Experimental Analysis of Pipelining Community Detection and

Recommender Systems

Ryan Dutra de Abreu

1 a

, Laura Silva de Assis

1 b

and Douglas O. Cardoso

2 c

Celso Suckow da Fonseca, Federal Center of Technological Education, Rio de Janeiro, RJ, Brazil

Smart Cities Research Center, Polytechnic Institute of Tomar, Tomar, Portugal

Keywords:

Collaborative Filtering, Social Networks, User Proﬁling, Collective Behavior.

Abstract:

Community detection and recommender systems are two subjects of the highest relevance among data-oriented

computational methods, considering their current applications in various contexts. This work investigated how

pipelining these tasks may lead to better recommendations than those obtained without awareness of implicit

communities. We experimentally assessed various combinations of methods for community detection and

recommendation algorithms, as well as synthetic and real datasets. This targeted to unveil interesting patterns

in the behavior of the resulting systems. Our results show that insights into communities can signiﬁcantly

improve both the effectiveness and efﬁciency of recommendation algorithms in some favorable scenarios.

These ﬁndings can be used to help data science researchers and practitioners to better understand the beneﬁts

and limitations of this methodology.

1 INTRODUCTION

The Internet’s signiﬁcant growth in the last three

decades has transformed how society conducts busi-

ness and communicates (Castells and Cardoso, 1996).

By 2020, its widespread usage had become a global

norm (Roser et al., 2015), resulting in new rela-

tionship structures, habits, consumption patterns, and

markets (Paul and Aithal, 2018).

Although beneﬁcial in many aspects, this digital

evolution has led to information overload for users,

making it challenging to ﬁnd relevant data and make

informed decisions (Dias, 2014). In this context, rec-

ommendation systems offer promise in managing this

overload, simplifying decision-making by suggesting

relevant items (Panteli et al., 2019).

Grounded in cognitive science and information re-

trieval, this systems are crucial for generating person-

alized suggestions based on user preferences (Rich,

1979; Sanderson and Croft, 2012). They can employ

various approaches, including collaborative ﬁltering,

content-based ﬁltering, and hybrid ﬁltering (Burke,

2000; Adomavicius and Tuzhilin, 2005; Pazzani and

Billsus, 2007; Schafer et al., 2007).

https://orcid.org/0009-0007-8339-1631

https://orcid.org/0000-0003-3081-9722

https://orcid.org/0000-0002-1932-334X

Collaborative ﬁltering (CF) techniques are widely

used due to its efﬁciency, recommending items based

on similar users’ preferences (Gorripati and Vat-

savayi, 2017). Yet, many commercial systems rely

on large, sparse user-item matrices, posing challenges

such as the “cold-start” problem (Guo, 2013).

To tackle these challenges, researchers have ex-

plored strategies like incorporating additional data

to model user preferences, introducing new similar-

ity measures, and using community detection algo-

rithms to uncover network structures (Guo, 2013;

Keerthi Gorripati and Angadi, 2018; Xie and Szy-

manski, 2013; Bourhim et al., 2018; Bhaskaran et al.,

2022). While integrating community detection with

recommender systems has shown promise in enhanc-

ing recommendation accuracy, there is still a wide

range of experimental opportunities in the ﬁeld.

This study investigates combining these tasks to

improve recommendations through implicit commu-

nity awareness. We explore novel algorithm combi-

nations, evaluating their efﬁciency and effectiveness

with synthetic and real datasets to provide insights for

data science researchers and practitioners.

Remaining sections include an overview of re-

lated literature (Section 2), details of the community-

oriented recommendation methodology (Section 3),

experimental descriptions and results (Section 4), and

a research summary (Section 5).

496

Dutra de Abreu, R., Silva de Assis, L. and Cardoso, D.

Experimental Analysis of Pipelining Community Detection and Recommender Systems.

