A Dominance-based Rough Set Approach to customer behavior in the airline market

(1)

A Dominance-based Rough Set Approach to customer behavior in the

airline market

James J.H. Liou

a

, Gwo-Hshiung Tzeng

b,c,d,*

a

Department of Air Transportation, Kainan University, No. 1 Kainan Road, Luchu, Taoyuan County 338, Taiwan b

Graduate Institute of Project Management, Kainan University, No. 1 Kainan Road, Luchu, Taoyuan County 338, Taiwan c

Department of Business and Entrepreneurial Management, Kainan University, No. 1, Kainan Road, Luchu, Taoyuan County 338, Taiwan d

Institute of Management of Technology, Chiao Tung University, Ta-Hsuch Road, Hsinchu 300, Taiwan

a r t i c l e

i n f o

Article history: Received 9 May 2008

Received in revised form 6 September 2009 Accepted 23 January 2010 Keywords: Dominance Rough set Customer behavior Data mining Marketing strategy

a b s t r a c t

Market segmentation is a crucial activity in the present business environment. Data mining is a useful tool for identifying customer behavior patterns in large amounts of data. This information can then be used to help with decision-making in areas such as the airline market. In this study, we use the Dominance-based Rough Set Approach (DRSA) to provide a set of rules for determining customer attitudes and loyalties, which can help managers develop strategies to acquire new customers and retain highly valued ones. A set of rules is derived from a large sample of international airline customers, and its predictive ability is evaluated. The results, as compared with those of multiple discriminate analyses, are very encouraging. They prove the usefulness of the proposed method in predicting the behavior of airline customers. This study demonstrates that the DRSA model helps to iden-tify customers, determine their characteristics, and facilitate the development of a market-ing strategy.

1. Introduction

In the face of a highly competitive and fast-changing airline market, managers must not only provide high-quality service but also react appropriately to changes in customer needs. However, it would be helpful if, instead of targeting all customers equally or offering the same incentives to all customers, enterprises could target only those customers who meet certain proﬁtability criteria based on their individual needs or purchasing behaviors[5]. Customer behavior is the result of complex interactions between a number of factors, which can include the level of marketing activity, the competitiveness of the envi-ronment, brand perception, the inﬂuence of new technologies, and individual needs[34]. The characteristics and behaviors of airline customers are even more complex, and customer perception and behavior are affected by many factors, such as safety, service, technology, environment, price and many others. Hence, it is crucial that management determine the most important factors that affect the attitude and loyalty of airline customers. In the past, researchers have generally made use of statistical surveys to determine customer behavior. In such surveys, natural language or linguistic variables (e.g., ‘‘although the airline service is satisfactory, the price of the product being offered is high, the individual’s decision is not to purchase”) are used to describe customer patterns. Unfortunately, this can create an environment of imprecision, uncertainty, and par-tiality with regard to knowledge. These linguistic variables are then transformed into quantitative values, after which factor, cluster, and discriminant analyses are conducted. However, the semantic imprecision of natural languages leads to problems

* Corresponding author. Address: Department of Business Administration, Kainan University, No. 1, Kainan Road, Luchu, Taoyuan County 338, Taiwan. Tel.: +886 3 3412430; fax: +886 3 3412456.

E-mail addresses:ghtzeng@cc.nctu.edu.tw,ghtzeng@mail.knu.edu.tw(G.-H. Tzeng).

Contents lists available atScienceDirect

Information Sciences

(2)

of computation, especially when the information described in a natural language is beyond the reach of existing bivalent logic and probability theory techniques[41].

Recently, data mining techniques have been adopted to predict customer behavior[6,30]. Data mining is one stage in Knowledge Discovery in Databases (KDD), involving the application of specific algorithms for pattern extraction[21]. Mar-keting managers can develop strategies to attract new customers and retain highly valued ones based on this mined knowl-edge. The Dominance-based Rough Set Approach (DRSA), originally developed by Greco et al. [8,9], is a relatively new approach in data mining that is very useful for data reduction in qualitative analysis. The rough set theory, a kind of natural language computation, is particularly useful for dealing with imprecise or vague concepts[23]. Basically, natural language computation is a system in which the objects of computation are simply predicates and propositions drawn from a natural language. A set of decision rules is generated by applying the rough set approach to analyze the classification data. These decision rules are in the form of logic statements of the type ‘‘if conditions, then decision”. The set of decision rules represents a preference model for the decision-maker that is expressed in a natural and understandable language. According to Zhu et al.[40], the rough set method does not require additional information about the data; it can work with imprecise values or uncertain data, is capable of discovering important facts hidden in that data, and has the capacity to express them in nat-ural language. The rough set theory has been successfully applied in a variety of fields, including medical diagnosis, engineer-ing reliability, expert systems, empirical studies of material data[15], evaluation of bankruptcy risk[29], machine diagnosis

[39], business failure prediction[1,4], network intrusion detection[40], travel demand analysis[7], mining stock price[35], the insurance market[28], and accident prevention[36].

