Predicting tourism loyalty using an integrated Bayesian network mechanism

Expert Systems with Applications 36 (2009) 11760–11763

Contents lists available at ScienceDirect

Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa

Predicting tourism loyalty using an integrated Bayesian network mechanism Chi-I Hsu a, Meng-Long Shih b, Biing-Wen Huang c,*, Bing-Yi Lin a, Chun-Nan Lin d a

Dept. of Information Management, Kainan University, No. 1 Kainan Road, Luchu, Taoyuan County 338, Taiwan Dept. of Social Studies Education, National Taitung University, Taiwan Dept. of Applied Economic, Chung-Hsing University, Taiwan d Dept. of Tropical Agriculture and International Cooperation, National Ping-Tung University of Science and Technology, Taiwan b c

a r t i c l e

i n f o

Keywords: Tourism management Loyalty Bayesian networks Linear structural relation model

a b s t r a c t For effective Bayesian networks (BN) prediction with prior knowledge, this study proposes an integrated BN mechanism that adopts linear structural relation model (LISREL) to examine the belief or causal relationships which are subsequently used as the BN network structure for predicting tourism loyalty. Four hundred and fifty-two valid samples were collected from tourists with the tour experience of the Toyugi hot spring resort, Taiwan. The proposed mechanism is compared with back-propagation neural networks (BPN) or classification and regression trees (CART) for 10-fold cross-validation. The results indicate that our approach is able to produce effective prediction outcomes. Ó 2009 Elsevier Ltd. All rights reserved.

1. Introduction With the prevalence of tourism in Taiwan, the development of tourism industry has facilitated local economy and increased the employment opportunity. Thus, tourism becomes an industry that is valued and actively developed by the government. However, when facing a more competitive tourism environment, how to attract the customers and further transform them into loyal ones will be the key for the operation of leisure business. This research seeks to determine the factors that influence tourism loyalty. Moreover, this study proposes an integrated mechanism that combines LISREL (Joreskog & Sorbom, 1993) and BN (Pearl, 1986) to predict a tourist’s loyalty level. Valid samples were collected from 452 tourists with the tour experience of the Toyugi Hot Spring Recreational Village, which is located at the eastern region of Taiwan. The village is managed by Taitung County Farmers Association. With an area of 15 hectare, it is the largest hot spring recreational village in Chihben hot spring area. The village is full of rich ecological resources such as spatial grasslands, varied plants, wild birds and butterflies. Therefore, the village provides the visitors the combined leisure functions such as recreation, conference, experience, education and training by the unique hot spring, landscapes and farm produces. Visitors can enjoy the services, herbs and various agricultural products in hot spring hotels. Besides, the village also provides junior high schools and elementary schools a teaching space for rural village

* Corresponding author. Tel.: +886 422840349; fax: +886 422860255. E-mail addresses: [email protected] (Chi-I Hsu), [email protected] (M.-L. Shih), [email protected] (B.-W. Huang), m93221001@ ms2.knu.edu.tw (B.-Y. Lin), [email protected] (C.-N. Lin). 0957-4174/$ - see front matter Ó 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2009.04.010

ecology and experience, and the information of agricultural tourism. It is a recreational village with different features. The data analysis was conducted in two stages: verifying relationships in the research model and predicting the level of tourism loyalty. In the first stage, LISREL was used to verify the belief or causal relationships in the research model of tourism loyalty. LISREL is a structural equation modeling (SEM) technique used to determine whether a research model is valid by examining the goodness-of-fit between the model and raw data. It has been widely applied in social-science research. In the second stage, the supported relationships of the LISREL analysis are used as the BN network structure to predict a tourist’s loyalty level. The predicted results were also compared with those generated by BPN and CART. 2. Tourism loyalty Loyalty refers to the repurchase will of certain products and services (Jones & Sasser, 1997). Fornell, Johnson, Anderson, Cha, and Bryant (1996) suggested that ‘‘repurchase possibility” and ‘‘repetitive purchase” are two critical factors for loyalty assessment. Sirohi, Mclaughlin, and Wittink (1998) indicated the following as the indexes to assess loyalty: (1) continuous purchase; (2) increase purchase in the future; (3) recommendation for others’ purchase. This research defines tourism loyalty as the visitors’ will to revisit and recommend the destinations to others after arriving at the tourist attraction. The measurement indexes include continuous revisiting, revisiting will and recommendation to others. Loyalty is one of the targets of strategic marketing and it allows companies to enhance the competitive advantages (Craft, 1999). The benefits of loyalty include below: (1) customers’ repurchase

