batman]) is equal to 3

(resp. 2 and 1). Thus, the number of interactions N

min

(resp. N

med

, N

max

) is equal to 1 (resp. 2, 2, 2)

and the number of days P

min

(resp. P

med

, P

max

) is

equal to 1 (resp. 2, 2, 3).

The resulting item-item dynamic network is then

considered as a generic temporal network that repre-

sents the evolution of users’ preferences over time.

3.2 Step 2: Dynamic Community

Detection

Given the temporal network that models the dynamic

evolution of users’ interests (i.e., the output of step1),

this step consists in using a community detection

algorithm to extract communities. The community

identiﬁcation process strongly depends on the nature

of the network (i.e., dynamic or static) and the over-

lapping aspect of the network entities (if the entities

may belong to different subgroups at the same time).

It is well understood that items in the recommender

system are naturally characterized by multiple com-

munity memberships since they may satisfy the needs

of several users with different tastes. Thus, overlap-

ping community detection might be more appropri-

ate to reveal interesting and useful patterns of similar

items.

Moreover, the users’ interests in items are chang-

ing over time. We need an algorithm that can con-

sider the dynamic aspect of this behavior when de-

tecting communities. The ongoing changes of a net-

work may be among the most interesting properties

to consider to detect more signiﬁcant and appropriate

communities. We consider for instance the evolution

of two steps (t

and t

) of a small graph composed of

ﬁve nodes represented in Figure 3. If we only have

a simple view of the last step (t

), we will not be

able to identify the appropriate community structure

in the network. This is mainly because this network

is strongly connected since it merges the whole ob-

served links during the studied period. However, if we

have an overview of the evolution process, we will be

able to detect the network community structure that

represents the real situation. Hence, by seeing t

and

, we can say that we probably have two communities

(i.e., (abc) and (de)). Consequently, a considerable

impact on recommendations is envisaged depending

on the type of community detection algorithm.

c b

Figure 3: Example of two different steps of a network evo-

lution.

In this step, we focus on community detection al-

gorithms that deal with two basic characteristics: the

dynamic behavior of the network and the overlapping

aspect of communities. In this work we focus on al-

A New Dynamic Community-Based Recommender System

129

gorithms that can support the temporal network (i.e.,

output of step 1) as input, namely iLCD (Cazabet and

Amblard, 2011) and AFOCS (Nguyen et al., 2011).

• In (Cazabet and Amblard, 2011), the authors pro-

pose an Intrinsic Longitudinal Community Detec-

tion (iLCD) that uses a multi-agent-based method

to handle the evolution of communities. The gen-

eral principle of the algorithm is essentially based

on local computations. Indeed, the communi-

ties can perceive only the nodes that they contain

and they can interact only with other communities

having at least a node in common. When they per-

ceive a topological change in their local environ-

ment (edge/node addition/removal), the candidate

communities take decisions to add/lose nodes,

split into two distinct communities, or merge with

another community based on a set of rules:

– For any link addition (i, j) in the network, if i

is in a community C and j does not belong to

C, iLCD checks if j must be integrated into C

and it ﬁnds then the set of these newly-formed

communities by the appearance of this new link

(i.e., resulted from the fusion of the modiﬁed

communities).

– For any deletion of a link, if this edge is found

within a community, iLCD checks if the con-

cerned community C loses one or many nodes

which may lead to self-division.

• In (Nguyen et al., 2011), the authors proposed

the Adaptively Finding Overlapping Community

Structure (AFOCS) algorithm that can detect

overlapping communities in a dynamic network

through two phases. The ﬁrst phase consists in

discovering all possible basic communities in the

original input network by extracting at ﬁrst all

possible densely connected sub-graphs of the net-

work and then combining those that share a sig-

niﬁcant common substructure. The second phase

consists in examining, after every network topo-

logical change, the internal density of the current

community structure to update the basic commu-

nity structure:

– If new node addition (V + v): AFOCS checks

if v should (i) be considered as outliers or (ii)

joins existing communities or (iii) forms a new

community with a substructure of the network.