DOI: 10.5220/0012237000003584

In Proceedings of the 19th International Conference on Web Information Systems and Technologies (WEBIST 2023), pages 496-503

ISBN: 978-989-758-672-9; ISSN: 2184-3252

2 RELATED WORK

Integrating community detection algorithms with rec-

ommender systems has gained recent attention. These

approaches use data structure and relationships to en-

hance recommendation accuracy and diversity. This

section provides an overview of relevant works in this

area.

Pham et al. (Pham et al., 2011) proposed a cluster-

ing approach that improves recommendation quality

by incorporating social information. Their user-based

algorithms make predictions by aggregating ratings

from the active user’s nearest neighbors with weights.

The technique outperformed traditional CF methods

in evaluations using DBLP and Epinion data.

A graph-based CF approach treating users as a

network of highly similar sub-communities was pre-

sented in (Bourhim et al., 2018; Bourhim et al., 2019).

By utilizing key-nodes and the Louvain modularity-

based algorithm to group users, signiﬁcant perfor-

mance gains were achieved compared to classical rec-

ommendation approaches.

Ciftc¸i et al. (C¸ iftc¸i et al., 2021) used Louvain and

Label Propagation community detection algorithms to

identify artist clusters, aligning well with Spotify’s

artist lists. Additionally, Kherad and Bidgoly intro-

duced a recommendation method (Kherad and Bid-

goly, 2022) combining matrix factorization, graph

analysis, and deep learning techniques. They em-

ployed the Louvain algorithm for community detec-

tion and HITS for identifying important nodes in the

user communication network, resulting in superior

performance compared to existing methods.

In (Patra et al., 2014), a similarity-based CF

method was introduced to address the user cold-start

issue in sparse data situations, delivering reliable rec-

ommendations for new users with limited ratings,

substantially improving accuracy compared to exist-

ing solutions.

Similarly, Gorripati and Vatsavayi (Gorripati and

Vatsavayi, 2017) proposed a community-based CF

approach that tackles both data sparsity and cold-

start problems. By examining user preferences within

the same community, they enhance recommendation

quality. Their method involves identifying user pat-

terns, creating a network based on co-rating relation-

ships, identifying community structures, and comput-

ing neighborhood similarity.

In the realm of music recommendation sys-

tems, (Cao et al., 2020) addressed data sparsity and

cold-start issues. The authors introduced a com-

munity detection algorithm based on the Louvain

method, which exhibited higher accuracy, stability,

and shorter running times compared to alternative ap-

proaches, validating the effectiveness of their cold-

start algorithm.

While these works have shown the potential of in-

tegrating community detection algorithms with rec-

ommender systems to enhance recommendation ac-

curacy and address challenges, they have some limita-

tions. Furthermore, there is still a research gap in ex-

ploring the combinations of community discovery and

recommendation algorithms, as well as their impact

on system performance. Our work aims to advance

this integration by exploring novel algorithm combi-

nations, assessing their effectiveness and efﬁciency

using synthetic and real datasets, and gaining insights

into system behavior. This contributes to more effec-

tive and personalized recommendation systems.

3 PROPOSED METHODOLOGY

This work’s core inspiration is to enhance recommen-

dation algorithms using community detection tech-

niques. This approach is highly adaptable and can

be employed with a range of prediction algorithms,

including heuristics, neighborhood methods, matrix

factorization, and other CF techniques. The method-

ology involves four steps, which will be detailed in

the following subsections.

3.1 Graph Projection

In this subsection, we outline the process of trans-

forming a bipartite graph of user-item ratings into

a unipartite weighted graph focused solely on users.

This projection enables us to capture user relation-

ships and interactions while considering the weighted

connections derived from their ratings.