11761

C.-I Hsu et al. / Expert Systems with Applications 36 (2009) 11760–11763

and promotion willingness can lead to revenue growth of the firms and the increase of market share; (2) reduction of costs; (3) increase of employees’ work satisfaction (Jacoby, 1994). In order to increase the customers’ loyalty, companies need positive customer relationship management (CRM). CRM means the enterprises find the customers’ real needs with the support of process and technology, and improve the products and services that are devoted to the enhancement of customer loyalty (Kalakota & Robinson, 1999). Spengler (1999) also suggested that CRM integrates planning, marketing and customer service by information technology, and provides customized services to increase customer loyalty and corporate operational benefits. In addition, Hui, Wan, and Ho (2007) indicated the characteristics of tourist attractions, such as interesting cultures, attractive urban sightseeing, interesting night life and attractive natural and scenic aspects might increase customer satisfaction and revisiting will. According to the literature review discussed above, this research proposes three factors which might increase tourism loyalty: customer service, web function with the support of technology and local characteristics. Three factors influence travelers’ tourism loyalty are proposed based on a literature review, including (1) Customer Service (CS): the service consumers received from employees; (2) Web Function (WF): the functions providing by the tour web site; (3) Local Characteristics (LC): the consumer’s perception of the local tourism characteristics; and Tourism Loyalty (TL): the loyal degree regarding a tourist revisit a destination. This study suggests that the greater the degree to which a tourist perceived regarding customer service, web function and local characteristics of the destination, the greater is his/her tourism loyalty, which refers to repeat of visit, willing to revisit and recommend to others. Therefore, H1–H3 is established as below: H1: Customer service positively influences tourism loyalty. H2: Web function positively influences tourism loyalty. H3: Local characteristics positively influences tourism loyalty. All questionnaire items are shown in Table 1. Each tourist was asked to rate on a scale of 1–5 his or her degree of agreement with each item. 3. The integrated Bayesian network mechanism To overcome the difficulty of constructing a BN structure when learning from data, this study proposes a novel approach that combines LISREL and BN to predict a tourist’s loyalty level. LISREL is an advanced statistical technique in the social and behavioral sciences

Table 1 The questionnaire items.

to verify the hypothesized relationships, but it is seldom combined with other machine-learning algorithms. This study uses LISREL to aid BN in discovering a suitable network architecture for prediction. 3.1. LISREL LISREL is one SEM technique which combines the concepts of both factor analysis and path analysis. It is especially appropriate to use LISREL to analyze the data in social and behavioral research fields. While multiple regression can estimate the parameters of only one linear equation at a time, LISREL can simultaneously process multiple sets of variable relationships to estimate the parameters in an entire system of linear equations in a model. The LISREL model and equations are shown in Fig. 1 in which E is disturbance; g is the vector of endogenous latent variable; a is intercept; B is the matrix of regression coefficient for endogenous latent variable; C is the matrix of regression coefficient for exogenous latent variable; n is the vector of exogenous latent variable; and f is the vector of latent disturbance. The analysis consists of two steps: (1) measurement model analysis, which aims to analyze the loading relationships between latent variables and their corresponding observable variables, and (2) structural model analysis, which aims to analyze the hypotheses relationships among latent variables. 3.2. BN A BN is a graphical model of variables and their relationships based on probability theory. It is also called a belief network or causal network. A BN uses prior probabilities and probabilities in sample space to estimate posterior probabilities. In a BN graph, arrows between nodes are used to represent a directed acyclic graph (DAG) (Niedermayer, 1998). Each parent node represents the cause of an event, a child node represents the outcome, and an arrow represents the causal relation. As shown in Fig. 2, the set of parent nodes of a node TL is denoted by parents(TL) and the joint distribution of the node values can be written as the product of the local distributions of each node and its parents. Advantages of BN include the ability to analyze problems with incomplete data and to combine domain knowledge and data (Hackerman & Wellman, 1995). However, without prior understanding or knowledge about the problem domain, the required significant computational effort of an NP-hard task in exploring a

E1

CS1 V1

E2

CS2 V2

E3

V1 CS3

Factor

Item

Content

E4

CS4

Customer service

CS1 CS2 CS3 CS4 CS5

Quick response to customers’ suggestions Fluent service flow Understanding customers’ need Quickly solving customers’ problems Listening to customers’ complaint

