Consumer Behaviour
Consumer Behaviour
Consumer Behaviour
J. Casillas
Dept. Computer Science and Artificial Intelligence, Computer Engineering School, University of Granada,
E-18071, Granada, Spain
Phone: (+34) 958 240804, Fax: (+34) 958 243317, e-mail: casillas@decsai.ugr.es
F.J. Martínez
Dept. Financial Economy and Accounting, Business Faculty, University of Huelva,
E-21071, Huelva, Spain
Phone: (+34) 959 017891, e-mail: francis@uhu.es
F.J. Martínez-López
Dept. Marketing, Business Faculty, University of Granada,
E-18071, Granada, Spain
Phone: (+34) 958 242350, Fax: (+34) 958 240695, e-mail: fjmlopez@ugr.es
Summary
The present paper tries to point out the necessity that firms focused on consumer markets have to understand and
predict in a higher efficiently manner the behaviour of their target population. Thus, it is convenient that models of
consumer behaviour in which firms are based to take their decisions to be closer to what a real Marketing Management
Support System should offer. In this sense, we propose a new application for consumer behaviour modelling, based on
fuzzy association rules (FAS) for adjusting the data, as a complementary way to the results obtained by using the
classic technique of model estimation based on Structural Equation Modelling (SEM). With this aim, an online
behavioural model is proposed being later tested making use of both FAS and SEM. Finally, a comparative analysis of
the results is done, focusing specially on our proposal of methodological application.
Keywords: consumer behaviour, marketing modelling, model estimation, structural equation modelling, fuzzy
association rules, knowledge discovery
1. INTRODUCTION
It is been deeply highlighted the vital necessity stated by both, marketing academics and practitioners, for knowing and
explaining the consumer’s behaviour patters in a manner increasingly efficient. Firms focused on final markets are
immersed in highly competitive systems in which it is needed that their decision processes to be as correct as possible.
In this sense, models of consumer behaviour, inasmuch as they are marketing models, are considered as a specific case
of Marketing Management Support System (MkMSS), and throughout the time have demonstrated to be a source of
transcendental relevance for the development of marketing science [see van Bruggen and Wierenga (2000)].
Notwithstanding, actual models of consumer behaviour do not seem to cover all the necessities that it should
supposedly satisfy a model which aim to aid on the marketing decision making. With respect to this, based on
Gatignon (2000), future models, considering both their theoretical and technical aspects, which try to explain
consumers’ decision making will have to be clearly focused on users’ (demand side) requirements of such models.
That is to say, models must be more complete, flexible, and customized to the strategic singularities of the competitive
environment which their users operate in. Thus, as the main problem that actually face firms oriented to consumer
markets is not the availability of information (data), but the possession of necessary level of knowledge to take the right
decisions, the use of avant-garde behavioural models able to exploit it may represent an essential source of competitive
advantage.
On the other hand, it is expected that MkMSS will tend to improve their performance taking advantage of synergies
caused by the integration of modelling estimation techniques based on classic econometric with expert systems based
on artificial intelligence.
In this respect, considering the three pillars in which marketing modelling is based [see Roberts (2000)], and more
specifically the consumer behaviour modelling, we focus our work on one of them, i.e.: the modelling estimation
techniques and its improvement, presenting theoretically and empirically the potentials that methods of estimation
based on fuzzy association rules (FAS) present to propose and obtain estimated models of behaviour more exhaustive,
complex, flexible, interactive and which offer much more quantity of qualitative information than preceding estimation
techniques used in this field [Gatignon (2000); van Bruggen and Wierenga (2000)]. In this sense, FAS can be a
plausible alternative to complement, in principle, the results obtained by using Structural Equation Modelling (SEM)
techniques which have been the ones usually used in the last decades to estimate complex models of consumer
behaviour. Furthermore, even considering the evident utility offered by SEM to test theoretical models proposed, it
shows a series of limitations, that could be solved by FAS estimation techniques, inasmuch as: (1) supposes a linear
relation among variables of the model, so it does not allow to analyze nor, therefore, interpret relations among several
variables when such relations are contemplated with different degrees of intensity; (2) usually works with “simple” or
recursive causal models, without considering possible reverse relations among variables, so parameters obtained
making use of SEM specific statistical software (e.g.: LISREL, AMOS, etc.) will tend to be biased; and at last but not
least (3) estimation techniques based on SEM are useful to test theoretic proposals for a consumer behaviour model,
though its utility to support marketing decisions (MkMSS) seems to be restricted by its own results derived from its
process of estimation [see Laurent (2000); Steenkamp and Baumgartner (2000)].