Consider a weighted bipartite graph G =

(U, I , E, r) where U is the set of users nodes, I is

the set of items nodes representing the catalog items,

E is the set of edges connecting nodes from U to I

and r represents the ratings, i.e., the weights assigned

to each edge. It can be evaluated the dissimilarity

between two users u

and u

, where (u

, u

) ∈ U

by iterating over the items that both users have rated

and evaluating the mean squared difference between

their corresponging ratings. This will result a unipar-

tite weighted graph G

′

= (U

′

, E

′

, w), where U

′

rep-

resents the set of nodes in the weighted projection –

which corresponds to the set of nodes U in the bi-

partite graph, E

′

represents the set of edges in the

weighted projection, capturing the connections be-

tween pairs of nodes in U

′

which rated at least one

item in common, and w represents the weight func-

tion in the weighted projection, assigning numerical

Experimental Analysis of Pipelining Community Detection and Recommender Systems

497

Figure 1: Example of a bipartite graph: (a) representing

user-item relationships and its corresponding projection (b)

showing the connections among users based on their rat-

ings.

values to the edges in E

′

Mathematically, the weight w

is deﬁned as

shown in Equation (1):

|G[u

] ∩ G[u

∑

i∈G[u

]∩G[u

]

− r

)

(1)

where, G[u

] and G[u

] represent the sets of items

rated by users u

and u

, respectively. r

and r

represent the ratings of i-th item given by users u

and

, respectively. To give the just mentioned weights

a positive connotation, we apply the transformation

shown in Equation (2), which also acts as a normal-

ization to the (0, 1] interval:

′

1 + w

(2)

Larger w values result in smaller normalized

weights w

′

, indicating lower user similarity, while

smaller w values lead to larger normalized weights,

indicating higher similarity. These weighted connec-

tions reﬂect user relationships in the bipartite graph

and serve as the basis for constructing the projected

user-user graph. This transformation helps us focus

on user-user relationships for recommendation pur-

poses.

Figure 1 visually illustrates a bipartite graph G

and its corresponding projection G

′

, generated from a

ﬁctitious dataset of 10 users and 10 items with ratings

from 1 to 5. The left side shows user-item relation-

ships, while the right side depicts connections among

users based on their ratings.

3.2 Community Detection

Community detection aims to identify clusters of

users within the graph who exhibit similar patterns or

behavior. By uncovering these communities, we can

Figure 2: Graph projection with identiﬁed communities.

gain insights into the underlying structure and dynam-

ics of user interactions, which can further enhance the

effectiveness of our recommendation algorithm.

To detect communities in the projected graph, we

can employ various community discovery algorithms

in order to identify densely connected groups (Choud-

hary et al., 2023). In the context of recommenda-

tion systems, particular algorithms have been widely

used for community detection, like Louvain (Blondel

et al., 2008), Girvan-Newman (Girvan and Newman,

2002), Infomap (Rosvall and Bergstrom, 2008) and

Label Propagation (Raghavan et al., 2007).

This clusting process is illustrated in Figure 2,

where the Louvain algorithm is applied to the pro-

jected graph generated in the previous subsection

(Figure 1(b)). This application results in a graph visu-

alization where each node is assigned a speciﬁc color

representing its community membership.

One can gain valuable insights into the users’ col-

lective behavior through community detection tech-

niques, leading to a more personalized and satis-

factory user experience. Once communities are de-

tected, they represent user subsets with shared charac-

teristics, allowing for personalized recommendations

based on community afﬁliations.

3.3 Model Instantiation and Training

This section focuses on enhancing the recommenda-

tion model training phase using community informa-

tion. After completing graph projection and commu-

nity detection, we aim to incorporate community in-

sights to capture collective user behavior and prefer-

ences.

Let us denote the training dataset as D, the rec-

ommendation algorithm as A, and the set of commu-

nities as C . The community-based recommendation

model training process can be outlined as follows.