E5

CS5

E6

WF1

WF1 WF2 WF3

Web information attracted people Ease of use Web site security

LC1 LC2 LC3

Farm products have local characteristics The brand and image of farmers’ association is relieved Service and produce have geographic features

TL1 TL2 TL3

Repeat of visit Revisiting will Recommendation to others

Web function

Local characteristics

Tourism loyalty

E7

WF2

E8

WF3

E9

LC1

E10

LC2

E11

LC3

CS

TL

TL1

E12

TL2

E13

TL3

E14

WF

LC

η = αη + Bη + Γξ + ς Fig. 1. LISREL: verifying the believe relationships.

11762

C.-I Hsu et al. / Expert Systems with Applications 36 (2009) 11760–11763

CS

WF

collected after deducting the sample with more than three questions not answered. The effective response rate was 82.2%. The resultant sample was then randomly split into two subsamples. The first subset was used for exploratory factor analysis (EFA) to identify the factor structure hidden in the data collected. The other subset was used for LISREL analysis. SPSS 13.0 for Window version was used for EFA analysis. As shown in Table 2, using the Eigen value rule, a 4-factor structure emerged. The reliability and validity are examined as shown in Table 3. The Cronbach’s a values for all constructs are all above the 0.7 level (Cronbach, 1947). We examined convergent validity with the composite reliability (CR) and the average variance extracted (AVE) by the constructs. A CR greater than .60 is preferred (Fornell & Larcker, 1981) and all constructs in this study met this requirement. The AVE for all constructs in this study exceeded the preferred 0.5 (Fornell & Larcker, 1981).

LC

TL

P (CS ,WF , LC , TL ) = ∏ P (TL parents (TL)) Fig. 2. BN: predicting the outcome level.

Table 2 The result of EFA. Factor

Item

1

2

3

4

Customer service

CS1 CS2 CS3 CS4 CS5

.796 .692 .648 .751 .700

.128 .284 .380 .081 .188

.014 .061 .301 .281 .364

.181 .065 .123 .159 .141

Web function

WF1 WF2 WF3

.274 .191 .258

.794 .895 .742

.106 .155 .259

.251 .142 .220

Local characteristics

LC1 LC2 LC3

.131 .287 .158

.205 .108 .142

.764 .649 .766

.257 .262 .113

Tourism loyalty

TL1 TL2 TL3

.044 .163 .265

.256 .191 .100

.240 .145 .211

.701 .821 .769

4.2. Verifying relationships in the research model We employed LISREL to examine the measurement and structural models. Regarding whether the measurement model and structural model are good of fit, the criteria of goodness-of-fit measures are as follows: v2 =df is suggested to be smaller than 2 for good of fit (Carmines & McIver, 1981) and 3 for acceptable fit (Chin & Todd, 1995); CFI is suggested to be greater than 0.95 for good fit (Bentler, 1995); NFI, NNFI greater than 0.9 for good fit (Hu & Jen, 2005); IFI greater than 0.9 for good fit (Hu & Jen, 2005); GFI and AGFI greater than 0.9 for good of fit, and 0.8 for acceptable fit (Subhash, 1996); PGFI greater than 0.5 for good fit (Mulaik, James, Van Altine, Bennett, & Stilwell, 1989); SRMR smaller than 0.08 for good fit (Hu & Jen, 2005); and RMSEA smaller than 0.05 for good fit and 0.08 for acceptable fit (McDonald & Ho, 2002). The goodness-of-fit indexes for both measurement and structural models are acceptable as shown in Table 4. As shown in Table 5, all path coefficients are significant at the 0.05 level in the structural model. The results indicate that the hypothesized relationships are supported. Square multiple correlations (SMC) are also reported, which is 0.73 indicating the explained proportion of variance in tourism loyalty.

Table 3 Cronbach’s a, CR, and AVE.

CS WF LC TL

a

CR

AVE

0.849 0.873 0.731 0.770

0.899 0.785 0.764 0.875

0.642 0.549 0.520 0.700

previously unknown network is costly and inefficient (Niedermayer, 1998). BN is becoming an increasingly important solution for practical problems in the field of Artificial Intelligence (Korb & Nicholson, 2003). The applications of BN include the areas of maintenance project delays based on specialists experience (de Melo & Sanchez, 2008), detecting firms that issue fraudulent financial statements (FFS) and identifying factors associated to FFS (Kirkos, Spathis, & Manolopoulos, 2007), and other problem domains.