– If new edge addition (E + e): AFOCS checks

if e (i) has no impact on the current structure

or (ii) allows the combination of two existing

communities or (iii) allows an existing commu-

nity to grow up.

– If node deletion (V −v): AFOCS checks if this

change (i) has no impact on the current struc-

ture or (ii) allows an existing community to

contract or to disappear.

– If edge deletion (E −e): AFOCS checks if this

change (i) has no intact on the current structure

or (ii) allows an existing community to split

into two distinct communities.

3.3 Step 3: Recommendation

This step consists in generating the recommendation

list for the active user based on the network commu-

nity structure of items (i.e., the output of step 2). The

discovered communities are exploited to (i) determine

the candidate items that can suit the active user’s latest

feedback and (ii) select the items involved to compute

the active user’s preferences prediction for these can-

didate items.

Typically, the preference prediction is computed

for each unseen item. However, it is not necessary

to handle the items that could not match the active

user’s current needs. Indeed, the user’s preferences

change over time and what was interesting before may

not be the case now. For this reason, we assume

that the active user would be interested in items sim-

ilar to his/her current taste, namely candidate items.

The candidate items are therefore identiﬁed based

on the communities of the recently appreciated item.

By doing this, (i) we avoid browsing all the unseen

items and eventually overcome the scalability prob-

lem. And (ii) we keep away from sparse cases since

there is at least one similar item already liked by the

active user. Then, the active user’s preference for each

candidate item is computed to generate the top k rec-

ommendation list.

u,i

∑

j∈C

s(i, j) r

u, j

∑

j∈C

|s(i, j)|

(1)

where C is the set of items pertaining to the communi-

ties of i, r

u, j

is the rating given by the active user u to

the item j and s(i, j) is the similarity degree between

items i and j.

There are several methods to compute the relative

similarity among items in the literature (Sarwar et al.,

2001). Basically, the similarity of items i and j is

computed by looking essentially to the co-users who

have rated both these two items. We present a brief

description of the most common similarity measures

as follows:

• Pearson correlation measures the similarity be-

tween items based on the correlation between the

co-users ratings.

s(i, j)=

∑

u∈U

u,i

−R

)(R

u, j

−R

)

√

∑

u∈U

u,i

−R

)

√

∑

u∈U

u, j

−R

)

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

130

where U is the set of co-users, R

u,i

is the rating

given by user u to item i, R

is the average ratings

given to i and R

is the average ratings given to j.

• Cosine measure estimates the similarity of two

items by looking to the distance between the fea-

ture vectors of these items.

s(i, j)=cos(i,j)=

i∗j

∥

∗

∥

• Tanimoto coefﬁcient is the implementation of Jac-

card similarity that computes the similarity be-

tween two items based on the intersection of their

preferences.

• Likelihood ratio computes the similarity between

items based on the overlapping users’ preferences.

• Adjusted cosine similarity uses the average of the

co-user’s ratings to make the correlation be-

tween the co-users’ preferences. However, in this

work we focus on computing the similarity of two

items within the same community. Intuitively, we

modify the Adjusted cosine similarity to integrate

the community structure in addition to the co-

users rating information. In fact, we replace the

mean of the user’s ratings with the mean of the

user’s ratings for items belonging to the common

communities to measure how much the co-users

like one item relative to the other similar items

they rated.

s(i, j)=

∑

u∈U

u,i

−R

u,C

i j

)(R

u, j

−R

u,C

i j

)

√

∑

u∈U

u,i

−R

u,Ci j

)

∑

u∈U

u, j

−R

u,C

i j

)

where U is the set of co-users, R

u,i

is the rating

given by user u to item i, R

u,Ci j

is the average rat-

ings given by user u to items belonging to com-

mon communities of i and j.

The preference estimation is also affected, on the

one hand, by the number of items in the same com-

munities and, on the other hand, by the number of

common communities. Algorithm 1 presents the rec-

ommendation step.

Example 3. Let us consider the network commu-

nity structure composed of 2 communities C

and C

Movies Ted2, Batman and X-men belong to C

and

movies Gladiator, Batman and Ted2 belong to C

. Ta-

ble 2 presents the rating data of Jane on the available

movies.

