Artist Recommendation based on Association Rule Mining and

Community Detection

Okan C¸ iftc¸i, Samet Tenekeci and Ceren

Ulgent

urk

Department of Computer Engineering,

Izmir Institute of Technology,

Izmir, Turkey

Keywords:

Association Rule Mining, Community Detection, Recommender Systems, Graph Databases.

Abstract:

Recent advances in the web have greatly increased the accessibility of music streaming platforms and the

amount of consumable audio content. This has made automated recommendation systems a necessity for

listeners and streaming platforms alike. Therefore, a wide variety of predictive models have been designed to

identify related artists and music collections. In this paper, we proposed a graph-based approach that utilizes

association rules extracted from Spotify playlists. We constructed several artist networks and identiﬁed related

artist clusters using Louvain and Label Propagation community detection algorithms. We analyzed internal

and external cluster agreements based on different validation criteria. As a result, we achieved up to 99.38%

internal and 90.53% external agreements between our models and Spotify’s related artist lists. These results

show that integrating association rule mining concepts with graph databases can be a novel and effective way

to design an artist recommendation system.

1 INTRODUCTION

With the widespread use of the Internet, music dis-

tribution channels have diversiﬁed. As a result, more

and more music becomes consumable every day. At

this point, just as listeners want to discover new songs

and artists, music streaming platforms seek to make

the right recommendations to increase the time their

users spend on the application. For this purpose,

many recommendation systems have been developed

that make automatic music and artist suggestions us-

ing users’ activities, common behavior patterns, or

other metadata.

These systems adopt state-of-the-art methods of

graph theory, statistics, data mining, machine learn-

ing, and deep learning. They can be classiﬁed by

the type of input features they used. In the litera-

ture, music and artist recommendation systems have

been designed as content-based (based on text and

audio features) (Cano et al., 2005a; Cano et al.,

2005b; Yoshii et al., 2006; Chen et al., 2011), context-

aware (based on playlist and category metadata) (Han

et al., 2010; Hariri et al., 2012; Pichl et al., 2015),

location-aware (based on geospatial data) (Schedl and

Schnitzer, 2014; Kaminskas et al., 2013; Cheng and

Shen, 2016), culture-aware (based on cultural meta-

data) (Baumann and Hummel, 2003; Baumann and

Hummel, 2005; Zangerle et al., 2018), graph-based

(based on topological features) (Cano et al., 2006;

Celma and Herrera, 2008; Yin et al., 2012), or in-

tegrating multiple features (Neumayer and Rauber,

2007; Wang et al., 2012; Schedl, 2013).

A collection of related artists can be modeled as

a social network where each node represents an artist

and each edge represents the strength of the relation-

ship between an artist pair. Graph-based models are

effective in representing such complex networks. Ad-

ditionally, the topological properties of graphs can be

used to detect communities in these networks. In this

context, we proposed a graph-based approach for the

problem of artist recommendation.

Our approach can be included in the class of

context-aware methods since it relies on playlist data.

It utilizes well-known graph theory and data mining

techniques on Spotify playlists to extract association

rules and corresponding artist communities in a graph

database. The workﬂow of our method includes 4

main steps: (1) identify frequent artist sets and asso-

ciation rules in Spotify playlists, (2) construct a graph

database using the support and conﬁdence values cal-

culated in Step 1, (3) discover related artist communi-

ties by running state-of-the-art community detection

algorithms on the constructed artist network, (4) per-

form internal and external cluster validations on dis-

covered communities. Among the similar artist rec-

ommendation systems, our approach is novel as it is

Çiftçi, O., Tenekeci, S. and Ülgentürk, C.

Artist Recommendation based on Association Rule Mining and Community Detection.

DOI: 10.5220/0010678600003064

In Proceedings of the 13th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2021) - Volume 1: KDIR, pages 257-263

ISBN: 978-989-758-533-3; ISSN: 2184-3228

257

the ﬁrst attempt to perform community detection on

graph databases built using the concepts of support

and conﬁdence.

The remainder of this paper is organized as fol-

lows: Section 2 introduces the related work. Section

3 describes materials, methods, and evaluation met-

rics used, along with the running environment and

performance. Section 4 includes experimental results

as well as the discussions and threats to validity. Fi-

nally, Section 5 concludes the paper and presents the

future work.