Therefore, this paper proposes the use of fuzzy systems as a knowledge discovery tool to allow marketing academics
and practitioners to improve the understanding of consumer behaviour. With this aim, the paper is organized as
follows. Section 2 highlights the relevance that methods based on FAS can have to help satisfying the previous idea
when applied to model consumer behaviour. Thus, a methodological application is proposed based on the Cooperative
Rules methodology. Section 3 describes a theoretical model of online consumer behaviour that will be used as a
benchmark to show how our methodology works. Section 4 shows the results obtained by making use of both, our
application and SEM, and it treats the main implications that results coming from our application have to understand
the marketing problem posed. Finally, Section 5 highlights, from a methodological point of view, several interesting
findings resulting from our application.
2.2. METHODOLOGY FOR CONSUMER BEHAVIOR MODELING WITH FUZZY ASSOCIATION RULES
This section introduces the process we propose to perform the knowledge discovery by fuzzy association rules.
Basically, it consists on preparing the data and on fixing the scheme we follow to represent the knowledge existing in
the data. Once defined these aspects, a specific learning method is used to automatically design the fuzzy association
rule system.
The process is the following:
• Data collection: as traditionally done in marketing, it is extracted by means of a questionnaire in a similar
way to the models estimated by structural equation modelling.
• Data processing: it is necessary to adapt the collected data to a scheme easily tractable by fuzzy system
learning methods. Thus, when more than one item is used to asses the same concept, the arithmetic mean
value is used.
• Representation: fuzzy association rules are used to represent the relationships between the variables. Once
fixed by the marketing expert the structural model, a fuzzy rule base is used to relate input with output
variables. For example, given the following structural model defined by three latent variables; i.e.: (1)
COGAD: cognitive valuations of certain ad, (2) ATTAVSING: overall opinion about advertising, (3)
ATTAD: overall opinion about certain ad (see figure 1).
COGADD
ATTAD
ATTAVSING
We use variables COGADD and ATTAVSING as inputs (antecedent) and ATTAD as output (consequent),
arising rules as the following:
IF COGADD is high and ATTAVSING is low THEN ATTAD is medium
Regarding to the membership functions, the numerical scale used in the questionnaire has been translated to a
fuzzy semantic as shows the following example (figure 2) for 3 linguistic terms and a 1 to 7 scale:
Small Medium Large
Membership degrees
1
0
1 2 3 4 5 6 7
Although some tuning methods have been proposed in the literature to automatically design the membership
functions [Karr (1991)], we think that it should be maintained invariable in order to get the maximum
interpretability, a very important issue in knowledge discovery.
• Learning method: One of the most interesting characteristics of our methodology is that any fuzzy rule
learning method for fuzzy rule-based systems can be used as optimization/design process. Anyway, since we
force the obtained fuzzy system to be very rigid due to the membership functions are not optimized, a
sophisticated fuzzy rule learning method should be used.
In this paper we propose to use a simple and quick method that has shown to obtain good results in fuzzy
modelling keeping a high interpretability of the derived fuzzy models: the method COR (cooperative rules)
[Casillas et al. (2002)]. It arises as an effort to exploit the accuracy ability of linguistic fuzzy rule based-
systems by exclusively focusing on the fuzzy rule base design. In this case, the fuzzy membership functions
and the fuzzy model structure keep invariable, thus resulting in the highest interpretability. The method has the
final goal of enhancing the interpolative reasoning the fuzzy rule-based system develops. This is one of the
most interesting features of fuzzy rule-based systems and plays a key role in their high performance, being a
consequence of the cooperative action of the linguistic fuzzy rules. To do that, the fuzzy rule base design is
made using global criteria that consider the action of the different rules jointly. It is attained by means of a
strong, smart reduction of the search space. The main advantages of the COR methodology are its capability
to include heuristic information, its flexibility to be used with different optimization or search techniques, and
its easy integration within other derivation processes.