We can note that the Ted2 and Batman movies

are more similar than any other movies since they

have more common communities. For instance,

Batman will be considered twice in the preference

computation as it has two common communities with

Ted2 while X−men will be considered just once as it

shares one common community with Ted2. Knowing

Algorithm 1: Pseudocode for the recommendation step.

Require: : Network community structure C =





Active user u, active user’s ratings R

Ensure: : Recommendation list L

∗

← Max pre f erred(R

)

∗

←target communities(C,i

∗

)

for j ∈ candidate items(C

∗

) do

u, j

← Pre f erence(R

,C, j)

end for

L ←Max Pre f erence(P

u, j

)

% Max pre f erred(R

): Returns the most recent liked item by the user u

% target communities(C,i

∗

): Returns the communities of i

∗

% candidate items(C

∗

): Returns the set of items belonging to C

∗

and has not yet been rated by u.

% Pre f erence(R

,C, j): computes the preference prediction as the

weighted sum of the u’s ratings given to the items belonging to

the communities of j.

% Max Pre f erence(P

u, j

): Selects the top-k items with the higher

preference prediction values in the recommendation list.

Table 2: Jane’ Ratings data.

Movie title Rating(1 −5)

Batman 3

X-men 5

Ted2 ?

Gladiator ?

that the similarity between the pair of items (Ted2,

Batman) (resp. (Ted 2, X−men) and (Gladiator,

Batman)) is 0.4 (resp. 0.53 and 0.6), the preference

of Jane on movies Ted2 and Gladiator are computed

as follows:

Jane,Ted2

∑

b∈C

s(Ted2,b)∗r

Jane,b

∑

b∈C

|s(Ted2,b)|

(0.4∗3)+(0.4∗3)+(0.53∗5)

0.4+0.4+0.53

= 3.79

Jane,Gladiator

∑

b∈C

s(Gladiator,b)∗r

Jane,b

∑

b∈C

|s(Gladiator,b)|

0.6∗3

= 3

Thus, the movie Ted2 is selected as the best rec-

ommendation alternative to Jane.

4 EXPERIMENTAL STUDY

To evaluate the accuracy of the proposed framework,

we conduct experiments on two real-world data sets

A New Dynamic Community-Based Recommender System

131

namely, MovieLens data set

and Video games re-

views data set

. All the experiments are implemented

in Java JDK 1.8 on a windows 10 based PC with in-

tel Core i3 processor having a speed of 2.40 GHz and

4GB of Ram and compiled in eclipse framework. In

what follows, we ﬁrst describe the data sets and the

evaluation metrics. Then, we present the experimen-

tal results.

4.1 Datasets and Metrics

The data sets used in our experiments are:

• MovieLens data set extracted from movies recom-

mender system containing 10 million of ratings

collected from 72 000 users on 10 000 movies

(items) from 2000 to 2009. Each user has rated

at least 20 items and made a rating at ﬁve discrete

levels from 1 to 5.

• Video game reviews data set extracted from Ama-

zon including 1 million ratings spanning from

June 1997 to July 2014. The ratings are collected

from 129 496 customers on more than 1371 video

games, and they are on a numeric ﬁve-point scale

(i.e., from 1 to 5).

Each data set is divided into two subsets: a train-

ing set and a testing set. The training set is used to

construct the dynamic network of items, detect com-

munities and ﬁnally, predict the user’s rating for an

item to generate the recommendation list. The test-

ing set is used to evaluate the closeness of the pre-

dicted rating compared to the user’s provided rating.

To this end, we ﬁrst split the data chronologically

into testing (20% of the recent instances) and training

(the remaining ones). The efﬁciency of the proposed

framework is evaluated using classiﬁcation accuracy

measure, namely F-measure and predictive accuracy

metric, namely Mean Absolute Error (MAE) (Chena

and Liu, 2017). F-measure denotes the weighted har-

monic mean of the precision and recall of the gener-

ated recommendations. MAE measures the deviation

of the estimated preference from the true preference

value speciﬁed by the active user. The lower the MAE

is, the better the generated recommendations are. For-

mally:

F −measure =

2 ∗Precision ∗Recall

Precision + Recall

Precision =

RR + FR

Recall =

RR + RN

https://movieLens.umn.edu

https://snap.stanford.edu/data/web-Amazon.html

where RR denotes relevant and recommended items,

RN denotes relevant and not recommended items,

and FR denotes not relevant and recommended items.