Although the Classical Rough Set Approach (CRSA) is a powerful tool for handling many problems, it is not able to deal with inconsistencies originating from the criteria, e.g., attributes with preference-ordered domains (scale) like product qual-ity, market share, and debt ratio[10]. However, the DRSA has an advantage over the CRSA in that it has access to an infor-mation table that displays comprehensive dominance relations. It is able to deal with inconsistencies where decisive classes are not consistent with their criteria. The aim of this study is to mine data regarding airline customer behavior using the DRSA. The derived knowledge can help airlines identify valuable customers, predict future behavior, and enable ﬁrms to make proactive, knowledge-driven decisions.

2. Customer behavior

In most research on customer behavior, customer demographic variables are applied to analyze customer behavior[30]. However, Rayport and Sviokla[27]suggest that a customer’s perception of product or service value is comprised of three basic elements: the product or service that a company offers, the context in which a company offers this product or service, and the infrastructure that enables the transaction to take place. In the traditional marketplace, content, context, and infra-structure are bundled together, usually offered by the same company; however, due to new technologies and new selling patterns, it is easy to separate these three elements in the air transport market. The importance of valuable customer behav-ioral variables—recency, frequency, and monetary (RFM)—has been extensively studied[20,31,32]. Researchers have ob-served that RFM variables are not only useful for the analysis of customer behavior but can also be effectively used to investigate customer value and niche markets. Hsieh[14]proposed a method that integrated data mining and behavioral scoring models for the management of banking customers. He divided customers into three groups according to their shared behaviors, characteristics and profitability. Marketers infer the profiles of each group of customers and propose management strategies appropriate to the characteristics of each group. Chen et al.[3]integrated customer behavioral variables, demo-graphic variables, and transaction databases to establish a method of mining changes in customer behavior in the retail mar-ket. In their study, customer behavior patterns are first identified using association rule mining. After the association rules for customer behavior are discovered, changes in customer behavior are identified by comparing two sets of association rules generated from two datasets from different periods. The changes in patterns thus identified can then be investigated and assessed to provide a basis for formulating marketing strategies. Customer behavior analysis in Internet marketing has al-ready been investigated by many researchers[16,18,19]. In most of such studies, data mining technologies are applied to produce a generalized customer profile of the Internet shopper and to further explore the Web usage pattern of the online consumer. The knowledge obtained through data mining helps foster informed Internet marketing decision-making and al-lows for the refinement of Web content and infrastructure to improve Internet marketing[18]. Wang and Hong[34], through the use of data mining techniques, developed a Customer Profitability Management (CPM) system for achieving marketing goals by leading customers to progress along pre-determined and desirable paths. Their system emphasizes continuous interplay between active and reactive monitoring procedures, from which any shift in customer behavior can be identified. Just as in the conventional marketplace, airlines need to build customer loyalty as well as attract new customers. How-ever, the unique characteristics of air transportation have altered the rules in the airline market. In a traditional business environment, where the seller meets a buyer in person, he can understand his behavior (intention to purchase, choice of product, etc.) from facial expressions, body language and verbal communication. The salesperson accumulates this knowl-edge while dealing face to face with the customer, and then uses this knowlknowl-edge to increase the customer’s satisfaction

[16]. However, the average airline customer books his/her ticket via the Internet or through a travel agency. His/her decision may be based on previous experience with that airline’s service, word-of mouth, the airline’s safety record, convenience, and so on. Also, the products offered by an airline are not physical objects; rather they are performance and reliability. Therefore,

(3)

the behavior of airline customers is different from that of customers in a traditional market or even Internet shoppers. There-fore, in this study we hope to ﬁll in this gap: to extend the knowledge of customer behavior through data mining. The infor-mation thus obtained can be used for better decision-making in the airline market.

3. Basic concepts of the Dominance Rough Set Approach

The rough set theory, ﬁrstly introduced by Pawlak[22]in 1982, is a valuable mathematical tool for dealing with vague-ness and uncertainty[23]. For a long time, the use of the rough set approach and other data mining techniques was restricted to classiﬁcation problems where the preference order of the evaluations was not considered. This is due to the fact that this method cannot handle inconsistencies that occur as a result of the violation of the dominance principle[10]. In order to deal with this kind of inconsistency, it was necessary to make a number of methodological changes to the original rough set the-ory. Greco et al.[8]proposed an extension of the rough set theory based on the dominance principle that would permit it to deal with inconsistency. This method is mainly based on the substitution of the indiscernibility relation for a dominance relation in the rough approximation of decision classes. It is more general than the classic functional or relational model and is more understandable for users because of its natural syntax[10]. The basic concepts of DRSA are described in the fol-lowing[2,8–10,12,13,17,22–26,33,37,38].

3.1. Data table

For algorithmic reasons, the information regarding the objects is supplied in the form of a data table whose separate rows refer to distinct objects (actions) and whose columns refer to the different attributes or criteria (attributes with preference-ordered domains) considered. Each cell in this table indicates an evaluation (quantitative or qualitative) of the object placed in that row by means of the attribute/criterion in the corresponding column.

