Application of an Automatic Data Alignment & Structuring System

for Intercultural Consumer Segmentation Analysis

Fumiko Kano Glückstad

Dept.of International Business Communication, Copenhagen Business School, Dalgas Have 15, Frederiksberg, Denmark

Keywords: Ontology Alignment, Knowledge Structuring, Nonparametric Bayesian Relational Modelling,

Cross-cultural Data Analysis, Opinion Survey Analysis, Consumer Segmentation.

Abstract: This position paper introduces a conceptual framework of our ambitious international research project

where the aim is extraction and alignment of heterogeneous consumer segment structures across a

multiplicity of markets and cultures. We argue that an automatic data alignment and structuring system

employing a non-parametric Bayesian relational modelling is an ideal approach that can address challenges

in the conventional cross-cultural data analysis. The paper presents an example of our preliminary work that

applies this approach to the analysis of opinion survey responses given by male populations in Sweden and

Japan. The framework successfully extracts groups of males who express similar but also dissimilar

response patterns from the two selected countries. Based on these preliminary studies, the paper discusses

potential contributions and future challenges of the international consumer analysis project.

1 INTRODUCTION

The establishment of semantic relations across

boundaries of e.g. organizational-, linguistic-, or

societal communities (Fensel et al., 2004) has been a

central topic of the ontology matching discipline

(Euzenat and Svaiko 2007) and substantial numbers

of ontology matching frameworks have been

introduced in recent years (Berlin and Motro, 2002;

Cheng et al., 2008; Doan et al., 2004; Euzenat, 1994;

Stumme and Mädche, 2001; Wang et al., 2004;

Mørup et al., 2014). However, a fully automatic

ontology alignment has remained a highly

challenging task since it typically involves similarity

computations between objects belonging to different

knowledge systems. As indicated in previous works

(Isaac et al., 2007; Pirrò and Seco, 2008; Pirrò and

Euzenat, 2010; Ngo et al., 2013; Cross et al., 2013;

Glückstad et al., 2014; Mørup, 2014), similarity

measures influence on the performance of ontology

alignment tasks. Mørup et al. (2014) addresses this

ambiguity by introducing a new framework applying

the nonparametric Bayesian relational modeling to

statistically analyze feature matches of all

combinations of objects belonging to different

knowledge systems, i.e., to align them while jointly

partitioning them into clusters. This new way of

analyzing “relatedness” between objects belonging

to different knowledge systems engenders an

additional ability to automatically align three or

more ontologies simultaneously (Mørup et al.,

2014).

Whereas this ability to automatically align and

structure multiple datasets contributes substantially

to the knowledge engineering and ontology

development discipline (e.g. Glückstad et al., 2014;

Mørup et al. 2014), such ability could also be useful

in other research disciplines which requires the

analysis of “relatedness” of objects across multiple

datasets. This position paper introduces our

ambitious international research project proposal

which sheds light on the potential contributions of

such automatic data alignment and structuring

framework to other research disciplines, namely, the

cross-cultural consumer segment analysis.

The next section first introduces the conceptual

framework of the proposed project followed by

potential contributions of the automatic data

alignment and structuring approach to the extraction

of heterogeneous consumer segment structures

across a multiplicity of cultures. Finally, we briefly

show an example of our preliminary study

examining the applicability of Mørup’s framework

to analyse cross-cultural opinion survey responses.

Glückstad, F..

Application of an Automatic Data Alignment & Structuring System for Intercultural Consumer Segmentation Analysis.

In Proceedings of the 7th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2015) - Volume 2: KEOD, pages 251-256

ISBN: 978-989-758-158-8

251

2 CONCEPTUAL FRAMEWORK

International marketing professionals who have

extensive professional experience in the Far East

markets are aware that consumers in China or Japan

cannot simply be categorized by a standardized

consumer segmentation model universally used in

the Western marketing practice. They need scientific

and empirical documentation comprehending

similarities and differences of consumer segments

across boundaries of societies. Cleveland et al.,

(2011a) emphasizes that: “a proportion of

individuals worldwide develop bicultural identities:

one based in local traditions combined with an

identity connected to an emerging global culture”

and “as corporations globalize, the key challenge for

managers is to institute an effective marketing

orientation across a composite of cultures”.