MAE =

∑

i=1

−q

where p

denotes the predicted rating for item i, q

denotes the true rating for item i, n: the corresponding

ratings-prediction pairs.

4.2 Experimental Protocol

Before presenting the experimental results and com-

paring them with state-of-the-art recommendation al-

gorithms, we propose to test the impact of some pa-

rameters on the quality of the proposed framework.

The parameters representing the optimum values are

then taken for the rest of the experiment.

4.2.1 The Parameters N and P

To detect the values that can capture communities

with the most interesting size and internal topologies,

we count, for each pair of items, the number of in-

teractions N and the corresponding period P. Subse-

quently, we select various signiﬁcant values (i.e., the

minimal, the median, the 3rd quartile, and the maxi-

mal) for these parameters as follows:

• For MovieLens data, we found 171 787 pairs of

movies interacted together. The maximal num-

ber of interactions N

max

(resp. the minimal N

min

the median N

med

, and the 3rd quartile N

) is 414

(resp. 1, 2 and 2). The maximal period of days

max

(resp. the minimal P

min

, the median P

med

and the 3rd quartile P

) is 1035 (resp. 1, 119 and

517).

• For video game data, we found 3580 pairs of

video games that have interacted together. The

maximal number of interactions N

max

(resp. the

minimal N

min

, the median N

med

, and the 3rd quar-

tile N

) is 55 (resp. 1, 1, 2). The maximal period

of days P

max

(resp. the minimal P

min

, the median

med

, and the 3rd quartile P

) is 5329 (resp. 1,

2290 and 2664).

A temporal network is constructed for each pos-

sible combination between P and N (i.e., 4

tempo-

ral networks built for each dataset). We ﬁrst apply

iLCD and AFOCS to detect dynamic communities

from each temporal network. The resulting communi-

ties are then used to compute recommendations. We

kept only the valid temporal network of items (i.e.,

with existing links and a non-empty community struc-

ture) since it is hard to ﬁnd connections that hold on

between nodes if N is maximal and/or P is minimal.

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

132

Table 3: The impact of N and P on the recommendation

quality in MovieLens.

[N,P]

Algo

iLCD AFOCS

MAE F −

measure

MAE F −

measure

[1,119] 0.68 0.267 0.88 0.15

[1,517] 0.6 0.29 0.7 0.2

[1,1035] 0.56 0.299 0.66 0.22

[2,517] 0.683 0.277 0.81 0.154

[2,1035] 0.48 0.39 0.57 0.258

We present in Tables 3 and 4 the accuracy of the rec-

ommendation for different community detection algo-

rithms and different temporal networks in MovieLens

and Video Game reviews data sets, respectively. Table

3 shows that the prediction accuracy in MovieLens

dataset reaches its maximum when N is ﬁxed to 2 and

P is ﬁxed to 1035. While Table 4 shows that the pre-

diction accuracy in the Video Games data set reaches

its maximum when N is equal to 2 and P is equal to

5329. Thus, we conclude that the prediction accu-

racy reaches its maximum when N is the median value

and P takes the maximum for both databases. Indeed,

we need to consider the most frequent value of item-

item interactions and the maximal period P to keep

the highest number of connected nodes. We can also

observe from Tables 3 and 4 that by using the dynamic

iLCD algorithm the recommendation quality reaches

its maximum compared to AFOCS. This is justiﬁed

by: (i) its aptitude for considering the network evolu-

tion over time that has a major impact on the detected

network community structure and (ii) its ability to dis-

tinguish between representative nodes and hubs (i.e.,

highly centralized nodes) to decide against merging

communities with many outside neighboring nodes.