Formally, a data table is in the form of a 4-tuple information system IS = (U, Q, V, f), where U is a ﬁnite set of objects (uni-verse), Q = {q1, q2, . . . , qm} is a ﬁnite set of attributes/criteria, Vqis the domain of the attribute/criterion q, V =Sq2QVqand f:

U Q ? V is a total function such that f(x, q) 2 Vqfor each q 2 Q; x 2 U, called the information function. The set Q is usually

divided into set C of condition attributes and set D of decision attributes. 3.2. Rough approximation by means of the dominance relations

Let qbe an outranking relation to U with reference to criterion q 2 Q, such that x qy means that ‘‘x is at least as good as

y with respect to criterion q”. Suppose that qis a complete preorder, i.e., a strongly complete (such that for each x, y 2 U, at

least one of x qy and y qx is veriﬁed and thus x and y are always comparable with respect to criterion q) and transitive

binary relation. Moreover, let Cl = {Clt, t 2 T}, T = {1, . . . , n} be a set of classes of U such that each x 2 U belongs to one and only

one class Clt2 Cl. We assume that all r, s 2 T, such that r > s, and each element of Clris preferred to each element Cls. In other

words, if is a comprehensive outranking relation on U, then it is supposed that

½x 2 Clr;y 2 Cls;r > s ) x y; where x y means x y and not y x.

We can deﬁne unions of classes relative to a particular dominated or dominating class; these unions of classes are called upward and downward unions of classes and are deﬁned, respectively, as

ClP t ¼ [ sPt Cls; Cl 6 t ¼ [ s6t Cls:

It is said that object x P-dominates object y with respect to P # C (denotation xDpy); if x qy for all q 2 P, and DP= \q2pq,

then the dominance relation DPis a partial preorder. Given P # C and x 2 U, let Dþ

PðxÞ ¼ fy 2 U : yDPxg;

D

PðxÞ ¼ fy 2 U : xDPyg;

represent the P-dominating set and the P-dominated set with respect to x, respectively.

The P-lower and P-upper approximations of ClPt;t 2 f1; . . . ; ng, with respect to P # C (denotation PðCl P

t Þ and PðCl P t Þ,

respectively), are deﬁned as

PðClP t Þ ¼ fx 2 U : D þ PðxÞ # Cl P t g; PðClPt Þ ¼ [ x2ClP t Dþ PðxÞ ¼ fx 2 U : D PðxÞ \ Cl P t – £g:

The P-lower approximation of an upward union ClPt ;PðCl P

tÞ, is composed of all objects x from the universe such that all

ob-jects y, having at least the same evaluations on all of the considered criteria from P, also belong to class Cltor better. Thus, one

can say that if an object y has at least as good an evaluation on the criteria from P as object x belonging to PðClP

tÞ, then

cer-tainly, y belongs to class Cltor better. The P-upper approximation of an upward union Cl P t ;PðCl

P

(4)

from the universe whose evaluations on the criteria from P are not worse than the evaluation of at least one other object y belonging to Cltor better. Analogously, the P-lower and P-upper approximations of Cl

6

t, t 2 {1, . . . , n}, with respect to P # C

(denotation PðCl6

tÞ and PðCl 6

tÞ, respectively) are deﬁned as PðCl6 tÞ ¼ fx 2 U : D PðxÞ # Cl 6 tg; PðCl6 tÞ ¼ [ x2Cl6 t DPðxÞ ¼ fx 2 U : D þ PðxÞ \ Cl 6 t – £g:

The P-boundaries (P-doubtable region) of ClP t and Cl 6 t are deﬁned as BnPðClPt Þ ¼ PðCl P t Þ PðCl P t Þ; BnPðCl 6 tÞ ¼ PðCl 6 tÞ PðCl 6 tÞ:

We deﬁne the accuracy of approximation of ClPt and Cl 6

t for all t 2 {1, . . . , n} and for any P # C, respectively, as

a

pðClPt Þ ¼ jPðClPt Þj jPðClPt Þj ;

a

pðCl 6 tÞ ¼ jPðCl6 tÞj jPðCl6 tÞj : The ratio

c

pðClÞ ¼ jU S t2f2;...;ng BnpðClPt Þ ! j jUj ¼ jU S t2f1;...;n1g BnpðCl 6 tÞ ! j jUj

defines the quality of approximation of the classification Cl by means of the criteria from set P # C, or, briefly, the quality of classification where jj means the cardinality of a set. This ratio expresses the proportion of all P-correctly classified objects— i.e., all of the non-ambiguous objects to all of the objects in the data table. Every minimal subset P # C such that

c

P(Cl) =

c

C(Cl) is called a reduct of C with respect to Cl and is denoted by REDCl(P). Again, a data table may have more than

one reduct. The intersection of all of the reducts is known as the core, denoted by CORECl.

3.3. Decision rules

The end result of the DRSA is a representation of the information contained in the considered data table in terms of simple ‘‘if. . ., then. . .” decision rules. For a given upward union of classes ClPt , the decision rules that are induced under the

hypoth-esis that actions belonging to PðClP

t Þ are positive (while the others are negative) suggest an assignment to ‘‘at least class Clt”.