Values are crucial for explaining the

motivational basis of human attitudes and behaviour

(Schwartz, 2012). Individuals act, make decisions,

emotionally react to advertisements, purchase and

assess products based on their identity-based

motivations. However, the identity formation and

thereby value formation of modern consumers are

becoming increasingly complex due to their

belonging to local, national and global communities

accessible via e.g. contemporary media

technologies. Hence, it is crucial to map out and

comprehend the complexity of consumers, i.e. the

heterogeneity of consumer segment structures across

societies (Cleveland et al., 2011a; Cleveland et al.,

2011b). Our hypothesis is that Schwartz’s Theory of

the Ten Basic Human Values (STBH: Schwartz,

2012) is a useful measurement scale for contrasting

consumers across a multiplicity of cultures (Fischer

and Schwartz, 2011). We hypothesize that, by

analysing the existing data from the World Value

Survey (WVS) including the question items of

STBH, it is feasible to develop a new measurement

scale that can effectively extract heterogeneous

consumer segment structures across cultures.

From a technical aspect, our hypothesis is that

the nonparametric Bayesian relational modelling is a

powerful and suitable approach (Schmidt and

Mørup, 2013; Mørup and Schmidt, 2012) for

extracting and aligning consumer segments existing

across a multiplicity of markets. The model jointly

partitions consumers in the respective markets into

segments representing either universal value patterns

across cultures or culturally specific value patterns.

By inspecting behavioural data connected to

members of the value-based consumer segments

extracted by the proposed approach, the project will

eventually be able to infer the preferences of

targeted consumers and their predicted behaviours.

Accordingly, we eventually intend to develop a data-

analytic platform and test its usability by

questioning:

• What type of consumer segments will potentially

be interested in [specific products/services] in

selected foreign markets?

• How can the potential consumer segment of

[specific products/services] be characterized?

Are the characteristics universal across cultures

or culturally specific?

• Are the newly emerged segments, such as the

“Millenials”, a transnational- or a nationally

specific phenomenon across these markets?

Figure 1: The conceptual framework.

Figure 1 illustrates the conceptual framework of

the intercultural consumer segmentation system. In

this framework, feature vectors corresponding to

responses to the question items of STBH given by

each consumer are used for segmenting consumers.

By analysing the “relatedness” between responses of

consumers (feature vectors) within and across

cultures, a methodological framework employing the

nonparametric Bayesian relational model

simultaneously extract consumer segments similar

but also dissimilar across cultures. The next section

briefly reviews this joint clustering mechanism.

3 METHODOLOGICAL

FRAMEWORK

The automatic data alignment framework introduced

in Mørup et al., (2014) employs a statistical model

belonging to a family of stochastic block-models

widely used in the social network analysis (Faust

and Wasserman, 1992; Wasserman and Anderson,

1987) and also applied in the complex brain

connectivity network (brainconnectivity.compute.

KEOD 2015 - 7th International Conference on Knowledge Engineering and Ontology Development

252

dtu.dk, Mørup et al., 2010; Mørup and Schmidt,

2012; Schmidt and Mørup, 2013; Schmidt et al.,

2012). The principle of the stochastic block-model is

to partition a set of objects into clusters whose

members are stochastically equivalent and possess

similar relations to other clusters to be extracted in

the partitioning process. In Mørup’s framework

(Mørup et al., 2014), an extended version of the so-

called Infinite Relational Model (IRM: Kemp et al.,

2006; Xu et al., 2006), a family of binary stochastic

block-models is employed. As is explained in Kemp

et al. (2006), the uniqueness of the IRM is to

automatically extract an appropriate number of

clusters based on the so-called Chinese Restaurant

Process (CRP: Aldous, 1985; Pitman, 2002).