Example 4. For instance, if we consider Figure

4. Titanic, Troy, and Tourist are 3 movies of the

same community C

(i.e., many users have evaluated

Tourist similarly as Titanic and Troy). Moreover,

movies Belle, Titanic and Troy belong to the same

community C

). However, Tourist and Belle cannot

be similar (i.e., be in the same communities C

∪C

)

just because both are similar to Titanic and Troy.

These latter are popular legendary movies and can

be appreciated by many users with different tastes.

Hence, we chose to use dynamic iLCD algorithm

(Cazabet and Amblard, 2011) to detect communities

in the remaining experiments. Figure 5 shows two

snapshots of the network community structure de-

tected by iLCD with the aforementioned optimal re-

sults of N and P.

Table 4: The impact of N and P on the recommendation

quality in Video games.

[N,P]

Algo

iLCD AFOCS

MAE F −

measure

MAE F −

measure

[1,2290] 1.22 0.263 1.34 0.19

[1,2664] 1.2 0.27 1.29 0.194

[1,5329] 1.19 0.284 1.268 0.2

[2,2290] 1.4 0.192 1.6 0.166

[2,2664] 1.02 0.293 1.24 0.225

[2,5329] 0.99 0.32 1.18 0.3

𝐶

Figure 4: Example of a network structure.

4.2.2 The Effect of the Similarity Measure

In this experiment, we choose to test the similarity

measures namely, Pearson correlation, Cosine, Tani-

moto coefﬁcient, Likelihood ratio and adjusted cosine,

as they are commonly used in recommender systems

in our data sets, and study their effects on the obtained

prediction quality. We have implemented the Ad-

justed cosine similarity measure, and we use Apache

Mahout’s implementations of other measures. Figure

6 shows the evaluation results.

We can see that for both MovieLens and Video

Games data sets, using the adjusted similarity in the

proposed framework achieves the lowest MAE and

the highest F-measure value compared to other simi-

larity measures. This result can be justiﬁed by the fact

that is more interesting to focus on how much the co-

users liked similar items instead of all the other items

to measure the similarity of items of the same com-

munities. Therefore, the adjusted cosine similarity is

used in what follows.

4.3 Comparison with Related Work

In this section, we propose to compare the recom-

mendation accuracy of our dynamic and overlapping

community-based recommendation framework with

existing related work approaches presented in Table

A New Dynamic Community-Based Recommender System

133

(a) MovieLens.

(b) Video Games reviews.

Figure 5: Network Community structure of iLCD with

med

max

The static community-based CF (so-called

S.community) (Fatemi and Tokarchuk, 2013) em-

ployed the Louvain method to extract communities

of movies. The static clustering-based CF (so-called

S.cluster) (O’Connor and Herlocker, 2001) consists

in using the Average link clustering algorithm to

group items into clusters based on the similarity

matrix. The dynamic clustering-based CF (so-called

D.cluster) (Hamzaoui et al., 2012) included a time

weight parameter in item-item similarity measure

to penalize the impact of historical data and then

take into account the dynamics of the active user’s

purchase habits. The timeSVD++ (Yehuda, 2009)

refers to the time-sensitive algorithm that considers

temporal information beyond the rating matrix by

introducing time-variant biases for each user and

each item. The idea is to decay the weight of the

user’s old ratings to have no effect on his status at the

current time.

Figure 7 displays the prediction accuracy results

over the state-of-the-art recommendation methods.

First, the prediction quality of S.cluster is un-

surprisingly worse than item-based CF as stated by

the literature (O’Connor and Herlocker, 2001). For

instance, applied to the MovieLens data set, the

clustering-based approach yields an MAE of 0.77

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

Adjusted Tanimoto Pearson Cosine Likelihood

MAE F-measure

(a) MovieLens data set.

0,2

0,4

0,6

0,8

1,2

Adjusted Tanimoto Pearson Cosine Likelihood

MAE F-measure

(b) Video Games data set.

Figure 6: Sensitivity of similarity measures.

Table 5: State-of-the-art methods.

Name

Overlapping Time Method

item-based No No KNN

S.community Yes No Louvain

S.cluster No No Average link

D.cluster No Yes K-means

TimeSVD++ No Yes −

and the item-based CF yields an MAE of 0.739.