Consider a dataset subset D

is created for each com-

munity C

∈ C, containing ratings and preferences of

users in that community. Using D

, the recommen-

WEBIST 2023 - 19th International Conference on Web Information Systems and Technologies

498

dation model M

which is associated with commu-

nity C

can be trained using the recommendation al-

gorithm A. The model M

learns from the user-item

interactions within D

, capturing the preferences and

patterns speciﬁc to community C

After training the model M

, it is stored as the

trained model speciﬁc to community C

. All this

process is repeated for each community C

∈ C ,

resulting in trained recommendation models M =

, M

, ··· , M

|C|

} for each of all |C| communities.

The Algorithm 1 summarizes the steps involved in

community-based recommendation model training.

1 for community C

∈ C do

2 Prepare the training dataset D

, the subset

of D regarding users in C

;

3 Train the recommendation model M

using D

and algorithm A;

4 Store the trained model M

speciﬁc to C

;

5 end

6 return Trained recommendation models

, M

, ·· · , M

|C|

for each community

Algorithm 1: Community-Based-RMT(D, A, C ).

As an illustrative example, for the two-

communities ﬁctitious projected graph mentioned

earlier (Figure 2), this iterative training process

would generate two training instances: one for the

ﬁrst community (users 1, 3, 4, 5, 6, 9, and 10) and

another for the second community (users 2, 7, and

8). By incorporating community information into

training, we enhance the accuracy and effectiveness

of the recommendation model. It learns not only

from individual user-item interactions but also

from collective behavior within each community,

capturing nuanced preferences and improving overall

recommendations.

3.4 Rating Prediction

This section covers the prediction phase of the pro-

posed recommendation methodology. Once the list of

communities and a trained model for each community

is obtained, the next phase can predict recommenda-

tions for any user based on their community member-

ship.

For a user u for whom we want to provide recom-

mendations, we can ﬁnd the community C

to which

user u belongs. Once community C

is identiﬁed, we

retrieve the trained recommendation model M

spe-

ciﬁc to community C

. Using model M

, we can gen-

erate personalized recommendations for user u. The

model utilizes user-item interactions and the learned

patterns within the community to provide recommen-

dations tailored to the preferences and behavior of

users in that community.

The prediction process involves using the trained

model M

to calculate the predicted ratings or scores

for items that user u has not yet interacted with.

This can be done using various techniques, such as

CF, matrix factorization, or deep learning-based ap-

proaches. The speciﬁc method depends on the rec-

ommendation algorithm chosen during the training

phase. Once the predicted ratings or scores are ob-

tained, it is possible to classify the items or select

the best-ranked recommendations to present to user

u. These recommendations are based on the pref-

erences and interests of users within the same com-

munity as u, capturing the collective behavior and

preferences of that community. Algorithm 2 summa-

rizes the prediction process for a single user, taking

as input a user u, the set of communities C, where

C = {C

, C

, ·· · , C

|C|

}, and the trained recommenda-

tion models M = {M

, M

, ·· · , M

|C|

1 C

← ﬁndCommunity(u);

2 M

← Trained recommendation model

speciﬁc to C

;

3 for each item i not rated by user u do

4 score

← predictScore(M

, u, i);

5 end

6 Sort items by score in descending order;

7 return top-k recommendations for user u

Algorithm 2: Community-Based-RP(u, C , M , k).

By leveraging community structures and trained

recommendation models, we can offer personalized

recommendations inﬂuenced not just by individual

user behavior but also by their community’s behavior.

This captures community-speciﬁc nuances, leading to

more accurate, relevant recommendations.

4 EXPERIMENTAL

ASSESSMENT

This section covers experiments evaluating the pro-

posed community-based recommendation methodol-

ogy. We conducted experiments on synthetic and real

datasets to gauge its effectiveness and robustness. We

also compared the results achieved with this approach

to default settings of established recommendation al-

gorithms to assess the impact of community detec-

tion. For transparency and reproducibility, the source

code is publicly accessible

https://github.com/ryan-dutra/Recommender-System/

Experimental Analysis of Pipelining Community Detection and Recommender Systems

499

4.1 Datasets

The experiments were conducted using both a syn-

thetic dataset generated speciﬁcally for this study, and

a well-known real-world dataset, Movielens 100k

The choice of these datasets was motivated by the

need to evaluate the performance and robustness of

our community-based recommendation methodology

across different data scenarios.

