CLUSTERING FACE DATA OF SELECTED FEMALE AVATAR USING

K-MEANS

Citra Dewi Megawati1st, Eko Mulyanto Yuniarno2nd, Supeno Mardi Susiki Nugroho3rd
1
Institut Teknologi Sepuluh November, [email protected], Surabaya Indonesia
2
Institut Teknologi Sepuluh November,[email protected], Surabaya Indonesia
3
Institut Teknologi Sepuluh November, [email protected], Surabaya Indonesia
ABSTRACT
This game requires an avatar to make it look attractive. An avatar is the first remember
from a game. However, illustrators find it difficult to determine the avatar desired by game
lovers. Avatars can also represent a person's appearance or personality in a game or cyberspace.
Avatars can be humanized animals or humanized plants or humans who are female or male. But
in this study we chose a female genre human avatar because it was considered to be more
complete in terms of facial features, because this study was sulking on the clustering of
frequently selected female face avatars. In this study face data features avatars such as face
shape, eyebrow shape, eyes, nose, lips and ears.
In this research clustering data on female avatar faces using the K-means method, K
Means is an unsupervised data clustering algorithm that uses class partitioning methods that
have similarities. In this research the data is partitioned into 3 different classes based on the
similarity of shapes. The results of this study indicate that 39% choosing triangle face shape, circle
left eyebrow, up right eyebrow, slanted eyes with a sharp nose, wide lips and long lobe ears. Then
34% choose avatars with oval faces, triangle left eyebrows, short right eyebrows with
asymmetrical nosed, wide eyes, thick upper lips and ears with thick lobes, and the last 27%
choose avatars with square faces with thin eyebrows, round eyes with rounded noses, lips thick
and ears with small lobes. So with this it is proven that the k-means method is very effective
and fast in classifying the faces of women who are often chosen.
Keywords: Clustering; characters; female character; faces shape; K-means
1. INTRODUCTION
This game is an interactive activity, where one or more players follow rules that limit
the behavior of players, but this application is liked by consumers, besides playing they also
unconsciously learn and enter the game's story. The game has several components, one of the
most important components is the character or avatar. The game will look interesting if there is
a unique avatar.
Avatars are characters that are directly related and exist in the game world. Avatar
greatly influences the player's preference for playing games. Avatar design is one form of
illustration that comes with the concept of "human" male and female genre, "humanized
animal" or "humanized plant" with all its attributes (character, physical, professional, residential
and even destiny). In the design of the avatar, one of the important things that determine is the
face shape of the avatar. However, illustrators find it difficult to determine the face of a human
avatar as desired by the player. In this study we chose a human-shaped avatar because it was
considered more complete in terms of facial features, because this study was sulking on the
grouping of face data of female avatars that were often chosen. In this study face data in the
form of features on the avatar's face such as face shape, shape of eyebrows, eyes, nose, lips and
ears.
Users take care when creating unique avatars, even though face options are ready to
use. (Cheng, Farnham, & Stone, 2002; Taylor, 2002), The problem has been found in the
adjustment process that requires a lot of time (Cheng et al., 2002). Most of the research on
avatars focuses on how visual appearance and appearance of avatar behavior can influence
viewers' perceptions. (e.g., Garau et al., 2003; Nowak & Rauh, 2005; Nowak & Rauh, 2008).
The appearance of the avatar can also influence the perception and behavior of the owner. for
example, users who consider their avatars are more similar to their own appearance (Vasalou,
Joinson, & Pitt, 2007), proven from the quality displayed from the visual avatar can be reflected
in the behavior of the owner, because the avatar is a representation of someone who looks at the
interface. (Rauh, Polonsky, & Buck, 2004).
Research on the classification of facial shapes using 3 methods (N K Bansode and P
K Sinha and Pomthep Sarakon and Theekapun Charoenpong). Research on the classification of
facial forms using SVM, (S. C. Zhang, B. Fang, Y. Z. Liang, J. Wen, and L. Wu), research on
grouping facial data. But they only discuss for face shape, not discussing in full about other face
features such as eyes, nose, eyebrows etc.
In this research clustering face data of female avatars using the K-means method,
Kmeans is an unsupervised data clustering algorithm that uses class partitioning methods that
have similarities. The purpose of this study was to find the faces of female avatars that are often
chosen by game consumers. The game consumers here are undergraduate students from various
majors.
2. METHODS
The research process was carried out according to the K-means clustering method. The
proposed method is the k-means algorithm, which is a method of grouping algorithms that goes
into an unsupervised algorithm. K-means uses a method of partitioning data based on its
similarity in shape. Figure 1 is a block diagram of the steps to research face data.