Moreover, applied to the Video Games data set, the

clustering-based approach gives a higher MAE (i.e.,

1.42) and thus a lower accuracy compared to the ba-

sic item-based (i.e., 1.39). The decrease in accuracy

could be due to the fact that the clustering algorithms

group items to be exclusively in one cluster. However,

certain items may have signiﬁcant predictive value

for multiple clusters. Subsequently, the predictions

computed using a static and overlapping community

detection approach (S.community) are more accurate

than the basic item-based. This is explained by the

fact that the overlapping aspect of the items subgroups

(i.e., an item may be a member of more than one

community) may reveal interesting neighbors selec-

tion (i.e., containing similar items with different pop-

ularity and preferences). It can also be observed from

Figure 7) that the D.cluster, in both data sets, provides

a better quality of prediction compared to both item-

based and S.community because it considers the tem-

poral information of users’ preferences when com-

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

134

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,9

S.cluster item-based S.community D.cluster TimeSVD++ DO2CF

MAE F-measure

(a) MovieLens data set.

0,2

0,4

0,6

0,8

1,2

1,4

1,6

S.cluster item-based S.community D.cluster TimeSVD++ DO2CF

MAE F-measure

(b) Video Games data set.

Figure 7: The performance of the state-of-the-art recom-

mendation algorithms.

puting the similarity between items (i.e., by penaliz-

ing the impact of old rating data and highlighting the

recent data rating). However, the time-sensitive al-

gorithm TimeSVD++ performs better than D.cluster

since it considers the temporal dimension in a large

space (i.e., not only when computing the similarity).

Finally, the proposed framework for overlapping and

dynamic community-based CF (DO2CF) has a lower

MAE value (i.e., 0.48 in MovieLens and 0.99 in the

Video games data set) compared to D.cluster, and this

meant that it yields the best recommendation accu-

racy in comparison to all the other CF methods. This

is justiﬁed by the fact that by considering both the

temporal changes in users’ interests and the overlap-

ping aspect of items, the detected communities can

better reveal the items that represent the most recent

related neighbors. Accordingly, the quality of the

community structure has an impact on recommenda-

tion accuracy. Furthermore, DO2CF performs better

than TimeSVD++ since the latter aims to capture tem-

poral effects by incorporating a time-variant bias for

each user and item at every time window. Indeed,

TimeSVD++ cannot track the underlying pattern for

different biases of time windows.

5 CONCLUSIONS

In this paper, we propose a new framework for dy-

namic and overlapping community-based CF that al-

lows (i) modeling the dynamic aspect of users’ pref-

erences in a recommender system into a dynamic

network, (ii) identifying dynamic and overlapping

groups of items using a community-detection algo-

rithm, and (iii) providing a diverse recommenda-

tion list that could match the current users’ tastes.

Through this work, we highlight that by consider-

ing the network topology changes and the overlap-

ping membership of items, the detected communities

are more signiﬁcant and appropriate. Consequently,

a considerable impact on the recommendations’ qual-

ity is envisaged depending on the type of community

detection algorithm. In addition to the recommenda-

tion accuracy improvement, the experimental results

for both MovieLens and Video Games data sets indi-

cate that (DO2CF) can scale with large data and ad-

dress sparsity.

REFERENCES

Abrouk, L., Gross-Amblard, D., and Cullot, N. (2010).

Community detection in the collaborative web. Inter-

national Journal of Managing Information Technol-

ogy (IJMIT), 2(4).

Aston, N., Hertzler, J., and Hu, W. (2014). Overlapping

community detection in dynamic networks. Journal

of Software Engineering and Applications, 7(10):872–

882.

Cazabet, R. and Amblard, F. (2011). Simulate to detect:

a multi-agent system for community detection. 2011

IEEE/WIC/ACM International Conference on Web In-

telligence and Intelligent Agent Technology (WI-IAT),

2:402–408.

Chena, M. and Liu, P. (2017). Performance evaluation of

recommender systems. International Journal of Per-

formability Engineering, 13(8):1246–1256.

Deng, W., Patil, R., Najjar, L., Shi, Y., and Chen, Z. (2014).