The synthetic dataset was generated to simulate a

graph with a community structure, enabling the estab-

lishment of ground truth regarding community mem-

berships and user-item preferences. Creating this syn-

thetic dataset allowed for a systematic study of how

community detection affects recommendation perfor-

mance. The dataset generation employed a custom

function for simulating a bipartite graph, incorporat-

ing parameters like the number of users, communities,

items, categories, as well as connection probabilities

and strengths. These parameters enabled the creation

of synthetic datasets with desired community struc-

tures and interaction patterns.

To generate the communitarian recommendation

graph, speciﬁc steps were followed. First, community

labels were randomly assigned to users while ensur-

ing sorted labels for consistency, representing users’

community memberships. Community size distribu-

tion was calculated by counting users in each com-

munity. Similarly, category labels were randomly as-

signed to items, with sorted labels for consistency,

and category size distribution was computed based

on the assigned labels, representing item distribution

across categories.

An afﬁnity matrix L was constructed to capture

the likelihood of users liking items in speciﬁc cate-

gories, with entries generated randomly based on a

given probability α. This matrix quantiﬁed the de-

gree of afﬁnity between communities and categories,

forming the basis for modeling user-item interactions.

To consider varying liking strengths, the afﬁnity

matrix L was transformed into an adjusted afﬁnity

matrix L

′

. Values in L

′

were scaled by the liking

strength parameter β, which controlled the overall im-

pact of afﬁnity on recommendations. Additionally, re-

maining values in L

′

were adjusted to

(1−β)

, ensuring

a balanced inﬂuence of afﬁnity across communities

and categories.

By combining the adjusted afﬁnity matrix L

′

with

matrices of zeros, a probability matrix P was con-

structed. This matrix represented the probabilities of

user-item interactions based on community and cate-

gory afﬁliations. Leveraging user and item distribu-

tions along with the probability matrix P, a stochastic

https://movielens.org/

block model graph (sbg) was created using the Net-

workX library (Hagberg et al., 2008). The stochas-

tic block model graph depicted the recommendation

graph, with block sizes corresponding to user and

item distributions.

4.2 Experimental Protocol

To evaluate our community-based recommendation

methodology, we divided the datasets into training

and testing sets. Various test sizes were considered,

including 25%, 10%, and 1% of total ratings. We em-

ployed random permutation cross-validation to ensure

reliable and unbiased evaluations. This technique in-

volves shufﬂing the dataset and then splitting it into

multiple training and testing subsets, mitigating po-

tential biases due to data ordering. The scikit-learn

library’s random permutation cross-validation imple-

mentation was used with ﬁve splits to generate diverse

training and testing subsets, ensuring robust perfor-

mance assessment.

In each training instance, we followed several

steps for the recommendation process. First, we trans-

formed the bipartite graph into a projected graph, as

described in Section 3, focusing on user-user and

item-item relationships. Community detection algo-

rithms, speciﬁcally Louvain (Blondel et al., 2008) and

Paris (Xie et al., 2013), were applied to identify com-

munities based on connectivity patterns between users

and items.

Once communities were identiﬁed, recommenda-

tion models were instantiated and trained separately

for each community. We evaluated ﬁve recommenda-

tion algorithms available in the Surprise library (Hug,

2020), categorized as follows:

1. Matrix Factorization Algorithms: SVD and

NMF.

2. Neighborhood-Based Algorithms: k-NN and

Slope One.

3. Clustering Algorithm: Co-Clustering.

Additionally, the NormalPredictor algorithm

served as a random baseline. Trained models were

used to generate predictions for user-item pairs in

the training data, with each model speciﬁc to its cor-

responding community, following the approach out-

lined in Section 3.