Whereas the original IRM is suitable solely to

the ontology learning, the extended IRM, called the

multinominal IRM (mIRM) in Mørup et al. (2014),

is based on a multinominal observation likelihood

and is suitable to align multiple datasets. A

distinctive characteristic of the mIRM framework is

to first count all combinations of binary feature

matches between objects, which will be used for the

alignment across datasets during the partitioning

process employed in the mIRM. Specifically, when

aligning two datasets whose objects consist of

respective feature vectors, four combinations of

matches are counted: i) 1-1 match means that an

object from System A and an object from System B

possess the same binary feature; ii) 1-0 match means

that an object from System A possess a feature that

is not possessed by an object from System B; iii) 0-1

match means that an object from System B possess a

feature that is not possessed by an object from

System A; and iv) 0-0 match means that a feature is

possessed neither by an object from System A nor

by an object form System B. Following the same

principle, the alignment of three systems counts

eight combinations of matches: i) 1-1-1; ii) 1-1-0;

iii) 1-0-1; iv) 1-0-0; v) 0-1-1; vi) 0-0-1; vii) 0-1-0;

and viii) 0-0-0. This further implies that it is feasible

to align four or more systems by counting all

possible combinations of matches, i.e. 16

combinations in case of four systems, 32

combinations for five systems and so forth. This

count statistics of binary feature matches replaces

similarity measures discussed in the previous

literatures (Isaac et al., 2007; Pirrò and Seco, 2008;

Pirrò and Euzenat, 2010; Ngo et al., 2013; Cross et

al., 2013; Glückstad et al., 2014).

The count statistics obtained by this process is

further used for simultaneously partitioning and

aligning sets of objects between multiple systems.

For example, in the alignment of two systems, the

mIRM is defined as the following generative process

where objects in the two systems are simultaneously

partitioned into z and w clusters according to the

CRP.

~









, ~









,





~





,





,



~











,



.

The probabilities that the four types of matches are

observed between an extracted cluster l in System A

and cluster m in System B are defined as





according to the Dirichlet distribution. Finally, the

count statistics of the four types of matches





,



 between an object i in System A and an

object j in System B are drawn from the

multinominal distribution where the between-cluster

probabilities for each of the four combinations of

matches are defined as











. For those interested in

further details of the algorithms, please refer to:

Kemp et al., (2006; 2010) for the original IRM;

Schmidt and Mørup (2013) for the general review of

the nonparametric relational modelling; and Mørup

et al., (2014) for the mIRM.

Some of the major contributions demonstrated by

Mørup et al., (2014) are that i) the automatic

alignment of two as well as three knowledge

systems has been achieved and ii) the alignment of

systems have not been influenced by the choice of

similarity measure. While Mørup et al. demonstrated

the alignment of knowledge systems across two and

three legal systems (educational systems of two and

three countries defined by UNESCO), the present

work applies this analytical framework to a new type

of dataset, namely cross-cultural opinion survey

responses. The latter analysis demonstrates how the

alignment results of the four types of matches are

interpreted in the domain of opinion survey analysis.

4 PRELIMINARY STUDY

4.1 Datasets

For applying the mIRM framework to a cross-

cultural opinion survey analysis, the 6

Wave of

World Value Survey (WVS) datasets downloadable

from the WVS organization web-site are employed.

The data used in this analysis are collected from 557

Swedish and 954 Japanese males who responded to

all ten question items V79-89 corresponding to the

Portrait Values Questionnaire (PVQ) developed by

Schwartz (2012). For example, the portrait of “Self-

Direction value (V79)” describes: “Thinking up new

Application of an Automatic Data Alignment & Structuring System for Intercultural Consumer Segmentation Analysis

253

ideas and being creative is important to him. He

likes to do things in his own original way.”

Accordingly, respondents must select an answer

from the six-level Likert scale: 1. Very much like

me; 2. Like me; 3. Somewhat like me; 4. Little like

me; 5. Not like me; and 6. Not at all like me. In this

analysis, these six answer categories are

semantically divided into a positive answer covering

1-4 and a negative answer including 5 and 6. This

means that there are 2

patterns for answering these

ten PV Q questions.