4.3. Predicting the level of tourism loyalty In the above section, we describe LISREL analysis employment to verify the hypothesized relationships in the proposed research model. Subsequently, this study adopts the supported relationships as the BN network structure to predict a tourist’s loyalty level. Based on the LISREL analysis results, the nodes CS, WF, and LC represent the antecedents of the outcome node TL. The input data for each node is the average value of the corresponding questionnaire items. The BN software Netica 2.05 (NORSYS, 2000) was used to construct the probability model using the 452 valid samples. The same software was used to predict the outcome. All the learning/ testing results were obtained for 10-fold cross-validation. For comparison purposes, the BPN and CART modules in SPSSClementine 8.1 software package were used to predict the same outcome TL. The system’s default control parameters were adopted. The input variables are the same constructs in the research model including CS, WF, and LC.

4. The experiment results 4.1. Sample and exploratory factor analysis This study develops a questionnaire based on the proposed items and delivered to 550 travelers with the tourism experience of the Toyugi hot spring resort of farmers’ association, Taitung County, Taiwan. Four hundred and ninety-eight travelers replied the questionnaires. Four hundred and fifty-two valid samples were Table 4 The goodness-of-fit indexes.

Criteria Measurement model Structural model

v2 =df

CFI

NFI

NNFI

GFI

AGFI

PGFI

RMSEA

SRMR

<3 1.80817 1.80817

>0.95 0.99 0.99

>0.90 0.97 0.97

>0.90 0.98 0.98

>0.80 0.99 0.92

>0.80 0.89 0.89

>0.50 0.63 0.63

<0.08 0.060 0.060

<0.08 0.042 0.042

C.-I Hsu et al. / Expert Systems with Applications 36 (2009) 11760–11763 Table 5 The hypothesis tests summary. Hypothesis

Content

Path coefficient

t-Value

Support

H1 H2 H3

CS ? TL WF ? TL LC ? TL

0.30 0.18 0.46

2.80 2.06 3.60

Yes Yes Yes

Table 6 The experiment results. LISREL–BN

BPN

CART

Training

Testing

Training

Testing

Training

Testing

1 2 3 4 5 6 7 8 9 10

0.143 0.265 0.263 0.280 0.285 0.278 0.265 0.263 0.283 0.283

0.239 0.348 0.370 0.217 0.174 0.239 0.348 0.370 0.196 0.196

0.667 0.663 0.688 0.663 0.690 0.666 0.666 0.690 0.649 0.690

0.804 0.733 0.511 0.444 0.600 0.733 0.756 0.622 0.289 0.609

0.667 0.673 0.715 0.678 0.690 0.676 0.676 0.715 0.681 0.687

0.761 0.489 0.400 0.178 0.533 0.444 0.556 0.400 0.178 0.413

MAE

0.261

0.270

0.673

0.610

0.686

0.435

The 10-fold cross-validation results for the mean average errors (MAE) for LISREL–BN, BPN and CART are shown in Table 6. It is indicated that our approach obviously outperforms BPN and CART in both the training (.261 vs. .673 and .686) and testing results (.270 vs. .610 and .435) with 10-fold average basis. 5. Conclusions 5.1. Tourism implications According to the LISREL results, we found that: (1) customer service, web function and local characteristics have significant and positive influence on tourism loyalty. Local characteristics are the most important, followed by customer service and web function. In order to increase tourism loyalty to attract visitors’ revisiting or recommendation to others, and increase the number of visitors, the key of tourism management is the reinforcement or use of the characteristics of the spots. With regard to customer service, tourism management should value the quick response to customers’ suggestions and quickly solve the customers’ problems. Finally, with the prevalence of Internet business application, the web function of tourism websites is commonly used by visitors. The tourism administrators should recognize that the enhancement and improvement of web function will increase tourism loyalty. 5.2. Methodology discussion Although LISREL and BN have been widely applied individually in many research studies, few of these studies have investigated combining them for predictive purposes. This research has proposed an integrated mechanism to predict a tourist’s loyalty level. The prediction performance of this approach has also been compared with the predictions of BPN and CART. The three methods LISREL, BN, and BPN represent relationships as networks. However, in terms of explanatory power, the BPN’s internal learning is processed in a black-box mode in which the internal weights are difficult to express in an explicit way which