BLFFOR
BLFSPE
GENATT TRUST
SHOP
BLFUPT
BLFPRI
On the other hand, results presented in Table 3 allows us to individually test the statistical significance of the different
relations set previously on the hypotheses formulation section, as well as the degree of variance of the endogenous
latent variables of the model (GENATT and TRUSTSHOP) explained by their antecedents (R2). Thus, the whole part
of the structural parameters in the model is significantly different to zero, so hypotheses have not been rejected, with the
exception of the non significant relation BLFUPT – GENATT (H1c). Furthermore, we verified that the exogenous
latent variables accounted for 40% of the variance in general attitudes towards Internet (GENATT), which, in turn,
were responsible for 20% of the variance referring to trust in Internet as a way to shop (TRUSHOP). We should also
note that two reverse items belonging to BLFSPE were transformed for the all three to have the same sense.
Table 3. Structural parameter estimates, (standard error) and t-value
DEPENDENT BLFFOR BLFSPE BLFUPT BLFPRI GENATT Error Var. R²
VARIABLE
0.61 0.16 -0.0038 -0.19 --- 0.54
GENATT (0.075) (0.077) (0.056) (0.046) --- (0.044) 0.40
8.14 2.07 -0.069 -4.13 --- 12.32
--- --- --- --- 0.36 0.45
TRUSHOP --- --- --- --- (0.042) (0.053) 0.20
--- --- --- --- 8.46 8.58
Therefore, this estimation technique has given to us information about both the significant relations and the degree of
intensity and sense of the relation. For example, as we hypothesized, BLFFOR and BLFSPE have a positive linear
relation with GENATT, being also BLFFOR the most influential factor in GENATT (parameter estimate of 0.61),
while BLFPRI presents a weaker negative relation with such an endogenous variable.
Each table also shows the mean square error (MSE) obtained by each fuzzy association rule set over the tackled data.
MSE is defined as:
1 N
MSE = ∑ ( F ( x1i ,K, xni ) − y i )2 ,
2 N i=1
with F ( x1i ,K, xni ) being the output obtained form the fuzzy system when the example data ei = ( x1i ,K, xni , y i ) is
used, and y i being the known desired output. The closer to zero the measure is, the higher the global performance and,
thus, the better the rule cooperation.
Table 4: System of rules for GENATT considering together the four antecedents
MSE = 0.416517
Rule BLFFOR BLFSPE BLFUPD BLFPRI GENATT
1 L L L S VL
2 L M L L VL
3 L M L M VL
4 L M L S VL
5 L M M M VL
6 L S L S VL
7 L S M L VL
8 L S M M L
9 L S M S VL
10 L S S S VL
11 M M L L L
12 M M L M L
13 M M L S VL
14 M M M M L
15 M M M S L
16 M M S S VL
17 M S L S M
18 M S M L M
19 M S M M L
20 M S M S VL
21 M S S M L
22 M S S S M
23 S M L L L
24 S S L M L
25 S S M M L
26 S S S L S
Table 5: System of rules for GENATT considering separately the four antecedents
MSE 0.556464 0.597823 0.588862 0.560186
Antecedent label BLFFOR-GENATT BLFSPE-GENATT BLFUPT-GENATT BLFPRI-GENATT
VL L L VL L
L VL VL L —
M L L L L
S L L L L
VS VS L L VL
In Table 7 several pairs of contradictory rules extracted from table 4 are showed. Furthermore, we reflect on both,
expected value that should have taken the consequent GENATT according to previous results, and consistency of rules
system obtained by our application analysing the behaviour of MSE associated with the model when modifying the
consequent label in function of such an expectation.