Figure.1. Research diagram block

3.1 Data on face features
Data on facial features is the design of facial features consisting of faces, eyebrows, eyes,
nose and ears. The design of facial features uses a manga design style, because the design style
of the manga can clearly show the difference between the features of his face. The design of
face features refers to " avatar factory games", "avatar maker", and "avatar generator".
Processing the design using a combination of Corel and Photoshop applications.
The design of the first face feature is the shape of the face. The face shape consists of 6
based on paper classification face shape based on similarty, correlation, and fractal dimensions
by bansode N.K and P.K sinha which consists of triangle face, diamond face, heart face, oval
face, round face and square faces. Table.1. is the design of the face of a female avatar, table.2. is
a female avatar's eyebrow design, table.3. is a female avatar's eye design, table.4. is a female
avatar's nose design, table.5. is a female avatar's lip design, table.6. is a female avatar's ear
design.
Table.1. Face shape female avatar

Triangle face Diamond face Heart face

Round Oval face Square

Table.2. Eyebrow shape

Down Top Thin Short

Thick Triangle Circular Curved Straight

Table.3. eye shape

Slanted eyes Long eyes Long eyes Long eyes

Little eyes Little eyes Round eyes Round eyes

Almond eyes Eye into Round eyes Round eyes

Almond eyes Almond eyes Almond eyes Wide eye

Wide eye Wide eye Wide eye Wide eye

Table.4. Design nose shape

Round nose Long nose Asyimetrical Pug nose Sharp nose Eagle nose
 Table.5. Design lips shape

