2015 IEEE International Conference on Systems, Man, and Cybernetics

Towards An Intuitionistic Fuzzy Agglomerative

Hierarchical Clustering Algorithm for Music
Recommendation in Folksonomy
Chun Guan Kevin Kam Fung Yuen Frans Coenen
Department of Computer Science Department of Computer Science Department of Computer Science
University of Liverpool and Software Engineering University of Liverpool
Liverpool, UK Xi’an Jiaotong-Liverpool University Liverpool, UK
chunguan@liverpool.ac.uk Suzhou, China coenen@liverpool.ac.uk
chun_guan@163.com kevinkf.yuen@gmail.com
kevin.yuen@xjtlu.edu.cn

Abstract—Folksonomy, a system for social tagging or Fuzzy set theory, used to represent the uncertainty membership
collaborative tagging, is popular in Semantic Web research. of items to groups, was established some fifty years ago [12,
Folksonomy is applied to items, such as music pieces, which their 13]. By adding the non-membership concept the idea of
personalized tags can be annotated by users. Recommendation
Intuitionistic Fuzzy Sets (IFS) was developed as the extension
systems can use these tags to produce meaningful information.
Clustering methods, such as the Agglomerative Hierarchical
of fuzzy set. IFS has been applied in various areas such as
Clustering (AHC) method, can be applied in the context of decision making, machine learning and pattern recognition.
recommendation system. This paper proposes the Intuitionistic This paper proposes to combine IFS and AHC to form
Fuzzy Agglomerative Hierarchical Clustering (IFAHC) Intuitionistic Fuzzy Agglomerative Hierarchical Clustering
algorithm for recommendation using social tagging. The (IFAHC) to cluster items according to their IFS values so as to
Intuitionistic Fuzzy Set (IFS) concept is used to represent tag give meaningful patterns. In this paper the proposed IFAHC is
values which are vague and uncertain. IFAHC can cluster items used to cluster music pieces according to a folksonomy tagging
represented by using IFS into different groups. The application
process. In IFAHC, IFS is applied to represent items with both
of IFAHC to music recommendation is used to demonstrate the
usability of the proposed method.
their membership degrees and their non-membership degrees.
The rest of this paper is organized as follows. Section II
Keywords-Agglomerative Hierarchical Clustering; Intuitionistic describes the proposed IFAHC algorithm. Section III presents a
Fuzzy Set; Folksonomy; Music Recommendation demonstration of applying IFAHC to music recommendation.
I. INTRODUCTION Section IV gives a summary of this research and further
research directions.
Clustering analysis [1, 2] is an important technique in data
mining. Hierarchical clustering [1-6] is a classical and popular
clustering algorithm since it was proposed in 1963 [1].
II. INTUITIONISTIC FUZZY AGGLOMERATIVE
Agglomerative Hierarchical Clustering (AHC) algorithms have
been progressively applied in many areas, as described in [3-6]. HIERARCHICAL CLUSTERING
AHC is the method to build a bottom-to-top hierarchical A. Object representation in IFS
decomposition of the data set on the basis of dissimilarities
The definition of IFS is given as follows with respect to [11,
between objects. The clustering result of an AHC algorithm is
14]. Let E be a fixed set. An Intuitionistic Fuzzy Set A in X is
typically illustrated using a dendrogram offering easy
expressed as below.
interpretation by a decision maker.
The system for social tagging and collaborative tagging was  A  { x,  A ( x), A ( x) x  E }   
firstly described using the term “folksonomy” by Vander [7].
Folksonomy allows users to annotate their favorite resources
and items with personalized tags [6]. Clustering algorithms where  A ( x ) is the membership degree of x , i.e.
have been applied to organize the large collections of data  A : X [0,1] , A ( x) is the non-membership degree of x , i.e.
using folksonomy [6, 8, 9]. Resources and items can be divided
 A : X [0,1] .  A ( x ) and  A ( x) satisfy the condition below.
into clusters according to their tags. The clustering results can
be used in a recommendation system. Since the tags are
personalized, defined by users with freely chosen vocabularies,  0  A ( x)  A ( x) 1 x X   
the tag values tend to contain fuzziness [10].
To deal with this fuzziness, the Intuitionistic Fuzzy Set (IFS) The intuitionistic index  A ( x ) represents the hesitancy
concept described in [11, 14] is used to represent the tag values. degree of x to A as below.