4.2 Analysis of the Aligned Results

The mIRM has been run 20 times with 5000

iterations in this analysis. The Normalized Mutual

Information (NMI) scores for the 20 runs are 0.91

for the Japanese and 0.92 for the Swedish. Figure 2

displays that the mIRM extracted in total 81 and 58

clusters respectively from the Japanese and Swedish

datasets.

Figure 2 displays four plots named as 1-1

matches, 1-0 matches, Jaccard and SMC (Simple

Matching Coefficient). The two plots names as “1-1

matches” and “1-0 matches” are examples of the

clusters extracted by the mIRM estimating an

appropriate number of clusters to be extracted based

on the count statistics of the four combinations of

the matches (1-1, 1-0, 0-1, and 0-0). The

intersections uniformly highlighted with yellow in

the plot of “1-1 matches” indicate that all cluster

members between the Japanese (Y axis) and Swedes

(X axis) agree either positive or negative on all the

ten PVQ question items.

Figure 2: Clusters extracted and aligned between Japanese

and Swedish males.

Figure 3: Cluster mean scores of the ten question items, size of the clusters, and the proportion of the positive and negative

answers indicated by the members of the respective clusters.

KEOD 2015 - 7th International Conference on Knowledge Engineering and Ontology Development

254

On the other hand, an intersection highlighted with

colours closer to yellow in the plot of “1-0 matches”

indicate that a Japanese and a Swede belonging to

the intersection disagree on many of the ten question

items. The plots named as “Jaccard” and “SMC”

displays the levels of similarity scores computed

between Japanese and Swedish respondents

belonging to the respective clusters extracted by the

mIRM. A noteworthy point is that, if these similarity

measures are used for co-clustering the same

respondents, the clustering results will be influenced

by the distribution of scores between the highest

“1.0” and the lowest “0.0” as shown in (Glückstad et

al., 2014).

Figure 3 further inspects the actual responses of

the cluster members to the ten PVQ question items

(V79-89). The four plots in Figure 3 show the

overall results for the STBH values for specific

cluster combinations: J1-S1; J2-S2; J4-S5; and J3-

S2. For example, Figure 3 indicates that all

members belonging to the Japanese cluster 1 (JM1:

47 respondents) and the Swedish cluster 1 (SM1: 76

respondents) answered positively to all ten PVQ

questions. On the other hand, all members in JM2

and SM2 responded negatively to the question

regarding the “Power” value. These two

combinations of clusters are the ones connected via

the uniformly yellow intersections in “1-1 matches”

of Figure 2. In the case of JM3-SM2 combination,

members within the respective clusters all agree on

either negative or positive responses. However,

disagreement appears between JM3 and SM2 in

“Stimulation” and “Achievement” values. These

clusters are linked with the intersection uniformly

coloured with Orange in “1-1 matches” of Figure 2.

Finally, in the combination between JM4 and SM5,

members of JM4 disagree on several questions such

as “Conformity” value within JM4 as well as

disagree with SM5 on PVQ items such as

“Stimulation” value. These clusters are linked with

the green coloured intersection, which indicate that

the relation between the two clusters is rather

ambiguous, i.e. neither similar nor dissimilar.

5 CONCLUDING REMARKS

The previous section demonstrated the joint

clustering of two datasets compiled from the ten

PVQ question items representing STBH values. The

mIRM framework applied in this work has extracted

groups of people who express similar but also

dissimilar response patterns from the respective

datasets. The extracted clusters are aligned between

the two culturally specific datasets, of which

between-cluster relations are observable in the

matrices displayed in Figure 2. A noteworthy point

is that these matrices illustrate the overall similarity

relations of respondents within- and across-clusters

based on the differentiated colours. This means that

these similarity relations might be further used for

effectively visualizing the hierarchical structures of

the extracted clusters by use of existing ontology

learning technologies.