Table 7: Contradictory rules found
Model MSE
SETTING RULE BLFFOR BLFSPE BLFUPD BLFPRI GENATT MSE Variation
7 L S M L VL
Observed 0.416517
8 L S M M L
Expected 8 L S M M VL 0.422708 +0.006191
15 M M M S L
Observed 0.416517
16 M M S S VL
15 M M M S VL 0.453628 +0.037111
Expected
16 M M S S L 0.422980 +0.006463
21 M S S M L
Observed 0.416517
22 M S S S M
22 M S S S L 0.419026 +0.002509
Expected
22 M S S S VL 0.427267 +0.010750
The first contradiction comes from matching rule 7 and 8. At the extent that a negative relation has been detected
between BLFPRI and GENATT, it is expected that, ceteris paribus, a reduction in BLFPRI will cause an improvement
in GENATT or, as minimum, invariability in such a consequent if the degree of variation of BLFPRI is not enough to
cause changes. However, our application has observed such a paradoxical situation by rule 8. Under these
circumstances, we consider that if a modification in the consequent (expected rule), adjusting it to its expected value, is
able to reduce the MSE of the model, then it would be a sign of that modification of rules system done according to
what it was expected fits in better with observed patterns of data, otherwise initial solution would be more appropriate.
In this sense, it can be seen that RMS of the partial model increases with this modification. The same occurs for the
other contradictory pairs of rules found, where the MSE variation has depended on the degree of modification
considered for the consequent.
Furthermore, it can be conclude at this respect that those contradictory effects shown by those pairs of rules in
GENATT were just apparent, and that results are not due to a local optimum fall of the search algorithm, since better
accuracy degrees are obtained with the apparent contradictions. So, a plausible argument to explain this is that of free
this application presented here from responsibilities and make responsible for this to a deficient specification of the
theoretical marketing model. Due to that fact, other relevant factors not considered explicitly in our conceptual model,
which can also influence in GENATT, would be the answer to contradictions found in our system of rules.
Finally, with respect to TRUSTSHOP, the other endogenous latent variable of the proposed behavioural model, table 6
shows the results after applying our algorithm of approximation (Five linguistic terms have been used for both
variables). In this sense, initially considering that users of our sample do not really show a considerable trust in Internet
shopping, pattern of relation between both variables does not present a linear behaviour. Thus, GENATT seems to be
an influential factor in TRUSTSHOP as the former causes a gradually improvement over its consequent when overall
opinion towards Internet increases from low to intermediate positions. However, GENATT is not influential when
taking intermediate or higher values at the extent that TRUSTSHOP keeps a moderate place.
4.4. IMPLICATIONS OF THE PREVIOUS RESULTS BASED ON THE PROPOSED APPLICATION FOR
EXPLAINING THE ONLINE CONSUMER BEHAVIOR MODEL
Results from using our application allows to confirm that, in one hand, the different beliefs towards the Internet
contemplated here have a significant influence over user´s general opinion towards the Internet, though their influential
power and intensity differ depending on both the factor and the value taken by each of them. On the other hand, attitude
towards the Internet has also been a significant factor to determinate users´s trust in Internet shopping, though a similar
reflection could be made on its degree of influence according to its value taken.
Thus, opinion about formal aspects of the Internet and web sites is been the most relevant belief to discriminate values
taken by the overall attitude towards the Internet. In this respect, that can be seen more clearly when considering their
relation in an isolated way. Furthermore, high opinions about formal aspects usually assure really good attitudes
towards the Internet. That is to say, when users present the highest opinion about this factor, the rest of beliefs do not
seem to be really influential. So, online business should take care of this aspect at the extent that, making use of a
parallel though, it will have a considerable effect on the users´s affective response toward their website.
Notwithstanding, the degree of influence of this factor is not so clear when it has an intermediate or low presence.