978-1-4799-8697-2/15 $31.00 © 2015 IEEE 2039

DOI 10.1109/SMC.2015.356
  A ( x)  1   A ( x)  A ( x)   
Euclidean distance for IFS, as proposed by [14], was used to
calculate the dissimilarities:
Objects in a folksonomy are labeled by user defined tags,
1 n    A ( xi )   B ( xi )  A ( xi )  B ( xi ) 
2 2
and the values of some tags may be fuzzy. This paper proposes
an approach to representing object tag values using IFS. d ( A, B)  
2n i 1   ( x )   ( x ) 2
  

Assume that an object x has been tagged by M users using a  A i B i 
collection of tags, T . An expert will pick up a tag t from
collection T as an attribute of x for clustering. Three kinds of where d ( A, B) is the normalized Euclidean distance between
relationships between object x and tag t can exist. cluster A and B .
 The relationship that a user tagged x by t , which is
similar to the membership relationship, could be
3) Combine the two closest clusters into a bigger cluster.
represented by  A ( x )
.
 The relationship that a user did not tag x by t , which 4) Compute dissimilarities between the new cluster and
is similar to the non-membership relationship, could be the other clusters whilst the remaining dissimilarities remain
represented by  A ( x) unchanged.
. Several types of measurement are suitable for measuring the
 The relationship that a user tagged x by t ' ( t '  t , but distance between clusters. As a widely used measure, the mean
distance, is used in IFAHC:
t is similar to t ' , and t ' T ), which is similar to the
hesitancy relationship, could be represented by  A ( x ) .
 d mean (C j , Ck )  d (m j , mk )   
The similarity can be defined by the overlap of
keywords.
where m j is the mean value of cluster C j , mk is the mean
Let M  , M and M  be the numbers of users with respect to
the above three relationships respectively, and value of cluster C k . d (m j , mk ) is the normalized Euclidean
M  M M   M . The equations below are defined for distance computed using the mean values.
transforming the tagged objects into IFSs. 5) Repeat Steps 3 and 4 until all objects are in one cluster
or user defined termination condition is satisfied.
M
 t ( x)    
M III. MUSIC RECOMMENDATION: A HYPOTHETICAL
APPLICATION
M
 t ( x)     Suppose that 10 heavy metal music pieces have been
M randomly chosen from a folksonomy. Four types of heavy
metal music genres have been chosen to be the attributes of the
M sample datasets in this study. The four genres are Folk Metal,
 t ( x)    
M Symphonic Power Metal, Melodic Death Metal and Traditional
Heavy Metal. The tags of each music piece are represented
Finally, the objects can be transformed to a collection of using IFSs as presented in Table I. The calculation steps, using
IFSs of the above form. IFAHC, are as follows.
A. Object representation in IFS
B. Objects clustering TABLE I. A DATA SET OF 10 MUSIC PIECES REPRESENTED BY IFSS
Agglomerative hierarchical clustering [3] is a bottom-up
Id Folk Symphonic Melodic Traditional
strategy. [15] briefly described the hierarchical clustering Power Death Heavy
methods. The method starts by initializing each object as an u v π u v π u v π u v π
atomic cluster and then merges them into larger clusters, until 1 0.5 0.3 0.2 0.8 0.1 0.1 0.0 1.0 0.0 0.2 0.6 0.2
all objects are in a single cluster or termination condition is 2 1.0 0.0 0.0 0.6 0.2 0.2 0.1 0.7 0.2 0.5 0.2 0.3
satisfied [2]. The proposed IFAHC operates in a somewhat 3 0.6 0.1 0.3 0.8 0.0 0.2 0.5 0.2 0.3 0.0 1.0 0.0
4 0.7 0.2 0.1 0.2 0.8 0.0 0.6 0.1 0.3 0.1 0.5 0.4
different manner to classical AHC. The operation is as follows. 0.2 0.8 0.0 0.6 0.2 0.2 0.0 0.9 0.1 0.9 0.0 0.1
5
6 0.9 0.1 0.0 0.7 0.3 0.0 0.2 0.6 0.2 0.5 0.5 0.0
7 1.0 0.0 0.0 0.0 0.9 0.1 0.7 0.3 0.0 0.8 0.1 0.1
8 0.5 0.4 0.1 0.2 0.6 0.2 0.1 0.8 0.1 0.0 0.9 0.1
1) Initialize each object as an individual cluster.
9 0.3 0.5 0.2 0.8 0.1 0.1 0.6 0.1 0.3 0.7 0.3 0.0
2) Determine dissimilarities between clusters. 10 0.1 0.8 0.1 0.1 0.6 0.3 1.0 0.0 0.0 0.2 0.6 0.2
Euclidean distance has been used to calculate the
dissimilarities in classical AHC. In IFAHC, normalized