Another important remark is that the application

of the ontology engineering to the conventional

cross-cultural data analysis may add substantial

values to this specific research domain.

Traditionally, cross-cultural data analyses in Social

Sciences have rather focused on the analysis of

mean values of predefined segments such as “a

group of Japanese female teenagers” and based on

researcher’s hypotheses and comparisons across

cultures. By use of ontology engineering

technologies that can focus on structuring relations

between individuals based on their feature vectors,

the heterogeneous segment structures can be

extracted in a data-driven manner. Such an approach

is particularly needed for comprehending and

comparing heterogeneity across societies. Recalling

that in our contemporary society, individuals’ values

are shaped and influenced by a multiplicity of layers

ranging from the mono-cultural (local), multicultural

(regional) and transcultural (universal) layers in the

modern global world. The present work has

demonstrated a pre-cursor for challenging such new

and exciting research endeavours.

ACKNOWLEDGEMENTS

I would like to express my sincere gratitude to Profs.

Morten Mørup and Mikkel N. Schmidt who have

originally developed the mIRM model and the

related analytical toolbox used in the present work. I

also thank Microsoft Denmark and Techila

Technologies who provided a high performance

cloud-computing infrastructure that has successfully

reduced the computational time in the present work

from 180 hours (under the “non” parallel computing

environment using CPU: Intel Core i5M480

@2.67GHz, Memory: 4 GB) to just 40 minutes.

REFERENCES

Aldous, D. 1985. Exchangeability and Related Topics. in

Application of an Automatic Data Alignment & Structuring System for Intercultural Consumer Segmentation Analysis

255

E´Cole D’e´Te´ De Probabilite´S Desaint-Flour

XIII1983, 1–198

Berlin, J. & Motro, a. 2002. Database Schema Matching

using Machine Learning with Feature Selection. in

Proceedings of the 14th International Conference on

Advanced Information Systems Engineering, Vol. 2348

of Lecture Notes in Computer Science, 452–466

Cheng, C.P., Lau, G.T., Law, K.H., Pan, J. & Jones, a.

2008. Regulation Retrieval using Industry Specific

Taxonomies. in Artif Intell Law 16, Springer. 277–

303.

Cleveland, M., Erdoğan, S., Arıkan, G., Poyraz, T. 2011a.

Cosmopolitanism, Individual-Level Values and

Cultural-Level Values: A Cross-Cultural Study. in

Journal of Business Research 64, 934–943.

Cleveland, M., Papadopoulos, N., Laroche, M. 2011b.

Identity, Demographics, and Consumer Behaviors:

International Market Segmentation across Product

Categories, in Intl. Marketing Review, 28(3), 244-266

Cross, V., Yu, X. and Hu, X. 2013. Unifying Ontological

Similarity Measures: a Theoretical and Empirical

Investigation. in International Journal of Approximate

Reasoning, Vol.54, 861–875.

Doan, A.H., Madhavan, J., Domingos, P. & Halevy, a.

2004. Ontology Matching: a Machine Learning

Approach. in: Staab S, Studer R (Eds) Handbook on

Ontologies, Chapter 18. Springer, Berlin, 385–404

Euzenat, J. 1994. Brief Overview of T-Tree: the Tropes

Taxonomy Building Tool. in Proceeding of the 4th

ASIS SIG/CR Workshop on Classification Research,

69–87

Euzenat, J. & Shvaiko, P. 2007. Ontology Matching.

Springer, Berlin.

Faust, K. & Wasserman, S. 1992. Blockmodels:

Interpretation and Evaluation. in Social Networks, 14

(1-2), 5-61

Fensel D., Davies, J., Bussler, C. and Studer R. (Eds.)

2004. the Semantic Web: Research and Applications.

in First European Semantic Web Symposium,

Heraklion, Crete, Greece, Springer-Verlag, Berlin.