Analogously, for a given downward union Cl6

s, the rules induced under a hypothesis that actions belonging to PðCl 6 sÞ are

po-sitive and that all others are negative suggest an assignment to ‘‘at most class Cls”. On the other hand, the decision rules

in-duced under a hypothesis that actions belonging to the intersection PðCl6 sÞ \ PðCl

P

t Þ are positive, and that all of the others are

negative, suggest an assignment to some class between Clsand Clt(s < t).

The following three types of decision rules can be considered: 1. DP: decision rules that have the following form:

if f ðx; q1Þ P rq1and f ðx; q2Þ P rq2and . . . f ðx; qpÞ P rqp; then x 2 ClPt :

These rules are supported only by objects from P-lower approximations of the upward unions of classes ClP_t . 2. D6: decision rules that have the following form:

if f ðx; q1Þ 6 rq1and f ðx; q2Þ 6 rq2and f ðx; qpÞ 6 rqp; then x 2 Cl 6 t:

These rules are supported only by objects from the P-lower approximation of the downward unions of classes Cl6 t.

3. DP6: decision rules that have the following form:

if f ðx; q1Þ P rq1 and f ðx; q2Þ P rq2 and . . . f ðx; qkÞ P rqkand f ðx; qkþ1Þ 6 rqkþ1and . . . f ðx; qpÞ 6 rqp; then x

2 Cls

[ Clsþ1

[

[Clt:

These rules are supported only by objects from the P-boundaries of the unions of classes Cl6 s and Cl

P

t, where P = {q1,

q2, . . . , qp} # C, (rq1, rq2, . . . , rqp) 2 Vq1 Vq2 Vqpand t 2 {1, . . . , n}.

The algorithm for induction regarding decision rules is obtained from Greco et al.[11]. It is also noted that the set of deci-sion rules induced from the approximations deﬁned using dominance relations gives a more synthetic representation of knowledge as contained in the decision table than the set of rules induced from the CRSA, which uses indiscernible relations. This is due to the more general syntax of the rules (‘‘P” and ‘‘6” are used instead of ‘‘=”).

(5)

4. Empirical study: a case of customer behaviors in the airline market

In order to demonstrate the effectiveness of the DRSA and our proposed approach, we carried out an empirical study that is described in this section. We produced a questionnaire about customer behavior in the airline market with single and mul-tiple-choice answers and used DRSA to explore the classiﬁcation problem. The results should provide airlines with useful information to help them develop marketing strategies and achieve their marketing goals.

4.1. Preparation of the information table for customer behavior in the airline market

Since the products of airlines are their performance and reliability rather than physical objects, their customer behavior is different from that in the conventional marketplace. Therefore, the first thing to do was to define the attributes/criteria for the information table. After consulting with senior management from travel agencies and airline marketing personnel, we listed 10 criteria to be covered by the first questionnaire. Ninety-two passengers at the Taoyuan International Airport in Tai-wan agreed to rate the importance of the decision criteria that affected their purchasing behavior using a five-point Likert-type scale that ranged from 5 (extremely important) to 1 (no effect). Afterwards, the four highest-scoring criteria were ex-tracted and used to construct information systems for airline market mining. Since demographic variables are still important references for airline customer behavior, they were considered as other attributes to be evaluated. Based on the above eval-uation process, the results of the first questionnaire indicated that safety, price, quality of service and schedule are the most important criteria that influence the buying behavior of airline customers. Therefore, the second questionnaire contained two parts: (1) demographic information about the participants and (2) criteria for the surveyed airlines. The questionnaire covered eight international airlines serving Taiwan. The attributes/criteria for customer behavior in the airline market are shown inTable 1. The preference regarding schedule, quality of service, safety and buying decisions were set as gains, while the price was set as cost (the lower the better). The personal profile attributes were all set to ‘‘no preference”. Each respon-dent was asked to rate his/her level of satisfaction for each criterion and his/her buying decision for the specified airline. 451 respondents completed the questionnaire. The demographics of the 451 respondents are shown inTable 2.

4.2. Results of the DRSA analysis

The results of the DRSA analysis consisted of four parts: quality of approximation, rule generation, rule validation and the signiﬁcance of condition attributes/criteria.

4.2.1. Quality of approximation

The accuracy of approximation for the three decision classes is shown inTable 3. The results indicate good accuracy for the different classes. In general, high values for the quality of classiﬁcation and accuracy mean that the attributes/criteria selected are an approximation of the classiﬁcation. The ‘‘at most 1” class is the ‘‘do not consider buying” class. There are 27 objects belonging to that class. The accuracy of approximation for ‘‘at most 1” is one. The ‘‘at most 2” class includes the ‘‘do not consider buying” and ‘‘maybe” classes, for which accuracy reaches 0.957. On the other hand, the ‘‘at least 3” class refers to the ‘‘will surely purchase” class, and its lower and upper approximations are 173 and 185, respectively. The accu-racy of the ‘‘at least 3” class is 0.935. The ‘‘at least 2” class is composed of the ‘‘maybe” and ‘‘will surely purchase” classes, and its accuracy is raised to one. The overall quality of approximation is calculated as follows: (451 12)/(451) = 0.973. 4.2.2. Rule generation

We establish a set of rules, the ‘‘minimum cover rules” (i.e., where the set does not contain any redundant rules), and these rules are certain, such that there are a total of 25 rules generated from the data.Table 4shows the minimum cover rules obtained after eliminating those with a ‘‘cover strength” (i.e., number of objects, in this case passengers, covered by

Table 1

Speciﬁcations of attributes/criteria related to customer behavior.