2040
 Assume that the music piece ID 1 was tagged by 10 users, 5 After 9 iterations, all music pieces are in a big cluster. The
users among them tagged ID 1 as “Folk Metal”, 3 users did not result can be illustrated in the form of a dengrogram, as shown
tag ID 1 as “Folk Metal”, and 2 users tagged ID 1 as the other in Fig. 1.
tags including “Folk”, but not “Folk Metal”.
The tags of each music piece can be represented by IFSs.
The membership degree of ID 1 for the “Folk Metal” attribute
is 0.5, the non-membership degree is 0.3, and the hesitancy
degree is 0.2. All the IFSs are computed by Eqs. 4-6 and shown
in Table I.
B. Objects clustering
The R language was used to implement the proposed
IFAHC algorithm. Firstly, the dataset of 10 music pieces was
initialized as ten individual clusters: {1}, {2}, {3}, {4}, {5}, Fig. 1. Dengrogram of sample music dataset
{6}, {7}, {8}, {9}, and {10}. The dissimilarities between each produced by IFAHCThe clustering results and dengrogram can
cluster are calculated by Eq. 7. The dissimilarity matrix is be used to recommend music pieces to users. For example, the
shown in Table II. clustering results at Iteration 6 suggest that the data set should
be separated into four clusters, {1}, {2, 6, 10, 3, 7}, {4, 8}, and
TABLE II. DISSIMILARITIES MATRIX OF THE MUSIC PIECE DATA SET {5, 9}. Assume a user listened to music piece 3, the system will
recommend music pieces in order of 7, 2, 6, 10 to the user from
{1} {2} {3} {4} {5} {6} {7} {8} {9} {10} the bottom-up direction of the dengrogram.
{1} 0 0.163 0.205 0.264 0.205 0.147 0.317 0.158 0.236 0.317
{2} 0 0.237 0.222 0.223 0.094 0.218 0.219 0.225 0.331
{3} 0 0.215 0.332 0.185 0.319 0.210 0.203 0.263
{4} 0 0.318 0.201 0.186 0.197 0.226 0.192
{5} 0 0.225 0.305 0.266 0.211 0.316 IV. CONCLUSIONS
{6} 0 0.215 0.186 0.188 0.299
This paper proposes the Intuitionistic Fuzzy Agglomerative
{7} 0 0.281 0.260 0.275
{8} 0 0.270 0.250 Hierarchical Clustering (IFAHC) algorithm which was
{9} 0 0.218 developed on the basis of the Intuitionistic Fuzzy Set (IFS)
{10} 0 concept and Agglomerative Hierarchical Clustering (AHC).
The tag objects are represented by IFS and divided into clusters
using IFAHC to produce a dengrogram which can then be used
According to Table II, the closest two clusters are {2} and for recommendation purposes. A music recommendation
{6}. Therefore the two clusters are combined into a bigger application of social tagging is used to demonstrate the
cluster {2, 6}. The mean value of cluster {2, 6} is computed by usability and validity of the proposed approach.
Eq. 8 and shown in Table III.