Incorporating community detection and clustering

techniques into collaborative ﬁltering model. The 2nd

International Conference on Information Technology

and Quantitative Management (ITQM’14), 31:66–74.

Fatemi, M. and Tokarchuk, L. (2013). A community

based social recommender system for individuals and

groups. 2013 International Conference on Social

Computing (SocialCom), pages 351–356.

Fortunato, S. (2010). Community detection in graphs.

Physics reports, 486(3):75–174.

Guo, L. and Peng, Q. (2014). A neighbor selection method

based on network community detection for collabo-

rative ﬁltering. the 13th International Conference on

Computer and Information Science (ICIS’14), (143-

146).

A New Dynamic Community-Based Recommender System

135

Hamzaoui, N., Sedqui, A., and Lyhyaoui, A. (2012). Multi-

criteria collaborative recommender. International

journal of computational linguistics research, 3(3).

Newman, M. E. J. and Girvan, M. (2004). Finding and eval-

uating community structure in networks. Phisical re-

view E, 69(2).

Nguyen, N. P., Dinh, T. N., Tokala, S., and Thai, M.

(2011). Overlapping communities in dynamic net-

works: Their detection and mobile applications. Pro-

ceedings of the 17th annual international conference

on Mobile computing and networking, pages 85–96.

O’Connor, M. and Herlocker, J. (2001). Clustering items for

collaborative ﬁltering. Proceedings of the ACM SIGIR

Workshop on Recommender Systems: Algorithms and

Evaluation.

Palla, G., Farkas, I., and Vicsek, T. (2005). Uncovering

the overlapping community structure of complex net-

works in nature and society. Nature, pages 814–818.

Qiang, H. and Yan, G. (2012). A method of personalized

recommendation based on multi-label propagation for

overlapping community detection. the 3rd Interna-

tional Conference on System Science Engineering De-

sign and Manufacturing Informatization, 1:360–364.

Qin, S., Menezes, R., and Silaghi, M. (2010). A recom-

mender system for youtube based on its network of

reviewers. the IEEE International Conference on So-

cial Computing, pages 323–328.

Sahebi, S. and Cohen, W. (2011). Community-based rec-

ommendations: a solution to the cold start problem.

Workshop on Recommender Systems and the Social

Web (RSWEB).

Sarwar, B. M., Karypis, G., Konstan, J., and Riedl, J.

(2001). Item-based collaborative ﬁltering recommen-

dation algorithms. the 10th international conference

on world wide web (www’01).

Sarwar, B. M., Karypis, G., Konstan, J., and Riedl, J.

(2002). Recommender systems for large-scale e-

commerce: Scalable neighborhood formation using

clustering. Proceedings of the 5th International Con-

ference on Computer and Information Technology

(ICCIT).

Vlachos, M., Fusco, F., Mavroforakis, C., Kyrillidis, A. T.,

and Vassiliadis, V. G. (2014). Improving co-cluster

quality with application to product recommendations.

Proc. CIKM, pages 679–688.

Xie, J., Chen, M., and Szymanski, B. K. (2013). Label-

rankt: Incremental community detection in dynamic

networks via label propagation. CoRR.

Xin, L., Haihong, E., Junde, S., Meina, S., and Junjie, T.

(2014). Book recommendation based on community

detection. Pervasive Computing and the Networked

World, pages 364–373.

Yehuda, K. (2009). Collaborative ﬁltering with temporal

dynamics. Proceedings of the 15th ACM SIGKDD In-

ternational Conference on Knowledge Discovery and

Datamining, pages 447–456.

Yin, B., Yang, Y., and Liu, W. (2014). Exploring social

activeness and dynamic interest in community-based

recommendation system. Proceedings of the 23rd

International Conference on World Wide Web, pages

771–776.

Ying, J.-C., Shi, B.-N., Tseng, V., Tsai, H.-W., Cheng,

K. H., and Lin, S.-C. (2013). Preference-aware com-

munity detection for item recommendation. 2013

Conference on Technologies and Applications of Ar-

tiﬁcial Intelligence (TAAI), pages 49–54.

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

136