2 RELATED WORK

Our related work consists of approaches that adopt

graph-based algorithms, with a few exceptions that

are noted.

In (Cano et al., 2006), the authors conducted a

study on different music recommendation systems by

means of complex network analysis. They examined

the common and distinctive topological features of

AllMusicGuide, MSN Entertainment, Amazon, and

Launch Yahoo! Music. Their results showed that

despite some common features, such as small world-

ness, different network characteristics exist, such as

the link degree distribution.

In (Celma and Herrera, 2008), two graph-based

approaches, namely Item- and User-centric, have

been proposed to evaluate the quality of novel rec-

ommendations. The authors tried to detect whether

the network topology has any pathology that hinders

novel recommendations and measure users’ perceived

quality of novel, previously unknown, recommenda-

tions. They also compared the content-based and

social-based recommendation methods.

In (Anglade et al., 2011), the authors designed a

recommendation system based on the music prefer-

ences of users’ social connections. To this end, they

used the social shufﬂe principle in graphs represent-

ing the social interactions between the users. They

developed an application called Starnet to verify their

claims. As a result, they proved the effect of social

communication networks on music preferences.

In (Yin et al., 2012), the authors proposed a graph-

based artist recommendation system that utilizes lis-

tening and trust preference networks (LTPN). They

combine listening and trust information provided by

users and unfold his/her reliable friends with similar

tastes to make better recommendations. Their exper-

imental results demonstrate LTPN can not only pro-

vide better recommendation but also help relieve the

cold start problem caused by new users.

In (Wang et al., 2014), the authors worked on a

graph-based recommendation system for social net-

works. They used rating and tag information and

charted the relationship based on the co-tagging be-

havior of users. They aimed to design a more accu-

rate recommendation system by supporting the ran-

dom walk with restart algorithm on tags. As a result,

they were able to achieve good performance and ac-

curate propositions.

In (Turnbull and Waldner, 2018), the authors tried

to ﬁnd a solution to the task of local music event rec-

ommendation. It is difﬁcult to ﬁnd the relationship

between local music artists since they tend to be ob-

scure long-tail artists with a small digital footprint. To

address this problem, they utilized Latent Semantic

Analysis (LSA). They embedded artists and tags into

a latent feature space and effectively modeled artist

similarity. They also introduced the concept of a Mu-

sic Event Graph that makes it easy and efﬁcient to

recommend events based on user-selected genre tags

and popular artists.

Lastly, in (Yakura et al., 2018), the authors pro-

posed a system to make background music sugges-

tions based on users’ feedback. They especially

worked on the concentration-enhancing background

music used while working. They did not use a graph-

based approach but emphasized the power of the

recommendation system as they proved the focus-

enhancing effect of the suggested songs.

3 MATERIALS AND METHODS

3.1 Dataset & Preprocessing

In 2018, Spotify helped organize the RecSys Chal-

lenge 2018, a data science research challenge fo-

cused on music recommendation. As part of that

challenge, Spotify introduced The Million Playlist

Dataset (Chen et al., 2018) a dataset of 1 million

playlists consisting of over 2 million unique tracks

by nearly 300,000 artists. This represents the largest

public dataset of music playlists in the world. From

this challenge we accessed a sample. It contains

approximately 663,000 tracks in 9999 playlists with

172,000 unique tracks and 36,000 unique artists. It

contains track, artist, album, playlist ids, track name,

album name and artist name. In this task we group

by each artist by playlist id then used each playlist as

a transaction to ﬁnd association rules between artists

therefore we only use playlist id and artist name from

given playlist.

KDIR 2021 - 13th International Conference on Knowledge Discovery and Information Retrieval

258

3.2 Frequent Itemsets & Association

Rule Mining

We utilize the well-known Apriori algorithm

(Agrawal and Srikant, 1994) to obtain frequent

itemsets and association rules for our dataset.

Apriori algorithm is often used for the analysis of

co-occurring items over relational databases. Essen-

tially, it uses Boolean association rules to evaluate the

features and transactions. It starts with identifying

the common individual items in the database and

proceeds by extending them to larger itemsets as long

as the frequency of these itemsets is higher than a

certain threshold (i.e. minimum support criterion).