TABLE III. IFSS OF CLUSTER {2, 6}

Folk Symphonic Melodic Traditional REFERENCES
Power Death Heavy
u v π u v π u v π u v π
0.95 0.05 0.00 0.65 0.25 0.10 0.15 0.65 0.2 0.50 0.35 0.15 [1] A.K. Jain, M.N. Murty and P J. Flynn, “Data clustering: a review,”
ACM computing surveys (CSUR), vol. 31, 1999, pp. 264-323.
[2] J. Han and M.Kamber, Data Mining: Concepts and Techniques,
Southeast Asia ed., Morgan kaufmann, 2006.
The output of the first iteration combination is: {1}, {2, 6},
[3] J.H. Ward Jr, “Hierarchical grouping to optimize an objective function,”
{3}, {4}, {5}, {7}, {8}, {9}, {10}. The next eight outputs for Journal of the American statistical association, 1963, vol. 58, pp. 236-
each loop are shown as follows. 244.
[4] G. Karypis, EH. Han and V. Kumar, “Chameleon: Hierarchical
 Loop 2:{1}, {2, 6}, {3, 7}, {4}, {5}, {8}, {9}, {10} clustering using dynamic modeling,” Computer, 1999, vol. 32, pp. 68-75.
 Loop 3:{1}, {2, 6}, {3, 7}, {4, 8}, {5}, {9}, {10} [5] Y. Zhao and G. Karypis, “Evaluation of hierarchical clustering
algorithms for document datasets,” Proceedings of the eleventh
 Loop 4:{1}, {2, 6}, {3, 7}, {4, 8}, {5, 9}, {10} international conference on Information and knowledge managemen,
ACM, 2002, pp.515-524.
 Loop 5:{1}, {2, 6, 10}, {3, 7}, {4, 8}, {5, 9} [6] A. Shepitsen, J. Gemmell, B. Mobasher and R.Burke, “Personalized
recommendation in social tagging systems using hierarchical clustering,”
 Loop 6:{1}, {2, 6, 10, 3, 7}, {4, 8}, {5, 9} Proceedings of the 2008 ACM conference on Recommender systems,
ACM, 2008, pp: 259-266.
 Loop 7:{1}, {2, 6, 10, 3, 7, 4, 8}, {5, 9} [7] TV. Wal, Folksonomy coinage and definition ,
http://vanderal.net/folksonomy.html, February 2007.
 Loop 8:{2, 6, 10, 3, 7, 4, 8, 1}, {5, 9}
[8] S. Niwa and S. Honiden, “Web page recommender system based on
 Loop 9:{2, 6, 10, 3, 7, 4, 8, 1, 5, 9} folksonomy mining for ITNG'06 submissions,” Information Technology:
New Generations, 2006. ITNG 2006. Third International Conference on.
IEEE, 2006, pp. 388-393.

2041
[9] L. Specia and E. Motta, “Integrating folksonomies with the semantic [12] L.A. Zadeh, “Fuzzy sets,” Inform. and Control, vol. 8, 1965, pp. 338-
web. The semantic web: research and applications,” Springer Berlin 353.
Heidelberg, 2007, pp. 624-639. [13] G. Klir and B.Yuan, Fuzzy sets and fuzzy logic. New Jersey: Prentice
[10] E. Portmann, “A fuzzy grassroots ontology for improving social Hall, 1995.
semantic web search,” Proceedings of 6th international summer school [14] E. Szmidt and J. Kacprzyk. “Distances between intuitionistic fuzzy sets,”
on aggregation operators, Benevento, 2011. Fuzzy sets and systems, vol. 114(3), 2000, pp. 505-518.
[11] K. Atanassov, “Intuitionistic fuzzy sets,” Fuzzy Sets and Systems, vol. [15] F. Murtagh, “A survey of recent advances in hierarchical clustering
20 , 1986), pp. 87-96. algorithms,” The Computer Journal, 1983, vol. 26, pp. 354-359.

2042