To assess performance, we employed the RMSE

metric, quantifying the accuracy of recommendation

models in predicting user-item ratings. We also eval-

uated the computational cost, including training and

testing times in seconds, to assess the efﬁciency of

our community-based recommendation approach.

WEBIST 2023 - 19th International Conference on Web Information Systems and Technologies

500

Figure 3: Predictive performance considering various datasets, community detection approaches and recommendation algo-

rithms.

Figure 4: Training time considering various datasets, community detection approaches and recommendation algorithms.

By conducting experiments on synthetic and real-

world datasets, we aimed to comprehensively evalu-

ate our approach’s performance across various data

settings. Subsequent subsections present and discuss

the results obtained.

4.3 Results and Discussion

This section presents computational experiment re-

sults evaluating the community-based recommenda-

tion approach’s performance, comparing it with de-

fault recommendation methods, and analyzing effec-

tiveness and computational cost across various sce-

narios.

To evaluate the community-based recommenda-

tion approach, tests were conducted on both synthetic

and real datasets, as previously explained. Figure 3

illustrates the performance comparison between the

proposed approach (utilizing Louvain or Paris algo-

rithms for community discovery) and the default al-

ternative (without a community discovery algorithm).

The bar heights represent the average RMSE of the

iterations, while the error bands show the variability

or uncertainty in these averages, calculated using the

standard deviation.

In the synthetic dataset scenario, our community-

based recommendation approach consistently outper-

formed the default method (without a community

detector), especially with neighborhood-based algo-

rithms, which showed signiﬁcant performance im-

provement when combined with community detec-

tion.

In the real dataset scenario (Movie Lens 100k),

our approach generally performed similarly to the de-

fault method with slight differences. The only ex-

ception was the Normal Predictor algorithm, which

showed lower RMSE when community detectors

were used.

These results indicate that community detec-

tors can beneﬁt recommendation tasks, especially

for datasets with clear network structures. Addi-

tionally, neighborhood-based algorithms showed the

most signiﬁcant improvement when integrated with

our community-based approach. This highlights the

potential effectiveness of combining neighborhood-

based algorithms with community-based recommen-

dation strategies.

In addition to performance evaluation, the com-

Experimental Analysis of Pipelining Community Detection and Recommender Systems

501

Figure 5: Prediction time considering various datasets, community detection approaches and recommendation algorithms.

putational cost of the community-based recommen-

dation approach was also analyzed. Two aspects were

considered: i) training time, and ii) prediction time.

Figure 4 and Figure 5 display the training and predic-

tion time for each algorithm in different scenarios for

both datasets. The bar heights correspond to the me-

dian execution times, while the error bands depict the

uncertainty associated with these median values, cal-

culated using the Median Absolute Deviation (MAD).

While the results for synthetic data indicated a

slight increase in computational cost, the use of detec-

tors notably reduced training time for all algorithms in

real data scenarios and decreased prediction time for

neighborhood-based algorithms.

In summary, our proposed solution approach

demonstrates promise in improving recommendation

algorithm performance, particularly in datasets with

highly connected networks, especially when employ-

ing neighborhood-based algorithms. Additionally,

even in datasets with fewer disconnected users, the

approach maintains similar performance to the default

method while signiﬁcantly reducing computational

costs. These ﬁndings highlight the approach’s effec-

tiveness in recommendation systems, enhancing per-

formance while optimizing computational resources.

5 CONCLUSIONS

In this paper, we introduced and evaluated a

community-based approach to enhance the effective-

ness and efﬁciency of recommendation algorithms.

Our approach yielded promising results, particularly

in data scenarios with clear network structures, con-

sistently outperforming the default method with lower

RMSE values. The synergy between community

detectors and neighborhood-based algorithms fur-

ther highlighted our approach’s advantages. Even

in datasets lacking distinct clusters, our approach

maintained comparable performance to the default

method while signiﬁcantly reducing computational

costs. These ﬁndings offer valuable insights for rec-

ommendation system researchers and practitioners,

showcasing the potential of community-based meth-

ods in improving accuracy and computational efﬁ-

ciency.