Attribute/criterion name Attribute/criterion value Value set Preference

Condition attributes

Gender Female; male {F, M} None

Age Below 30; 30–40; 41 and above {1, 2, 3} None

Occupation Government employee; employee of company; student; others {G, P, S, O} None Education High school and below; college; masters and above {H, C, M} None Income NT$40,000 and below; 40,001–90,000; 90,000 and above {1, 2, 3} None

Service quality Poor; medium; satisfactory {1, 2, 3} Gain

Schedule Bad; good {1, 2} Gain

Safety record Poor; medium; satisfactory {1, 2, 3} Gain

Price Low; medium; high {1, 2, 3} Cost

Decision criterion

(6)

the rule) of less than 10. The reduced rule set contains 18 rules, with two rules corresponding to class 1, nine rules to at most class 2, ﬁve rules to class 3, and two rules to at least class 2. We can see fromTable 4that if the price is high, the decision is ‘‘will not consider buying” and its cover strength is 126 (rule 2). This means that nearly one-fourth of the respondents will not consider purchasing from a speciﬁc airline if its price is high. Rule 5 suggests that the customer’s decision will be at most class 2 (maybe) when safety is rated as medium or less, the customer’s age is between 30 and 40, and the service quality is rated as medium or less. For students (rule 8), their decisions are at most class 2 when safety and service quality are equal to or less than medium. From rule 11, we can see that if the respondents are female, they are between 30 and 40, and the safety

Table 2

Background of the respondents.

Distribution Sample number Frequency (%)

Gender Female 244 54.1 Male 207 45.9 Age Less than 30 197 43.7 30–40 156 34.6 40 and above 98 21.7 Occupation Government employee 66 14.6 Employee of company 279 61.9 Student 52 11.5 Others 54 12.0 Education

High school and below 65 14.5

College 332 73.6

Masters and above 54 11.9

Income (NT)

Less than 40,000 225 49.9

40,000–90,000 204 45.2

90,000 and above 22 4.9

Table 3

Accuracy of classiﬁcation for customer behavior.

At most 1 At most 2 At least 2 At least 3

Lower approx. 27 266 424 173

Upper approx. 27 278 424 185

Boundary 0 12 0 12

Accuracy 1 0.957 1 0.935

Table 4

Minimum cover rules with a strength greater than 10.

No. Conditions Decision Strength

1 (Safety 6 1) & (Price P 2) D 6 1 27

2 (Price P 3) D 6 1 126

3 (Safety 6 2) & (Occupation = O) D 6 2 32

4 (Safety 6 2) & (Occupation = G) D 6 2 27

5 (Safety 6 2) & (Age = 2) & (Service 6 2) D 6 2 76

6 (Safety 6 1) D 6 2 38

7 (Safety 6 2) & (Education = H) & (Age = 3) D 6 2 18

8 (Occupation = S) & (Safety 6 6 2) & (Service 6 2) D 6 2 19

9 (Age = 3) & (Education = C) & (Service 6 2) D 6 2 23

10 (Safety 6 2) & (Price P 2) D 6 2 131

11 (Age = 2) & (Safety 6 2) & (Sex = F) D 6 2 53

12 (Safety P 3) & (Price 6 2) D P 3 164

13 (Education = M) & (Price 6 1) & (Service P 3) D P 3 13

14 (Price 6 2) & (Service P 3) & (Age = 3) D P 3 31

15 (Price 6 1) & (Service P 3) & (Occupation = P) & (Age = 1) D P 3 16 16 (Age = 2) & (Service P 3) & (Income = 2) & (Price 6 1) D P 3 25

17 (Safety P 2) D P 2 413

(7)

is rated medium or less, then their decision will be at most class 2, and the cover strength is 53. Rule 12 indicates that if the airline’s safety is satisfactory and the price is rated as medium or less, customers will surely purchase and the cover strength reaches 164. If the passenger’s income is between NT$ 40,000–90,000, his/her age is between 30 and 40 (which is considered middle age in this society), safety is rated satisfactory and the price is low, then he/she will deﬁnitely be a loyal customer. If safety is greater than or equal to medium and the price is low, then the respondent’s decision will be at least class 2, and the cover strength numbers are 413 and 202, respectively. The results indicate that a good safety record is attractive to 90% of the customers, while low price is attractive to about 50%. The set of minimum cover rules indicates that for those who are under 30 years of age, whether employed with private companies or part of student groups, price and service are more important. For higher-income, more highly educated or government-employee female respondents, safety is important. It is also worth noting that of the condition criteria, the schedule is the one that seems to have little effect. This may be because there were so few higher-income respondents in the surveyed population. The higher-income group is usually more con-cerned with time, but in our demographic data, there were only 22 respondents in the high-income group.