For example, given two items, X and Y , Apriori

algorithm deﬁnes the support and conﬁdence values

as:

Supp(X,Y ) =

Frequency(X ∪Y )

Con f (X,Y ) =

Supp(X,Y )

Supp(X)

(1)

where N is the total number of transactions in the

database. According to the Apriori algorithm, if a k-

itemset (i.e. an itemset with k elements) provides the

minimum support value, the subsets of this set also

satisfy the minimum support criterion. The frequent

itemsets determined by Apriori can be used to deter-

mine association rules which highlight general trends

in the database.

3.3 Graph Clustering

We perform dimensional reduction on our dataset to

extract the 100 artists with the strongest interactions.

By trial and error, we select a threshold of 0.23 for

support and conﬁdence values and ﬁlter out links with

weights below this threshold. In this way, we both

speed up the experiments and increase the reliability

of the results. For convenience, we call the 100 artists

we selected source artists.

We use both support and conﬁdence values as

edge weights on artist networks. To generate clus-

ters on each graph, we run Louvain community de-

tection algorithm (Zachary, 1977; Lu et al., 2015) and

Label Propagation algoritm (Xing et al., 2014) which

are available on Neo4j

graph database platform. For

each algorithm, we perform two different clustering

using both support and conﬁdence values. For con-

venience, we call these methods LouSupp, LouConf,

LabSupp, and LabConf.

http://neo4j.org

3.3.1 Louvain Algorithm

Louvain algorithm aims to detect communities in

large networks. It maximizes a modularity score

for each community, where the modularity quanti-

ﬁes the quality of an assignment of nodes to commu-

nities. Louvain uses a greedy optimization method

that runs in time O(nlog n), where n is the number of

nodes in the network. In Louvain algorithm, ﬁrst the

small communities found by optimizing modularity

locally on all nodes, and then each small community

is grouped each other.

3.3.2 Label Propagation Algorithm

LPA is a fast algorithm for ﬁnding communities in

a graph. It detects these communities using a graph

structure. Steps of the algorithm are as follows: (i) ev-

ery node initialize with unique community label, (ii)

labels propagate through the network, (iii) in every

iteration on propagation, each node updates its label

to the one that the maximum numbers of its neigh-

bours belongs to. Algorithm reaches convergence

when each node has the majority label of its neigh-

bours. It can stop either convergence or maximum

number of iterations is achieved. At the end of the

propagation process only few labels remain. Nodes

that have the same community label at convergence

are said to belong to the same community.

3.4 Retrieving Ground Truths from

Spotify

As ground truth, we use related artists from Spotify.

We retrieve 20 related artists for each of 100 artists

(i.e. the source artists) in our dataset using Spotipy

API

. Since we are only interested in the relationships

between our source artists, we exclude other related

artists from this collection. After ﬁltering, we remove

clusters that contain a single element. Further de-

tails about Spotify’s related artist collection is given

in Section 4.2.

3.5 Evaluation Metrics

We use the Adjusted Rand Index (ARI) (Hubert and

Arabie, 1985) and the Pairwise Overlap Coefﬁcient

(POC) which is a modiﬁcation of the Overlap Coef-

ﬁcient (OC) (Vijaymeena and Kavitha, 2016), for in-

ternal and external cluster validation. ARI works only

on disjoint clusters, while POC can be applied to both

disjoint and overlapping clusters. Thus, we can utilize

both metrics to calculate pairwise agreement ratios

https://spotipy.readthedocs.io

Artist Recommendation based on Association Rule Mining and Community Detection

259

Table 1: The contingency table.