With respect to the belief invasion of privacy when surfing the web, it neither shows a constant degree of influence
over general opinion towards the Internet. In this sense, it seems to be more influential when takes low or intermediate
values. So, a creation on an online environment able to highly respect the users´s privacy will also generate good
affective responses to certain website, at the extent that this issue has demonstrated to be more influential when being
good valuated.
Beliefs towards interaction speed, considering previously that our sample of users have not really showed good
perceptions about this due to that the whole part of individuals present an intermediate or bad valuation of this, have a
residual influence over the attitude towards the Internet. This fact is contradictory with ideas defended by Lin and Lu
(2000) when highlighting that this factor is the most for the development of the user´s beliefs toward certain a web site.
On the other hand, a similar influence has been found for the case of beliefs towards content updating of websites.
Thus, even when both previous factors do not determine in a transcendental manner formation of the overall opinion
towards the Internet, we have found certain cases, those when influence of factors formal aspects and invasion of
privacy do not have a influence so clear, where the one or the other can moderate the value of the dependent variable.
Finally, primarily considering that users´ population show certain reserves to get involved in Internet shopping yet, as it
can be deduced from observing the overall level of trust, overall attitude towards the Internet seems to be an important
variable to generate trust till certain point, and beyond there this variable does not discriminate such a trust. In this
respect, attitude toward the Internet does not improve trust in Internet shopping when taking the former values from the
intermediate to the highest level. So, the main problem should be focused on thinking about why the current situation
occurs and how to generate higher levels of consumer trust.
5. CONCLUDING REMARKS
As it was pointed out in the initial section related with the motivation of this work, a tool for estimating a proposed
consumer behaviour model should not be only useful to test a set of theoretical relationships which form such a model,
but it must be also able to be helpful for the marketing management function to have a good perspective of certain
marketing problem and to take the right decisions.
With that aim, we have empirically tested the theoretical relations of our model by making use of both SEM and FAS,
so several conclusions could be drawn when comparing them.
In one hand, when using SEM we obtain results that allow us to know how well the model fits to the data as well as
which are the significant relations. Furthermore, we get different coefficients (parameters) that give us information
about the sense and intensity of the relation.
However, after applying our methodological proposal we have confirmed some previous ideas. In this sense, when real
relations between variables are not linear, parameters resulted from SEM are deficient to really explain and give
information about how certain variable (antecedent) influences another (consequent), so it is not correct to consider that
a lineal coefficient of relation with the dependent variable keeps the same for all the values that the antecedent can take.
Therefore, our application allows the user’s model (e.g.: a marketing manager) to know which is the exact behaviour of
the relations among variables for certain market situation. In this sense, results from this application are pretty much
useful that SEM results for helping MkMSS, at the extent that it does not only allow to test the model proposed, but it is
also able to give deep qualitative information about the relations according to the degree of intensity of the antecedents.
On the other hand, SEM results can show, as it is the case for BLFUPT-GENATT in our example model, that a
relation is not statistically significant, while results from our application indicate that such a factor can be decisive in
moderating the consequent in certain cases. Thus, a lineal approximation to estimate a complex model of consumer
behaviour, though it is helpful, could be too much simplistic to get a true perspective of the relations.
Furthermore, a case for future treatment is that of why some factors are antecedents statistically significant and not
others by basing on SEM results, if when we take a look to the results from our application it can be seen that they
present a similar relational patter with the consequent. That could be due to the inherent philosophy of adjusting to the
data in which it is based SEM.
In conclusion, though a deeper work is necessary to purify the potentials of our application to the consumer marketing
modelling issue, basing on the previous results, it can be said that our application is able to give higher qualitative
information about how the relations among variables behave. Thus, in principle, it is pretty convenient to make use of
this kind of approximation to explain relations in a complex model of consumer behaviour as a complement to the
classic techniques based on SEM used till now.
References
Casillas, J., Cordón, O., Herrera, F. (2002), “COR: A methodology to improve ad hoc data-driven linguistic rule learning methods
by inducing cooperation among rules”, IEEE Transactions on Systems, Man, and Cybernetics—Part B: Cybernetics, vol.