4.2.3. Rule validation

We check the feasibility of the decision rules generated in this study through the application of the 10-fold cross-valida-tion technique. First, from a randomly chosen 90% of the data, we generate decision rules. The remaining 10% of the data are used to validate the hit rate of the generated decision rules—i.e., the percentage of correct predictions for each class. This procedure is repeated 10 times; the hit rate is shown inTable 5.

As shown inTable 5, the overall classiﬁcation error is only 4.4%, with 431 objects decided correctly and only 19 objects decided incorrectly. The nine ambiguous objects indicate that the object classiﬁcation occurs 10 times during the validation processes. This ambiguity may be due to the hesitation of the respondents. The effectiveness of the DRSA is shown by the results of discriminant analysis inTable 5. The hit rate has increased from 91.8% with discriminant analysis to 95.6% for DRSA.

4.2.4. Signiﬁcance of the condition attributes/criteria

Since there is no reduct generated from the information system, all of the condition attributes/criteria are essentially uti-lized to distinguish the classes. However, the signiﬁcance of the condition attributes/criteria can be measured by their pres-ence in the derived rules. When a condition attribute/criterion shows up more frequently in the cover rules, and the strength of these rules is higher, this is more likely to be a key factor in customer decisions. The frequency with which the attributes/ criteria and cover strength appear in the derived minimum cover rules (Table 4)—number of objects matching the rule—can be used as an indication of the importance of the attributes/criteria. Rules 17 and 18 indicate that safety and price are the most signiﬁcant criteria for customer decisions, with cover strengths of 413 and 202, respectively. The implication is that most of the respondents make purchase decisions mainly based on the airline’s safety record and ticket price.

Another approach to exploring the importance of attributes/criteria is to set the strength level higher than a certain threshold value. Each rule in the derived rule set is thus ensured to be above a certain percentage in terms of cover strength. If we choose to generate all possible rules where the relative strength of each class is greater than 40%, then the total number

Table 5

Hit rates for DRSA and discriminant analysis.

Class 1 Class 2 Class 3

Class 1 25(27) 0(0) 0(0)

Class 2 0(11) 237(219) 3(16)

Class 3 0(0) 10(14) 167(164)

Ambiguous decision 9(0)

Overall classiﬁcation error 0.044(0.092)

Correct hit rate 95.6%(91.8%)

*Parenthesis () indicates the results of discriminant analysis.

Table 6

Possible rules for strengths greater than 40% for each class.

No. Condition Decision Strength

1 (Service 6 1) & (Price P 2) D 6 1 10

2 (Safety 6 1) & (Price P 2) D 6 1 27

3 (Price P 3) D 6 2 126

4 (Safety 6 2) & (Price P 2) D 6 2 131

5 (Education = C) & (Service 6 2) & (Safety 6 2) D 6 2 152

6 (Safety P 2) D P 2 413

7 (Price 6 1) D P 2 202

8 (Service P 3) & (Price 6 2) D P 2 172

(8)

of rules will be reduced to nine (Table 6). In the resulting set of rules, safety and price are still the most crucial condition criteria. It is clear that the higher the satisfaction level in terms of safety, the higher the desire to purchase will be, and that a low price is still a good way to attract people’s attention.

5. Discussion

Using the same data set and dummy-variable technique, we also apply discriminant analysis. The results are shown in

Table 5. Clearly the DRSA shows better prediction ability than does discriminant analysis. The hit rate increases from 91.8% for discriminant analysis to 95.6% for DRSA. Furthermore, in contrast to discriminant analysis, the rough set theory requires no underlying statistical assumptions. In particular, the DSRA can handle both attributes/criteria with and without preference order. In this study, we use this advantage to analyze survey data (whether with quantitative or qualitative attri-butes, and with or without preference order) and to better understand customer behavior in the airline market. Although rough set theory has been successfully applied to a variety of areas, it is still not often applied in the customer behavior ﬁeld. In this study, we also consider inconsistencies within criteria. The empirical results show that it is appropriate to apply DSRA to the mining of knowledge regarding customer behavior for the air transport market. Although the selected attributes/cri-teria may be strongly inﬂuenced by the local environment and culture, the proposed factor structure can be easily adapted and extended to the conditions and culture in other markets in any particular environment.