Thin Wide Thick bottom Thick up

Little Heart Thick

Table.6. Design ear shape

Little ear lobe Long ear lobe Thick ear lobe Two lobe ear
3.2 Data shape vectorization
The next process is vectorizing the data on the female avatar's face features, before
manufacture on the face survey game. Vector data is needed to facilitate data processing with
the K-means algorithm. These facial data features are arranged according to similarities.
Table.7. data structure table for face shape, table.8.table for ear shape, table.9. table for
lips shape, table.10.table for eyebrows shape, table.11. table for eye shape, table.12. table for
nose shape. The point is to sort face shapes based on the similarity of data so that it is easier to
clustering using k-means.
Table.7. Data structure face shape
Face shape elbow side Curved side Blunt side Pointed side Lengthwise Round
Triangle 0 0 0 1 0 0
Diamond 0 0 0 1 1 0
Heart 0 0 1 1 0 0
Round 0 1 1 0 0 1
Oval 0 1 1 0 1 1
Square 1 0 0 1 1 0
Table.8. Data structure ear shape
Shape ear Big Wide Thick Long Thin Short Narrow Small
Little ear 0 0 0 0 1 1 1 1
Long ear 1 0 1 1 0 0 1 0
Thick ear 1 1 1 0 0 1 0 0
Two lobe ear 1 1 1 1 0 1 1 1
Table.9. Data structure lip shape
Lip type Thick up Thick down Big Wide Thin up Thin down Little
Thin lip 0 0 0 1 1 1 0
Wide lip 0 0 0 1 1 1 0
Thin upper lip 0 1 1 1 1 0 0
thick upper lip 1 0 1 1 0 1 0
Little lip 1 1 0 0 0 0 1
Heart lip 1 1 0 0 0 0 1
Thick lip 1 1 1 1 0 0 0
Table.10. Data structure eyebrows shape
Eyebrow shape Thick Long Straight Curved Thin Short Up down Even
Down 0 1 0 1 1 0 0 1 0
Top 0 1 0 1 1 0 1 0 0
Thin 0 1 0 1 1 0 1 0 0
Short 1 0 1 0 0 1 1 0 0
Thick 1 1 0 1 0 0 0 0 1
Triangle 1 1 0 1 0 0 0 0 1
Circular 1 1 0 1 0 0 0 0 1
Curved 1 1 0 1 0 0 1 0 0
Straight 1 1 1 0 0 0 0 0 1
Table.11. data structure eye shape
Eye shape Wide Big Long Short Narrow Little
Slanted eyes 0 0 1 0 0 0
Long eyes 0 0 1 0 1 0
Little eyes 1 0 0 1 0 1
Eyes into 1 0 1 0 0 1
Round eyes 1 1 0 1 0 0
The eyes of almond 1 1 1 0 0 0
Wide eyes 1 1 1 0 0 0
Table.12. Data structure nose shape
Nose shape Big long Wide Sharp Round Short Narrow Little Pug
Round 0 0 1 0 1 1 0 0 0
Narrow long 0 1 0 1 0 0 1 0 0
Asymmetral 1 1 0 0 0 0 0 0 1
Pug 1 1 0 0 0 0 1 1 1
Sharp 1 1 0 1 0 0 1 0 0
Eagle 1 1 1 1 0 0 0 0 0
3.3 Clustering of face data using K-means
The next process after collecting data is clustering data using K-means. Clustering of
face data using K-means based on the similarity of shapes will be divided into several clusters
and cluster centers randomly.
3.3.1 K-Means Clustering
K-Means is one of the most effective unsupervised clustering algorithm techniques
using the clustering partitioning approach. This method partitioned existing data into groups so
that data with the same characteristics were entered into the same group and data with different
characteristics were grouped into other groups.
In this research 3 clusters were seen according to the clustering with similarities of
shapes starting from the face shape divided by 3 clusters because those with similar shapes
include triangular faces similar to diamond faces and similar to the faces of hearts but different
from oval faces, oval faces them selves have similarities with a round face and different from a
square face, in the form of clusters 1 triangle faces, diamond faces and hearts faces while
cluster 2 faces oval and round faces and in cluster 3 there is a square face only. These centroids
must be placed in a cunning manner because different locations cause different results. So, the
best choice is to put them as much as possible from each other. The application in this study is
that the K point of the centroid is randomly chosen to take 3 centroid points from 100 samples.
The objective function has been calculated as follows:
(1)
Where xi = 1-100 survey data is separated from the feature, while yi is the K point of centroid
and separated by its features. Where as for i is a member or sample data.
3.4 Testing of K-means results with silhouette coefficien
Testing the results of K-means grouping with silhouette cofficien is useful to see
whether the k-means results are good or not. The calculation results of the silhoutte coefficient
value have a range between -1 to 1. The results can be said to be good if they are positive, this
means the points are already in the right cluster. Where as if this value is negative it indicates
overlapping so that the point is between two clusters.
3.4.1 Silhouette Coefficient
The silhouette coefficient method serves to test the quality of the cluster produced as
well as a method for validating a cluster that combines the cohesion method and the separation
method. To calculate the silhoutte coefisient value, the distance between objects is needed by
using the euclidean distance method. The coefficient silhoutte value can be determined using
the following formula: (1)

Where : Si: silhoutte coefficient value. Sedangkan bi: average one point distance with all data in
one cluster. Dan ai: minimum average distance from one point to another different cluster.
The calculation results of the silhoutte coefficient value have a range between -1 to 1.
The results can be said to be good if they are positive, this means the points are already in the
right cluster. Whereas if this value is negative it indicates overlapping so that the point is
between two clusters.
3. RESULTS AND DISCUSSION
The results of face data were obtained from a survey of 100 people consisting of 5
psychology students, 4 electro-education students, 5 Indonesian literary students, 11 fine arts
students, 9 English literature students, 15 dance art students, 29 visual communication design
students, 16 machine students and 6 electro students. Figure.2. is a mapping of the face data
survey results conducted with 100 student samples.

Figure.2. Result survey data

Figure.3. The results of the face data survey are grouped into 3 cluster regions.
 Figure.3. Mapping of survey result with 3 cluster region
The data obtained is then randomly randomized and formed into 3 classes. The class is
determined by the similarity of the shape of its features starting from:
1. The face shape for the triangle face, the diamond face and the heart face are joined together
in one class, then the oval face and round face in one class and the last is the square face
2. Form eyebrows down, eyebrows up, thin eyebrows put together in one class, then straight
eyebrows, triangle eyebrows, and short eyebrows into one class, and the last thick
eyebrows, circular eyebrows and curved eyebrows into one class.
3. The eyes for slanted eyes and long eyes become one class, then small eyes, eyes into and
round eyes into one class and the last is the eyes of almonds and wide eyes in one class.
4. The shape of the nose for a round and long nose is one class, then the nose is not
symmetrical and the nose is one class and the last one is the nose and eagle nose one class.
5. Shape lips for thin lips and wide lips one class, then thick bottom lips and thick lips over
one class, and the last small lips, heart lips and thick lips of one class.
6. The shape of the ear for a small lobe ear is one class, then the ear long lobe is one class and
the last ear is thick lobe and ear two lobes are made into one class.
Figure 4 is the distribution of data based on 3 clusters and the third position of the centroid

Figure.4. The results of clustering data of cluster 1, 2, and 3.