... Y

sums

... n

sums b

... b

of LouSupp, LouConf, LabSupp, and LabConf, that

all consist of disjoint clusters. This is called inter-

nal validation (see Section 4.1). On the other hand,

we can use only the POC to calculate similarity be-

tween our graph communities and Spotify’s related

artist sets, because of the fact that these sets are over-

lapping. The comparison with Spotify data is called

external validation (see Section 4.2)

We utilize adjusted rand score from scikit-

learn library (Pedregosa et al., 2011) for internal val-

idation of clusters. Given a set S of n elements,

and two clusterings of these elements, namely X =

,... ,X

} and Y = {Y

,... ,Y

}, the overlap

between X and Y can be summarized in a contin-

gency table (Table 1) where each entry n

i j

denotes

the number of elements in common between X

and

: n

i j

= |X

∩Y

|. Using the contingency table, ARI

can be formulated as:

∑

i j



i j



−

∑





∑









∑





∑





−

∑





∑









(2)

where n

i j

, a

, and b

are values from the contingency

table. Typically, ARI score is between −1.0 and 1.0.

It is close to 0.0 for randomly labeled clusters. In case

of perfect match (i.e. identical clustering) ARI is 1.0.

A negative ARI represents that the agreement is less

than what is expected from a random result. Further

details about ARI is available in (Hubert and Arabie,

1985). The internal validation results are presented in

Section 4.1.

We utilize POC, which is a modiﬁcation

of OC, for both internal and external valida-

tion of clusters. Given two sets of overlap-

ping clusters, X = {(a,b,c), (c,d,e)} and Y =

{(a,b,c),(b,d),(b,e), (c,a)}, the traditional OC is

deﬁned as:

OC(X,Y ) =

|X ∩Y |

min(|X|,|Y |)

(3)

and equal to 0.5, since only the (a,b, c) cluster is

matching. To calculate POC, on the other hand, we

ﬁrst generate pairs of elements (i.e. tuples) using the

elements in the same clusters. Then, we ﬁlter out

the duplicated and symmetric pairs in each set and

Table 2: Execution Times (ms) of Clustering Methods.

LouSupp LouConf LabSupp LabConf

Time 94 200 12 24

obtain reduced sets of pairs. Finally, we calculate

the OC between these sets. For X and Y , these sets

are X

= {(a, b),(a, c),(b, c),(c, d),(c,e),(d,e)} and

= {(a,b),(a,c),(b,c),(b,d), (b,e)}, respectively.

Hence, POC(X,Y ) = OC(X

) = 0.6.

In case of a set of disjoint clusters, X =

,... ,X

}, the total number of pairs is:

| =

∑

i=1

|(|X

| − 1)

(4)

In the case of overlapping clusters (as in the exam-

ple above), duplicated and symmetrical tuples are ex-

cluded from this set. After we form the sets of artist

pairs for both clusterings, we calculate the POC. In

this way, we determine to what extent the two given

clusterings agree in associating the given artist pairs.

The internal and external validation results are pre-

sented in Section 4.1 and Section 4.2, respectively.

3.6 Running Environment &

Performance

In this work, we mostly use Python programming

language and Jupyter Notebooks

for development.

We utilize Apyori

library to create association rules,

Cypher query language and neo4j platform to store

and manage our graph database, and scikit-learn

li-

brary to validate our clusters. We use netgraph

for

visualization. We run the experiments on a MacOS

Catalina device with i7 9700 3.6 GHz Intel processor,

16GB RAM, and 512GB SSD. Table 2 shows the ex-

ecution time of each clustering method. Two conclu-

sions can be drawn from our performance analyses:

(1) the clustering algorithms run about twice as fast

when support is used instead of conﬁdence, (2) the

Label Propagation algorithm is about 8 times faster

than the Louvain algorithm.

4 EXPERIMENTAL RESULTS

In this section, we present the experimental results

in three parts. First, we provide some important

statistics about graph-based communities generated

by Louvain and Label Propagation algorithms and

https://jupyter.org

https://github.com/ymoch/apyori

https://scikit-learn.org

https://github.com/paulbrodersen/netgraph

KDIR 2021 - 13th International Conference on Knowledge Discovery and Information Retrieval

260

Figure 1: Data distribution of four clustering methods.

give internal validation results based on ARI and POC

metrics. Then, we provide data statistics about Spo-

tify’s related artist collection and present external val-

idation results based on POC. In the last part, we dis-

cuss the results and list the threats to validity.

4.1 Graph-based Communities

(Internal Validation)

Among our 4 community detection methods,

LouConf detected 5 communities while each of the

others (LouSupp, LabSupp, and LabConf) detected 4.