Fischer, R. Schwartz, S.H. 2011. Whence Differences in

Value Priorities? Individual, Cultural, or Artifactual

Sources. in Journal of Cross-Cultural Psychology, Vol

42, 1127-1144

Glückstad, F.K., Herlau, T., Schmidt, N. M., & Mørup, M.

2014. Cross-Categorization of Legal Concepts across

Boundaries of Legal Systems: in Consideration of

Inferential Links. in Artif Intell Law 22 (1), Springer,

61-108.

Isaac, A., Meij, L.V.D., Schlobach, S. & Wang, S. 2007.

an Empirical Study of Instance-based System

Matching. in the Semantic Web, Lecture Notes in

Computer Science, Vol.4825, 253–266.

Kemp, C., Tenenbaum, J.B., Griffiths, T.L. Yamada, T. &

Ueda, N. 2006. Learning Systems of Concepts with an

Infinite Relational Model. in Proc. the 21st National

AAAI Conference on 1:381-388.

Mørup, M., Glückstad, F.K., Herlau, T. & Schmidt, N. M.

2014. Non Parametric Statistical Structuring of

Knowledge Systems using Binary Feature Matches. in

Proceedings of 2014 IEEE International Workshop on

Machine Learning for Signal Processing

Mørup, M. & Schmidt, M.N. 2012. Bayesian Community

Detection. in Neural Computation 24 (9), 2434-2456.

Mørup, M., Madsen, K.H., Dogonowski, a.M., Siebner, H.

& Hansen, L.K. 2010. Infinite Relational Modeling of

Functional Connectivity in Resting State Fmri. in

Neural Information Processing Systems, 1750-1758.

Ngo, D., Bellahsene, Z. & Todorov, K. 2013. Extended

Tversky Similarity for Resolving Terminological

Heterogeneities across Ontologies. in OTM 2013,

Lecture Notes in Computer Science, Vol.8185, 711–

718.

Pirrò, G. and Seco, N. 2008. Design, Implementation and

Evaluation of a New Semantic Similarity Metric

Combining Features and Intrinsic Information

Content. in OTM Conferences (2), 1271–1288.

Pirrò, G. & Euzenat, J. 2010. a Feature and Information

Theoretic Framework for Semantic Similarity and

Relatedness. in International Semantic Web

Conference (1), 615–630.

Pitman J. 2002. Combinatorial Stochastic Processes. in

Tech. Rep., Springer.

Schmidt, M.N., Herlau, T. & Mørup, M. 2012.

Nonparametric Bayesian Models of Hierarchical

Structure in Complex Networks. in Informatics and

Mathematical Modelling.

Schmidt, M.N. & Mørup, M. 2013. Non-Parametric

Bayesian Modeling of Complex Networks. an

Introduction. in IEEE Signal Processing Magazine,

Vol. 30(3), 110-128

Schwartz, S. H. 2012. an Overview of the Schwartz

Theory of Basic Values. in Online Readings in

Psychology and Culture, 2(1).

Stumme, G. & Mädche, a. 2001. Fca-Merge: Bottom-up

Merging of Ontologies. in Proceedings of 17th

International Joint Conference on Artificial

Intelligence (IJAI), 225–234

Wang, J., Wen, J.R., Lochovsky, F. & Ma, W.Y. 2004.

Instance-based Schema Matching for Web Databases

by Domain-Specific Query Probing. in Proc. of the

30th Intl. Conference on Very Large Data Bases, 408–

419

Wasserman, S. & Anderson, C. 1987. Stochastic a

Posteriori Blockmodels: Construction and Assessment.

in Sociological Networks, 9, 1-36.

WVS: World Value Survey Association 2009. World

Value Survey Official Data File V. 20090901.

Aggregate File Producer: ASEP/JDS, Madrid

Xu, Z., Tresp, V., Yu, K., & Kriegel, H.P. 2006. Learning

Infinite Hidden Relational Models. in Uncertainty in

Artificial Intelligence (UAI2006)

KEOD 2015 - 7th International Conference on Knowledge Engineering and Ontology Development

256