Clustering face data with k-means clustering by determining 3 centroids that are
randomly determined, the results of the data are divided into 3 clusters and then iterated
repeatedly until the cluster members do not move or converge.
 (a) (b) (c)
Figure.5. Centroid point cluster 1,2, and 3
Figure 5 (a) is a Centroid point cluster 1 with triangle face shape, circle left eyebrow,
up right eyebrow, slanted eyes, sharp nose, wide lips, narrow long lobe ears. Figure 5 (b) is a
Centroid point cluster 2 with oval face shape, triangle left eyebrow, short right eyebrow, wide
eyed , asymmetrical nose, thick upper lip, thick lobe ears. Figure 5 (c) is a Centroid point
cluster 3 square face shape, thin eyebrow, round eyes, round nose, thick lips, thick lobe ears.
Table 13 is result total of cluster members. For those who are members of cluster 1,
there are 3 psychology students, 1 Indonesian literature, 7 English literature, 3 fine arts, 8 dance
arts, 7 machines, 10 DKV, electro 1. Then for those who are members of cluster 2, psychology
1 students, education . Electrical 2, electro 4, Indonesian literature 1, English literature 1, visual
arts 6, dance 3, machines 4 and DKV 11 and members of cluster 3 namely electro education 2,
Indonesian literature 3, English literature 1, visual arts 2, machinery 5, dance 5, DKV 7 and
electro 1. Table.13. Result total of cluster members

Total of cluster members

Cluster 1 39
Cluster 2 34
Cluster 3 27

Figure.6. Face shapes in cluster 1 or selected face clusters

Figure 6 is a sample of 3 selected faces that are included in cluster 2, the first named
rafa, choose triangle faces, thin eyebrows, wide eyes, round nose, lips and thick lobe ears, then
the faces of the two students who choose named fitri choose diamond faces, thin eyebrows ,
wide eyes, pug nose, thick lips and then the third student who chose to be named fanda and
chose triangular faces, eyebrows up, eyes in, nose asymmetrical, thin lips and small lobe ears.

Figure.7. Face shapes in cluster 2

Figure 7 is the results of the several member of cluster 2, namely the student name Eric,
chose an oval face, thick eyebrows, almonds, thin nose and thin upper lip. Then the names of
bayu students choose round faces, eyebrows down, round eyes, asymmetrical nose, thick lips
and small lobe ears. And the third student named Natasha chose a round face, eyebrows down,
almonds eyes, round nose, thick upper lip, and small lobe ears.
 Figure.8. Face shapes in cluster 3
Figure 8 is the results from cluster 3 members, namely the first student named Anabela
chose a square face, eyebrows down, round eyes, an asymmetrical nose, small lips and ear ears,
then a second member named Aida chose a square face, circular eyebrows, wide eyes, the nose
is not symmetrical, the lips of the heart, and the third named eloise chooses a square face, right-
eyed almonds and slanted left eye, snub nose, thin upper lip and small lobe ears.
To test the results of the k-means overlap or not, it is tested using a silhouette
coefficient. If the positive results of the analysis can be said to be good, this means that the
points are already in the right cluster. But there are still negative results so the result of k means
that it is still not good and the various cluster members are not in the right cluster. Then stated
the calculation of grouping data using k-means is correct. Figure 9 is the results of the
silhouette coefficient for the calculation of this research found little overlapping between cluster
1 and cluster 2.