The smallest and largest communities were generated

by the Label Propagation algorithm, with 7 and 61

artists, respectively. On the other hand, Louvain’s

communities ranged in size from 14 to 30. Note

that the communities generated by each method are

disjoint among themselves. Figure 1 presents the

data distribution obtained by each clustering method.

Considering the community sizes, the Louvain algo-

rithm provides a more balanced distribution. Figure 2

illustrates the communities generated by the LouConf

algorithm. Considering the edge densities, the artist

network has a scale-free degree distribution. In other

words, it has a small number of highly connected

hub nodes (these are famous American rappers like

Drake, Kanye West, Kendrick Lamar, and Future)

and a large number of weakly connected nodes.

Table 3: ARIs for Internal Validation.

LouSupp LouConf LabSupp LabConf

LouSupp 1 0.7313 0.4485 0.4355

LouConf 0.7313 1 0.3228 0.3099

LabSupp 0.4485 0.3228 1 0.9872

LabConf 0.4355 0.3099 0.9872 1

Table 4: POCs for Internal Validation.

LouSupp LouConf LabSupp LabConf

LouSupp 1 0.8943 0.8239 0.8137

LouConf 0.8943 1 0.7789 0.7661

LabSupp 0.8239 0.7789 1 0.9938

LabConf 0.8137 0.7661 0.9938 1

Table 5: POCs for External Validation.

# of pairs # of overlaps POC

LouSupp 1283 379 0.5279

LouConf 1022 320 0.4457

LabSupp 2095 647 0.9011

LabConf 2100 650 0.9053

* Number of artist pairs in Spotify clusters is 718.

4.2 Related Artists from Spotify

(External Validation)

For external cluster validation, we compare Spotify’s

related artist collection with clusters generated by

Louvain and Label Propagation algorithms. We se-

lect 20 related artists for each source artist and ob-

tain a collection of 2100 artists (including the source

artists), 730 of which are distinct. We can also think

of this collection as 100 clusters, each consisting of

21 artists. After ﬁltering out the related artists that are

not in the source artists, we remove clusters that con-

tain a single element. As a result, we have 87 clusters

that range in size from 2 to 17.

To calculate POCs, we generate sets of artist pairs

for 100 artists featured in 87 overlapping clusters. Af-

ter we form all artist pairs, which are 1649 pairs ac-

cording to Equation 4, we ﬁlter out the duplicated and

symmetrical tuples and obtain a ﬁnal set of size 718.

4.3 Discussions & Threats to Validity

In internal validation, both ARI and POC analyses

show that agreement between support-weighted and

conﬁdence-weighted clusters is lower in the Louvain

algorithm (ARI = 73.13%, POC = 89.43%) compared

to the Label Propagation algorithm (ARI = 98.72%,

POC = 99.38%). On the other hand, the highest cross-

algorithm agreement is achieved using support values

(ARI = 44.85%, POC = 82.39%). As an interest-

ing side note, LouConf shows higher agreement with

LabSupp than LabConf, although the results are very

close for both metrics.

In external validation, regardless of the edge

weighting method used, the Label Propagation

algorithm (90.11% and 90.53% for support and con-

ﬁdence, respectively) signiﬁcantly outperforms the

Louvain algorithm (52.79% and 44.57% for support

Artist Recommendation based on Association Rule Mining and Community Detection

261

Figure 2: Communities generated by LouConf algorithm.

and conﬁdence, respectively), and it shows a higher

agreement with Spotify’s related artist clusters.

Both ARI and POC analyses show that the Label

Propagation algorithm on a graph with conﬁdence-

weighted edges has the highest cluster agreement,

both internally and externally. On the other hand,

our performance analyses show that Label Propaga-

tion executes signiﬁcantly faster than Louvain (see

Section 3.6). The statistical superiority of the Label

Propagation algorithm in clustering performance can

be attributed to the fact that it contains a large cluster

with 61 artists.