Future development of our approach should in-

volve exploring whether other algorithms or param-

eter combinations, both for community detection and

recommendation, can achieve similar or superior en-

hancements within the same pipeline. Additionally,

further investigation using different datasets, includ-

ing real datasets beyond Movie Lens 100k and syn-

thetic data with varying characteristics, can expand

our understanding of how our approach handles di-

verse community structures.

ACKNOWLEDGEMENTS

Douglas O. Cardoso acknowledges the ﬁnancial sup-

port by the Foundation for Science and Technol-

ogy (Fundac¸

ao para a Ci

encia e a Tecnologia, FCT)

through grant UIDB/05567/2020, and by the Euro-

pean Social Fund and programs Centro 2020 and Por-

tugal 2020 through project CENTRO-04-3559-FSE-

000158.

REFERENCES

Adomavicius, G. and Tuzhilin, A. (2005). Toward the

next generation of recommender systems: A sur-

vey of the state-of-the-art and possible extensions.

IEEE transactions on knowledge and data engineer-

ing, 17(6):734–749. 1

Bhaskaran, S., Hariharan, S., Veeramanickam, M.,

Bharathiraja, N., Pradeepa, K., and Marappan, R.

(2022). Recommendation system using inference-

WEBIST 2023 - 19th International Conference on Web Information Systems and Technologies

502

based graph learning–modeling and analysis. In 2022

2nd International Conference on Innovative Sustain-

able Computational Technologies (CISCT), pages 1–

5. IEEE. 1

Blondel, V. D., Guillaume, J.-L., Lambiotte, R., and Lefeb-

vre, E. (2008). Fast unfolding of communities in large

networks. Journal of statistical mechanics: theory

and experiment, 2008(10):P10008. 3, 5

Bourhim, S., Benhiba, L., and Idrissi, M. J. (2018). Towards

a novel graph-based collaborative ﬁltering approach

for recommendation systems. In Proceedings of the

12th International Conference on Intelligent Systems:

Theories and Applications, pages 1–6. 1, 2

Bourhim, S., Benhiba, L., and Idrissi, M. J. (2019). Investi-

gating algorithmic variations of an rs graph-based col-

laborative ﬁltering approach. In Proceedings of the

ArabWIC 6th Annual International Conference Re-

search Track, pages 1–6. 2

Burke, R. (2000). Knowledge-based recommender sys-

tems. Encyclopedia of library and information sys-

tems, 69(Supplement 32):175–186. 1

Cao, K. Y., Liu, Y., and Zhang, H. X. (2020). Improving the

cold start problem in music recommender systems. In

Journal of Physics: Conference Series, volume 1651,

page 012067. IOP Publishing. 2

Castells, M. and Cardoso, G. (1996). The network society,

volume 469. oxford: blackwell. 1

Choudhary, C., Singh, I., and Kumar, M. (2023). Com-

munity detection algorithms for recommendation

systems: techniques and metrics. Computing,

105(2):417–453. 3

C¸ iftc¸i, O., Tenekeci, S., et al. (2021). Artist recom-

mendation based on association rule mining and

community detection. In Proceedings of the 13th

International Joint Conference on Knowledge Dis-

covery, Knowledge Engineering and Knowledge

Management-KDIR. SCITEPRESS. 2

Dias, P. (2014). From ‘infoxication’to ‘infosaturation’: a

theoretical overview of the cognitive and social effects

of digital immersion. ambitos, 24. 1

Girvan, M. and Newman, M. E. (2002). Community struc-

ture in social and biological networks. Proceedings of

the national academy of sciences, 99(12):7821–7826.