Figure.9. The results of the silhouette coefficient calculation

4. CONCLUSION
This K-means algorithm has proven to be effective and quickly finds the grouping of
favorite women's face data. The result was 39% choosing triangle face shape, circle left
eyebrow, up right eyebrow, slanted eyes, sharp nose, wide lips, narrow long lobe ears, then
34% choosing oval face shape, triangle left eyebrow, short right eyebrow, wide eyed ,
asymmetrical nose, thick upper lip, thick lobe ears and 27% choose square face shape, thin
eyebrow, round eyes, round nose, thick lips, thick lobe ears. For those who are members of
cluster 1, there are 3 psychology students, 1 Indonesian literature, 7 English literature, 3 fine
arts, 8 dance arts, 7 machines, 10 DKV, electro 1, then for those who are members of cluster 2,
psychology 1 students, education . Electrical 2, electro 4, Indonesian literature 1, English
literature 1, visual arts 6, dance 3, machines 4 and DKV 11 and members of cluster 3 namely
electro education 2, Indonesian literature 3, English literature 1, visual arts 2, machinery 5,
dance 5, DKV 7 and electro 1. And silhouette coeffisient proves that the results of the
calculation of k-means analysis above are not good and there are points that are not in the right
cluster. Then it can be concluded that the favorite face is with a triangle face shape, circle left
eyebrow, up right eyebrow, slanted eyes, sharp nose, wide lips, narrow long lobe ears.
5. ACKNOWLEDGEMENT
We would like to thank Mr. Eko Mulyanto as the supervisor, Mr. Supeno Mardi as the
supervisor, Mr. Mitra Istiar Wardhana as the supervisor, all residents of the Technology
Institute Sepuluh November who helped complete the research and friends of the 2017
Intelligent Multimedia Network.
6. REFERENCES
[1] V. Faber, “Clustering and the Continuous k-Means Algorithm,” Los Alamos Sci., 1994.
[2] A. Kannan, J. Winn, and C. Rother, “Clustering appearance and shape by learning
jigsaws,” Adv. Neural Inf. Process. Syst., 2006.
[3] A. Hartigan and M. A. Wong, “A K-Means Clustering Algorithm,” J. R. Stat. Soc., 1979.
[4] S. C. Zhang, B. Fang, Y. Z. Liang, J. Wen, and L. Wu, “A face clustering method based
on facial shape information,” in International Conference on Wavelet Analysis and
Pattern Recognition, 2011.
[5] Biocca, F., & Nowak, K. (2002). Plugging your body into the telecommunication system:
Mediated embodiment, media interfaces, and social virtual environments.
[6] Bull, P., & Rumsey, N. (1988). The social psychology of facial appearance . New
York: Springer-Verlag
[7] Damer, B.(1997). Avatars!: Exploring and buildingvirtual worlds on the Internet.
Berkley, CA: Peachpit Press
[8] Dietrich, D.(1997). (Re)fashioning the techno-erotic woman: Gender and textuality in
the cybercultural matrix.
[9] Nowak, K. L., & Biocca, F. (2003). The effect of the agency and anthropomorphism on
users’ sense of telepresence, copresence, and social presence in virtual
environments. Presence: Teleoperators and Virtual Environments
[10] Rauh, C., Polonsky, M., & Buck, R. (2004). Cooperation at first move: Trust, emotional
expressiveness and avatars in the prisoner’s dilemma game.
[11] Schroeder, R. (2002). Social interaction in virtual environments: Key issues, common
themes, and a framework for research.
[12] Taylor, T. L. (2002). Living digitally: Embodiment in virtual worlds. In R. Schroeder
(Ed.), The social life of avatars: Presence and interaction in shared virtual environments
(pp. 40–62). London: Springer-Verlag
[13] Turkle, S. (1995). Life on the Screen: Identity in the age of the Internet. New
York: Simon & Schuster
[14] Waskul, D., & Douglass, M. (1997). Cyberself: The emergence ofself in on-line
chat. The Information Society , 13(4), 375–397.
[15] West, C., & Zimmerman, D. (1991). Doing gender. InJ.Lorber & S.
A. Farrell (Eds.), The social construction of gender(pp. 13–37). Newbury Park, CA: Sage
Publications
[16] Kristine L. Nowak Christian Rauh, " The Influence of the Avatar on Online Perceptions
of Anthropomorphism, Androgyny, Credibility, Homophily, and Attraction ," USA,2017
[17] N K Bansode and P K Sinha , “Face Shape Classification Based on Region Similarity,
Correlation and Fractal Dimensions l,” University of Pune, 2016
[18] Pomthep Sarakon and Theekapun Charoenpong., "Face Shape Classification from 3D
Human Data by using SVM" , Sri nakharin wijot University, 2014.