11763

is relevant to real-world problems. Our approach is able to explain the correlative relationships between variables and outperforms BPN and CART by achieving higher predictive accuracy. BPN and CART are appropriate for exploratory studies in which the relationship between variables is unclear, whereas our approach is suitable for data prediction in empirical research with a theoretical basis. Because constructing a BN structure when learning from data presents certain difficulties (Niedermayer, 1998), the approach proposed in this research is demonstrated to be a prospective way to aid a BN in discovering a suitable network architecture with better prediction performance. In the future, a sensitivity analysis would be helpful to understand the approach robustness. Further simulations are needed to be of general interest. As for the contribution of this research, LISREL is a widely used advanced statistical tool in the social and behavioral sciences, but it is seldom combined with other machine-learning algorithms. The proposed approach can be used as a good reference for related research in social-science fields. References Bentler, P. M. (1995). EQS structural equations program manual Multivariate Software. CA: Encino. Carmines, E. G., & McIver, J. P. (1981). Analyzing models with unobserved variables: Analysis of covariance structures. In G. W. Bohrnstedt & E. F. Borgatta (Eds.), Social measurement: current issues. Beverly Hills, CA: Sage. Chin, W. W., & Todd, P. (1995). On the use, usefulness and ease of use of structural equation modeling in MIS research: A note of caution. MIS Quarterly, 19(2), 237–246. Craft, S. H. (1999). Marketers gain by measuring true loyalty. Marketing News, 33, 18. Cronbach, L. (1947). Test ‘reliability’: Its meaning and determination. Psychometrika, 16, 1–16. de Melo, Ana C. V., & Sanchez, Adilson J. (2008). Software maintenance project delays prediction using Bayesian Networks. Expert Systems with Applications, 34(2), 908–919. Fornell, C., & Larcker, D. F. (1981). Evaluating structural equation models with unobservables and measurement error. Journal of Marketing Research, 18(1), 39–50. Fornell, C., Johnson, M. D., Anderson, E. W., Cha, J., & Bryant, B. E. (1996). The American customer satisfaction index: Nature, purpose, and finding. Journal of Marketing, 60(spring), 7–18. Hackerman, D., & Wellman, M. P. (1995). Bayesian networks. Communications of the ACM, 38(3), 27–30. Hu, K. C., & Jen, W. (2005). Applications of LISREL and neural network to analyze the passenger’s behavioral intentions. Logistics Research Review, 8, 43–55. Hui, T. K., Wan, D., & Ho, A. (2007). Tourists’ satisfaction, recommendation and revisiting Singapore. Tourism Management, 28, 965–975. Jacoby, R. (1994). Why some customers are more equal than others. Fortune, 130, 9–13. Jones, T. O., & Sasser, W. E. Jr., (1997). Why satisfied customers defect. Harvard Business Review, 73(16), 88–99. Joreskog, K., & Sorbom, D. (1993). LISREL 8: User’s reference guide. Chicago: Scientific Software International. Kalakota, R., & Robinson, M. (1999). E-business: Roadmap for success (first ed.). USA: Mary T. O’Brien. Kirkos, E., Spathis, C., & Manolopoulos, Y. (2007). Data mining techniques for the detection of fraudulent financial statements. Expert Systems with Applications, 32(4), 995–1003. Korb, K. B., & Nicholson, A. E. (2003). Bayesian artificial intelligence. Chapman & Hall/ CRC. McDonald, R. P., & Ho, M. R. (2002). Principles and practice in reporting structural equation analysis. Psychological Methods, 7, 64–82. Mulaik, S. A., James, L. R., Van Altine, J., Bennett, N., & Stilwell, C. D. (1989). Evaluation of goodness-of-fit indices for structural equation models. Psychological Bulletin, 105, 430–445. Niedermayer, D. (1998). An introduction to Bayesian networks and their contemporary applications. Retrieved 20.06.06. NORSYS. Netica 2.05. (2000). Retrieved 10.12.06. Pearl, J. (1986). Fusion, propagation, and structuring in belief networks. Artificial Intelligence, 29(3), 241–288. Sirohi, N., Mclaughlin, E. W., & Wittink, D. R. (1998). A model of consumer perception and store loyalty intentions for supermarket retailer. Journal of Retailing, 74(2), 223–245. Spengler, B. (1999). Eyes on the customer. Computer World, 33, 60–62. Subhash, S. (1996). Applied multivariate techniques. NY: John Wiley and Sons.