32, pp. 526-537.
Casillas, J., Cordón, O., Herrera, F., Magdalena, L. (Eds) (2003a), Interpretability issues in fuzzy modelling, Springer, Heidelberg,
Germany.
Casillas, J., Cordón, O., Herrera, F., Magdalena, L. (Eds) (2003b), Accuracy improvements in linguistic fuzzy modelling, Springer,
Heidelberg, Germany.
Casillas, J., Cordón, O., Herrera, F. (2003c), “COR methodology: A simple way to obtain linguistic fuzzy models with good
interpretability and accuracy”, in Casillas et al. (2003b).
Chen, Q. and Wells, W., 1999. Attitude toward the site. Journal of Advertising Research (September/October), 27-37.
Cheskin Research and Studio Archetype/Sapient, 1999. Ecommerce trust study.
Crawford, F. and Shern, S., 1998. Six Winning Strategies for Web Selling. Retail news, Ernst & Young (Spring), 1-3 and 8.
Farquhar, B., Gordon, L., Balfour, A., 1998. Consumer needs in Global Electronic Commerce. International Journal of Electronic
Markets 8 (2), 9-12.
Freitas, A.A. (2002), Data mining and knowledge discovery with evolutionary algorithms, Springer, Heidelberg, Germany.
Gatignon, H. (2000), “Commentary on Peter Leeflang and Dick Wittink´s “Building models form marketing decisions: past, present
and future””, International Journal of Research in Marketing, vol. 17, pp. 209-214.
Helander, M., Khalid, H., 2000. Modelling the customer in electronic commerce. Applied Ergonomics 31, 609-619.
Hoffman, D., Novak, T., Peralta, M., 1998. Building consumer trust in online environments: the case for information privacy.
Communications of the ACM 42 (4), 80-85.
Jarvenpaa, S.L., Tractinsky, N., 1999. Consumer trust in an Internet store: a cross-cultural validation. Journal of Computer-Mediated
Communications 5 (2), http://www.ascusc.org/jcmc/vol5/issue2/jarvenpaa.html
Jevons, C., Gabbott, M., 2000. Trust, Brand equity and brand reality in Internet business relationships: An interdisciplinary approach.
Journal of Marketing Management 16, 619-634.
Karr, C.L. (1991), “Genetic algorithms for fuzzy controllers”, AI Expert, vol. 6, pp. 26-33.
Laurent, G. (2000), “Improving the external validity of marketing models: a plea for more qualitative input”, International Journal of
Research in Marketing, vol. 17, pp. 177-182.
Lin, J., Lu, H. (2000), “Towards an understanding of the behavioural intention to use a website”, International Journal of Information
Management 20, pp. 197-208.
Lohse, G., Spiller, P., 1999. Internet retail store design: How the user interface influences traffic and sales. Journal of Computer-
Mediated Communications 5 (2), http://www.ascusc.org/jcmc/vol5/issue2/lohse.htm
Roberts, J.H. (2000), “The intersection modelling potential and practice”, International Journal of Research in Marketing, vol. 17, pp.
127-134.
Schiffman, L., Kanuk, L., 1997. Consumer Behavior (Fifth Edition). Prentice Hall.
Solomon, M., 1997. Comportamiento del Consumidor (3ª Edición) (Original version: Consumer Behavior Buying, Having and
Being). Prentice Hall.
Steenkamp, J., Baumgartner, H. (2000) “On the use of structural equation models for marketing modeling”, International Journal of
Research in Marketing , vol. 17, pp. 195-202.
Urban, G., Sultan, F., Qualls, W., 2000. Placing trust at the center of your Internet strategy. Sloan Management Review (Fall), 39-48.
Van Bruggen, G.H., Wierenga, B. (2000), “Broadening the perspective on marketing decision models”, International Journal of
Research in Marketing, vol. 17, pp. 159-168.
Witten, I.H., Frank, E. (2000), Data mining: practical machine learning tools and techniques with Java implementations, Morgan
Kaufmann Publishers, San Francisco, CA, USA.