A condition attribute/criterion can be dropped if its removal will have no impact on the quality of the approximation. However, in our empirical study, no reduct or core is generated. This means that all of the selected attributes/criteria are important factors for the quality of the approximation. This result may be partly due to the fact that all criteria were derived from the first questionnaire, which included only important criteria.Table 3shows that the accuracy of the classification for customer behavior is as high as 0.973, which indicates the narrow boundary between the lower approximation and the upper approximation. Thus, the granules derived from the refined data can be understood to properly represent customer behavior. The high accuracy rate also shows that the objects in that class have higher dependency among all condition attri-butes/criteria for those objects. The total number of minimum cover rules is 25, but there are seven rules whose strength is less than 10. This indicates that most of the data derived from this study can be classified into 18 rules with only a few un-ique data. In contrast to the classical rough set theory, which tends to generate too many rules with only a little cover strength, with the DSRA we seem to be able to obtain better decision rules with stronger cover strength. This could be due to the fact that the DSRA considers both attributes/criteria and their inconsistencies. When there are too many unique rules generated, it is difficult to understand the contents of the data sets that relate to the decision rules in the condition part. In this study of customer behavior in the air transport market in Taiwan, we find that two criteria dominate customer decision-making: safety and price. Although price and quality of service are usually of major concern in the air transport market, the airline’s safety record is the most critical factor affecting customer decision-making. This is not difficult to under-stand, as the less than desirable safety record of Asian airlines in general has left most air travelers in Taiwan worried about safety. Therefore, the key to a successful strategy is to rebuild the confidence of air transport customers. The decision rules indicate that if the airline performs well in terms of safety, the customer will consider that airline even if ticket price and quality of service are less competitive. The schedule has little influence on the customer’s decision-making. In order to isolate representative rules and an effective factor structure, we conducted a 10-fold cross-validation and possible rule generation. The results from both processes show the validity of the decision rules. There is very good agreement, with only a 4.36% over-all classification error. The estimation results show that the accuracy of the approximation, the quality of the approximation, and the hit rates are satisfactory. The category of condition attributes/criteria for targeted customers can be increased in the future to help the decision-maker make a more precise judgment. More categories of condition attributes/criteria generate more refined decision rules that can improve the quality of the decision-maker’s strategy.

6. Conclusions

This study illustrates the usefulness of the DRSA approach as an operational tool for the prediction of customer behavior in the air transport market. The proposed prediction model takes the form of decision rules. Since the derived rules are sup-ported by real examples, they describe only the most relevant attributes/criteria. The classical rough set theory handles attri-butes without preferences, a technique that does not always accurately represent the real world. The DRSA is constructed by extending the classical rough set theory to include qualitative reasoning for preference-based customer behavior analysis. This is done by replacing the indiscernibility relation with the dominance relation. In this way we can represent the conﬂict-ing preference relations that affect customer behavior more objectively without introducconﬂict-ing the equivalence class concept of the classical rough set theory. Compared with those of the traditional statistical method, the results indicate that the DRSA has better prediction ability. Moreover, the derived decision rules are in natural language form, which makes their meaning easier to understand than with traditional methods.

Acknowledgements

(9)

References

[1] M.J. Beynon, M.J. Peel, Variable precision rough set theory and data discretisation: an application to corporate failure prediction, Omega 29 (2001) 561– 576.

[2] J. Blaszczynski, S. Greco, R. Slowinski, Multi-criteria classiﬁcation – a new scheme for application of dominance-based decision rules, European Journal of Operational Research 181 (2007) 1030–1044.

[3] M.C. Chen, A.L. Chiu, H.W. Chang, Mining changes in customer behavior in retail marketing, Expert Systems with Applications 28 (2005) 773–781. [4] A.I. Dimitras, R. Slowinski, R. Susmaga, C. Zopounidis, Business failure prediction using rough sets, European Journal of Operational Research 114

(1999) 263–280.

[5] J. Dyche, J. Dych, The CRM Handbook: A Business Guide to Customer Relationship Management, Addison-Wesley, MA, 2001.

[6] P. Giudici, G. Passerone, Data mining of association structures to model customer behavior, Computational Statistics and Data Analysis 38 (2002) 533– 541.

[7] C. Goh, R. Law, Incorporating the rough sets theory into travel demand analysis, Tourism Management 24 (2003) 511–517.

[8] S. Greco, B. Matarazzo, R. Slowinski, A new rough set approach to evaluation of bankruptcy risk, in: C. Zopounidis (Ed.), Rough Fuzzy and Fuzzy Rough Sets, Kluwer, Dordrecht, 1998, pp. 121–136.

[9] S. Greco, B. Matarazzo, R. Slowinski, Extension of the rough set approach to multicriteria decision support, INFOR 38 (2000) 161–195.

[10] S. Greco, B. Matarazzo, R. Slowinski, Rough sets theory for multicriteria decision analysis, European Journal of Operational Research 129 (2001) 1–47. [11] S. Greco, B. Matarazzo, R. Slowinski, J. Stefanowski, An algorithm for induction of decision rules consistent with dominance principle, in: W. Ziarko, Y.

Yao (Eds.), Rough Sets and Current Trends in Computing, LNAI 2005, Springer-Verlag, Berlin, 2001, pp. 304–313.

[12] S. Greco, B. Matarazzo, R. Slowinski, Rough sets methodology for sorting problems in presence of multiple attributes and criteria, European Journal of Operational Research 138 (2002) 247–259.