The unbalanced cluster distribution in the Label

Propagation algorithm could be a threat to the va-

lidity of our results. Similarly, the date mismatch

between the dataset we work on (September, 2018)

and the related artist collection that we retrieved from

Spotify (June, 2021) could be a threat to the valid-

ity. Methodologically, limiting the dataset to a subset

of 100 artists in this work would be a threat when

scaling our model up to larger and more complex

networks. Lastly, using only the contextual features

(playlist metadata) to generate association rules could

be another threat to validity.

5 CONCLUSIONS & FUTURE

WORK

Recent advances in internet technology have greatly

increased the accessibility of music streaming plat-

forms and the amount of consumable audio content.

This has made automated recommendation systems a

necessity for both listeners and streaming platforms.

As a result, various models have emerged that use dif-

ferent input features and computational methods to

detect related artists and music collections.

In this work, we proposed a graph-based model

that relies on contextual features (i.e. playlist data)

and association rules. We used support and conﬁ-

dence metrics as edge weights in artist network. We

utilized Louvain and Label Propagation community

detection algorithms to identify clusters of related

artists. We performed internal and external valida-

tions of clusters using the Adjusted Rand Index (ARI)

and Pairwise Overlap Coefﬁcient (POC).

We achieved clustering agreements up to 98.72%

between support and conﬁdence metrics, 44.85% be-

tween Louvain and Label Propagation algorithms,

and 90.53% between our model and Spotify’s related

artists. These results show that integrating association

rule mining concepts with graph databases can be a

novel and effective way to design a recommendation

system.

In future work, association rules and links in the

artist network can be semantically enriched by inte-

grating contextual data with other input features like

textual, categorical, cultural, or geospatial metadata.

Additionally, this model can be extended using other

datasets and community detection algorithms.

ACKNOWLEDGEMENTS

We would like to thank Dr. Damla O

guz from

Izmir

Institute of Technology for their comments and sug-

gestions on this study.

REFERENCES

Agrawal, R. and Srikant, R. (1994). Fast algorithms for

mining association rules. In Proceedings of 20th in-

ternational conference on very large data bases, pages

487–499, Morgan Kaufmann Publishers Inc. VLDB.

Anglade, A., Celma,

O., Fields, B., Lamere, P., and McFee,

B. (2011). Womrad: 2nd workshop on music rec-

ommendation and discovery. In Proceedings of the

ﬁfth ACM conference on Recommender systems, pages

381–382.

Baumann, S. and Hummel, O. (2003). Using cultural meta-

data for artist recommendations. In Proceedings Third

International Conference on WEB Delivering of Mu-

sic, pages 138–141. IEEE.

Baumann, S. and Hummel, O. (2005). Enhancing music

recommendation algorithms using cultural metadata.

Journal of New Music Research, 34(2):161–172.

Cano, P., Celma, O., Koppenberger, M., and Buldu, J. M.

(2006). Topology of music recommendation net-

works. Chaos: An interdisciplinary journal of non-

linear science, 16(1):013107.

KDIR 2021 - 13th International Conference on Knowledge Discovery and Information Retrieval

262

Cano, P., Koppenberger, M., and Wack, N. (2005a).

Content-based music audio recommendation. In Pro-

ceedings of the 13th annual ACM international con-

ference on Multimedia, pages 211–212.

Cano, P., Koppenberger, M., and Wack, N. (2005b). An

industrial-strength content-based music recommenda-

tion system. In Proceedings of the 28th annual in-

ternational ACM SIGIR conference on Research and

development in information retrieval, pages 673–673.

Celma,

O. and Herrera, P. (2008). A new approach to

evaluating novel recommendations. In Proceedings of

the 2008 ACM conference on Recommender systems,

pages 179–186.

Chen, C. W., Lamere, P., Schedl, M., and Zamani, H.

(2018). Recsys challenge 2018: Automatic music

playlist continuation. In Proceedings of the 12th ACM

Conference on Recommender Systems, pages 527–

528.

Chen, Z. S., Jang, J. S. R., and Lee, C. H. (2011). A ker-

nel framework for content-based artist recommenda-

tion system in music. IEEE Transactions on Multime-

dia, 13(6):1371–1380.

Cheng, Z. and Shen, J. (2016). On effective location-aware

music recommendation. ACM Transactions on Infor-

mation Systems (TOIS), 34(2):1–32.