Gorripati, S. K. and Vatsavayi, V. K. (2017). Community-

based collaborative ﬁltering to alleviate the cold-start

and sparsity problems. International Journal of Ap-

plied Engineering Research, 12(15):5022–5030. 1, 2

Guo, G. (2013). Improving the performance of recom-

mender systems by alleviating the data sparsity and

cold start problems. In Twenty-Third International

Joint Conference on Artiﬁcial Intelligence. 1

Hagberg, A., Swart, P., and S Chult, D. (2008). Explor-

ing network structure, dynamics, and function using

networkx. Technical report, Los Alamos National

Lab.(LANL), Los Alamos, NM (United States). 5

Hug, N. (2020). Surprise: A python library for recom-

mender systems. Journal of Open Source Software,

5(52):2174. 5

Keerthi Gorripati, S. and Angadi, A. (2018). Visual based

fashion clothes recommendation with convolutional

neural networks. International Journal of Information

Systems & Management Science, 1(1). 1

Kherad, M. and Bidgoly, A. J. (2022). Recommenda-

tion system using a deep learning and graph analysis

approach. Computational Intelligence, 38(5):1859–

1883. 2

Panteli, M., Piscopo, A., Harland, A., Tutcher, J., and Moss,

F. M. (2019). Recommendation systems for news ar-

ticles at the bbc. In INRA@ RecSys, pages 44–52. 1

Patra, B. K., Launonen, R., Ollikainen, V., and Nandi, S.

(2014). Exploiting bhattacharyya similarity measure

to diminish user cold-start problem in sparse data.

In Discovery Science: 17th International Conference,

DS 2014, Bled, Slovenia, October 8-10, 2014. Pro-

ceedings 17, pages 252–263. Springer. 2

Paul, P. and Aithal, P. (2018). Digital society: Its foundation

and towards an interdisciplinary ﬁeld. In Proceedings

of National Conference on Advances in Information

Technology, Management, Social Sciences and Edu-

cation, pages 1–6. 1

Pazzani, M. J. and Billsus, D. (2007). Content-based rec-

ommendation systems. In The adaptive web, pages

325–341. Springer. 1

Pham, M. C., Cao, Y., Klamma, R., and Jarke, M. (2011). A

clustering approach for collaborative ﬁltering recom-

mendation using social network analysis. J. Univers.

Comput. Sci., 17(4):583–604. 2

Raghavan, U. N., Albert, R., and Kumara, S. (2007).

Near linear time algorithm to detect community struc-

tures in large-scale networks. Physical Review E,

76(3):036106. 3

Rich, E. (1979). User modeling via stereotypes. Cognitive

science, 3(4):329–354. 1

Roser, M., Ritchie, H., and Ortiz-Ospina, E.

(2015). Internet. Our World in Data.

https://ourworldindata.org/internet. 1

Rosvall, M. and Bergstrom, C. T. (2008). Maps of random

walks on complex networks reveal community struc-

ture. Proceedings of the national academy of sciences,

105(4):1118–1123. 3

Sanderson, M. and Croft, W. B. (2012). The history of in-

formation retrieval research. Proceedings of the IEEE,

100(Special Centennial Issue):1444–1451. 1

Schafer, J. B., Frankowski, D., Herlocker, J., and Sen, S.

(2007). Collaborative ﬁltering recommender systems.

In The adaptive web, pages 291–324. Springer. 1

Xie, J., Kelley, S., and Szymanski, B. K. (2013). Overlap-

ping community detection in networks: The state-of-

the-art and comparative study. Acm computing surveys

(csur), 45(4):1–35. 5

Xie, J. and Szymanski, B. K. (2013). Labelrank: A stabi-

lized label propagation algorithm for community de-

tection in networks. In 2013 IEEE 2nd Network Sci-

ence Workshop (NSW), pages 138–143. IEEE. 1

Experimental Analysis of Pipelining Community Detection and Recommender Systems

503