[13] S. Greco, B. Matarazzo, R. Slowinski, Rough set analysis of preference-ordered data, in: J.J. Alpigini, J.F. Peters, A. Skowron, N. Zhong (Eds.), Rough Sets and Current Trends in Computing, Springer-Verlag, Berlin, 2002, pp. 44–59.

[14] N.C. Hsieh, An integrated data mining and behavioral scoring model for analyzing bank customers, Expert Systems with Applications 27 (2004) 623– 633.

[15] A.G. Jackson, S.R. Leclair, M.C. Ohmer, W. Ziarko, H. Al-kamhwi, Rough sets applied to material data, ACTA Material 44 (1996) 4475–4484. [16] M. Jenamani, P. Mohapatra, S. Ghose, A stochastic model of e-customer behavior, Electronic Commerce Research and Applications 2 (2003) 81–94. [17] W. Kotlowski, K. Dembczynski, S. Greco, R. Slowinski, Stochastic dominance-based rough set model for ordinal classiﬁcation, Information Sciences 178

(2008) 4019–4037.

[18] I. Kwan, J. Fong, H.K. Wong, An e-customer behavior model with online analytical mining for internet marketing planning, Decision Support Systems 41 (2005) 189–204.

[19] H. Lu, C.C. Lin, Predicting customer behavior in the market-space: a study of Rayport and Sviokla’s framework, Information and Management 40 (2002) 1–10.

[20] J.R. Miglautsch, Thoughts on RFM scoring, International Society for Strategic Marketing, Journal of Database Marketing 8 (2000) 67–72.

[21] S. Mitra, P. Mitra, S.K. Pal, Evolutionary modular design of rough knowledge-based network using fuzzy attributes, Neurocomputing 36 (2001) 45–66. [22] Z. Pawlak, Rough sets, International Journal of Computer and Information Science 11 (1982) 341–356.

[23] Z. Pawlak, Rough sets, in: T.Y. Lin, N. Cercone (Eds.), Rough Sets and Data Mining: Analysis for Imprecise Data, Kluwer Academic Publishers, Norwell, MA, 1997.

[24] Z. Pawlak, A. Skowron, Rudiments of rough sets, Information Sciences 177 (2007) 3–27. [25] Z. Pawlak, A. Skowron, Rough sets and Boolean reasoning, Information Sciences 177 (2007) 41–73. [26] Z. Pawlak, A. Skowron, Rough sets: some extensions, Information Sciences 177 (2007) 28–40. [27] J.F. Rayport, J.J. Sviokla, Managing in the market-space, Harvard Business Review 72 (1994) 141–150.

[28] J.Y. Shyng, F.K. Wang, G.H. Tzeng, K.S. Wu, Rough set theory in analyzing the attributes of combination values for the insurance market, Expert Systems with Applications 32 (2007) 56–64.

[29] R. Slowinski, C. Zopounidis, Application of the rough set approach to evaluation of bankruptcy risk, International Journal of Intelligent Systems in Accounting, Finance and Management 4 (1995) 27–41.

[30] H.S. Song, J.K. Kim, S.H. Kim, Mining the change of customer behavior in an internet shopping mall, Expert Systems with Applications 21 (2001) 157– 168.

[31] H. Suh, K.C. Noh, C.K. Suh, Customer list segmentation using the combined response model, Expert Systems with Applications 17 (1999) 89–97. [32] C.Y. Tsai, C.C. Chiu, A purchase-based market segmentation methodology, Expert Systems with Applications 27 (2004) 265–276.

[33] T.L. (Bill) Tsenga, C.C. Huang, Rough set-based approach to feature selection in customer relationship management, Omega 35 (2007) 365–383. [34] H.F. Wang, W.K. Hong, Managing customer proﬁtability in a competitive market by continuous data mining, Industrial Marketing Management 35

(2006) 715–723.

[35] Y.F. Wang, Mining stock price using fuzzy rough set system, Expert Systems with Applications 24 (2003) 13–23.

[36] J.C. Wong, Y.S. Chung, Rough set approach for accident chains exploration, Accident Analysis and Prevention 39 (2007) 629–637. [37] Y. Yao, Y. Zhao, Attribute reduction in decision-theoretic rough set models, Information Science 178 (2008) 3356–3373. [38] Y. Yao, Y. Zhao, Discernibility matrix simpliﬁcation for constructing attribute reducts, Information Sciences 179 (2009) 867–882.

[39] L.Y. Zhai, L.P. Khoo, S.C. Fok, Feature extraction using rough set theory and genetic algorithms an application for the simpliﬁcation of product quality evaluation, Computers and Industrial Engineering 43 (2002) 661–676.

[40] D. Zhu, G. Premkumar, X. Zhang, C.H. Chu, Data mining for network intrusion detection: a comparison of alternative methods, Decision Science Journal 32 (2001) 635–660.

[41] L.A. Zadeh, A new frontier in computation—computation with information described in natural language, keynote speaker’s lecture notes, IEEE Conference on Systems, Man, and Cybernetics, Taipei, Taiwan, 2006.