Han, B. J., Rho, S., Jun, S., and Hwang, E. (2010). Music

emotion classiﬁcation and context-based music rec-

ommendation. Multimedia Tools and Applications,

47(3):433–460.

Hariri, N., Mobasher, B., and Burke, R. (2012). Context-

aware music recommendation based on latenttopic se-

quential patterns. In Proceedings of the sixth ACM

conference on Recommender systems, pages 131–138.

Hubert, L. and Arabie, P. (1985). Comparing partitions.

Journal of classiﬁcation, 2(1):193–218.

Kaminskas, M., Ricci, F., and Schedl, M. (2013). Location-

aware music recommendation using auto-tagging and

hybrid matching. In Proceedings of the 7th ACM Con-

ference on Recommender Systems, pages 17–24.

Lu, H., Halappanavar, M., and Kalyanaraman, A. (2015).

Parallel heuristics for scalable community detection.

Parallel Computing, 47:19–37.

Neumayer, R. and Rauber, A. (2007). Integration of text

and audio features for genre classiﬁcation in music in-

formation retrieval. In European Conference on Infor-

mation Retrieval, pages 724–727, Berlin, Heidelberg.

Springer.

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,

Thirion, B., Grisel, O., and Duchesnay, E. (2011).

Scikit-learn: Machine learning in python. The Journal

of machine Learning research, 12:2825–2830.

Pichl, M., Zangerle, E., and Specht, G. (2015). To-

wards a context-aware music recommendation ap-

proach: What is hidden in the playlist name? In 2015

IEEE International Conference on Data Mining Work-

shop (ICDMW), pages 1360–1365.

Schedl, M. (2013). Ameliorating music recommendation:

Integrating music content, music context, and user

context for improved music retrieval and recommen-

dation. In Proceedings of international conference on

advances in mobile computing & multimedia, pages

3–9.

Schedl, M. and Schnitzer, D. (2014). Location-aware music

artist recommendation. In International conference

on multimedia modeling, pages 205–213. Springer,

Cham.

Turnbull, D. and Waldner, L. (2018). Local music event

recommendation with long tail artists. arXiv preprint

arXiv:1809.02277.

Vijaymeena, M. K. and Kavitha, K. (2016). A survey on

similarity measures in text mining. Machine Learning

and Applications: An International Journal, 3(2):19–

28.

Wang, H., Li, G., and Feng, J. (2014). Group-based

personalized location recommendation on social net-

works. In Asia-Paciﬁc Web Conference, pages 68–80.

Springer.

Wang, X., Rosenblum, D., and Wang, Y. (2012). Context-

aware mobile music recommendation for daily activ-

ities. In Proceedings of the 20th ACM international

conference on Multimedia, pages 99–108.

Xing, Y., Meng, F., Zhou, Y., Zhu, M., Shi, M., and Sun,

G. (2014). A node inﬂuence based label propagation

algorithm for community detection in networks. The

Scientiﬁc World Journal.

Yakura, H., Nakano, T., and Goto, M. (2018). Focusmusi-

crecommender: a system for recommending music to

listen to while working. In 23rd International Confer-

ence on Intelligent User Interfaces, pages 7–17.

Yin, C. X., Peng, Q. K., and Chu, T. (2012). Personal artist

recommendation via a listening and trust preference

network. Physica A: Statistical Mechanics and its Ap-

plications, 391(5):1991–1999.

Yoshii, K., Goto, M., Komatani, K., Ogata, T., and

Okuno, H. G. (2006). Hybrid collaborative and

content-based music recommendation using proba-

bilistic model with latent user preferences. ISMIR,

6:296–301.

Zachary, W. W. (1977). An information ﬂow model for con-

ﬂict and ﬁssion in small groups. Journal of anthropo-

logical research, 33(4):452–473.

Zangerle, E., Pichl, M., and Schedl, M. (2018). Culture-

aware music recommendation. In Proceedings of the

26th Conference on User Modeling, Adaptation and

Personalization, pages 357–358.

APPENDIX

Relevant datasets and source codes are available at

https://github.com/okanvk/ArtistRecommendation.

Artist Recommendation based on Association Rule Mining and Community Detection

263