Mining Tourist Motive for Marketing Development via Twice-Learning

Abstract

Prediction of tourist decision-making processes, including tourist motive, attitude, behavior, and so on, is of great importance for the development of tourism marketing strategies. Recently, applying machine learning techniques to predict tourist decision-making processes has drawn much attention. However, many machine learning techniques applied in the tourist decision-making prediction task fail to address two practical yet important problems. One is the failure of constructing models that can generate accurate yet comprehensible predictions at the same time, and the other is the failure to accommodate the characteristics of data collected from tourists that is usually small yet noise prone. In this article, we address the two entangled problems using the twice-learning framework to predict tourist motive from tourist external and internal features data collected through on-site survey. The results indicate that, based on the two-phase learning process, we can predict tourist motive accurately as well as extract meaningful insights, which are useful for targeted marketing strategies development from the real-world data.

INTRODUCTION

With the tremendous increase in tourism demand in recent decades, tourism has become one of the world’s fastest growing industries. Prediction of tourist decision-making processes, including tourist motive, intention, behavior, and so forth, is widely regarded as a study of great importance for providing guidance for policy-making in the development of the industry (Hsu, Cai, and Li 2010; Lee 2009; Valle et al. 2012). Useful prediction models of the tourist decision-making process should not only generate accurate predictions but also explicitly explain why such predictions are made, in order to provide practitioners with meaningful insights (Song and Li 2008). The state-of-the-art prediction models used in tourism research include rough set models (Celotto, Ellero, and Ferretti 2012; Goh and Law 2003), structural equation models (Chi and Qu 2008; Gross and Brown 2008; Jalilvand et al. 2012) and so on. Although these models are comprehensible to some extent, they are somewhat more descriptive than predictive, and vulnerable to noise and model assumptions.

Since machine learning techniques were introduced to the tourism research in the late 1990s (Law and Au 1999; Pattie and Snyder 1996), the prediction performance in tourism research has been greatly improved. However, most of the machine learning techniques applied in tourism research are “black-box” in nature, such as neural network (Chen, Lai, and Yeh 2012; Law 2000; Palmer, Montaño, and Sesé 2006) and support vector regression (Chen 2011; Chen and Wang 2007; Hong et al. 2011). Little insight on the knowledge hidden beneath the data can be drawn from the model representation, which limits the application of machine learning in analyzing data in real-world tourist decision-making processes. To make machine learning suitable for tourist decision-making process prediction, a learner that is able to generalize well and comprehend the knowledge learned from data is highly demanded. To the best of our knowledge, no such models have ever been applied to tourist decision-making process prediction in empirical tourism research.

Moreover, another major obstacle for using machine learning to effectively predict the tourist decision-making process is that constraints in the data collection process and the data properties have never been given serious consideration during the learning process. The data related to tourists are usually taken from on-site surveys, and consequently have following inevitable flaws. (1) The amount of the dataset is limited. Questionnaires and interview surveys are the major ways to collect data from tourists. The data collection process is time and energy consuming, which inherently limits the amount of data that can be collected on site. Tourists might have no time or might be unwilling to complete the questionnaire surveys or be interviewed during their tour, which also accounts for the limited extent of the dataset. (2) The collected data are always noise-prone. Among those people who agree to be surveyed, some will not take it seriously and, therefore, will give the answers to the questions in a casual and unserious way. Also, if the questions are related to their privacy, they will be reluctant to provide answers. These situations lead to the fact that a large number of respondents will provide wrong answers or not answer the questions completely in the survey. Thus, the tourist decision-making process prediction tasks will usually end up with a small yet noisy dataset, which raises a big challenge to select an appropriate machine learning method for effective prediction. Unfortunately, up to now, most of the research related to tourist decision-making process prediction just applies existing machine learning methods in a straightforward way, without adapting any methods to the data properties in the tasks. Although some positive results might be achieved, the performances are still far from what could be done.

Therefore, to make effective predictions for the tourist decision-making process, a learner must have strong generalization ability and comprehension in order to process even small and noisy data. In this article, we address the problem with a twice-learning framework (Zhou and Jiang 2003), which is able to construct a comprehensible learner that generalizes well from essentially small and noisy datasets. In detail, a neural network ensemble is trained from the original dataset in the first learning phase, and then, in the second learning phase, virtual examples are generated from the trained neural network ensembles and used to train a set of decision rules. The empirical studies show that we can not only achieve more accurate predictions than the methods that are widely used in the empirical tourism research, but also discover useful rules that explicitly explain why such predictions are made. The discovered rules provides meaningful and effective insight into the development of the tourism industry.

The remainder of this article is organized as follows: the next section describes the twice-learning framework in detail; “Mining Practice” presents the mining practice to generate a comprehensible model, and then reports and analyzes the findings; “Related Work” presents the related work of the application of machine learning in empirical tourism research; the last section concludes this article and offers suggestions for future research.

TWICE-LEARNING FRAMEWORK

Because most of the traditional data mining approaches fail to produce models with strong generalization ability as well as high comprehensibility at the same time, Zhou and Jiang (2003) proposed a twice-learning framework, which elaborates on combining these two abilities in a new way. This learning framework includes two phases of work, as shown in Figure 1. In the first phase, a model with strong generalization ability is constructed using the original training data. This phase is aimed to preprocess the original dataset. In the second phase, another model with high comprehensibility is constructed from the preprocessed data in order to obtain the prediction that can be ultimately understood by human experts. Note that the noise contained in the original dataset would be “smoothed out” in the first phase, and the virtual examples are usually generated to augment the training data, which makes the learning process effective even if the training dataset is small. By employing the whole procedure, both strong generalization ability and high comprehensibility can be achieved at the same time. The algorithms of the framework used in each phase follow.

FIGURE 1 The twice-learning framework.

Let S= denote the training set. First, a model with strong generalization ability is employed to construct a predictive model N* from the training set S. And then N* is used to predict the concept label y_i′ for each example x_i (i = 1,2, …, n) in the training set. By replacing the true concept label y_i of x_i with y_i′, a new dataset S′ is obtained. Note that y_i′ may not be the same as y_i. Because of the strong generalization ability of N*, it can usually provide more accurate prediction than other predictive models. By replacing the true label with the predicted label, the labeling noise contained in S would be smoothed out by the accurate prediction produced by N*, and hence, the S′ is less noisy than the original training set S, which is beneficial for the learning algorithm with strong comprehensive ability in the second phase.

In the second phase, a model with high comprehensibility is used to generate the predictions that can be clearly understood by human experts. Because a large training set is usually required in the learning process, the current training set S′ is enlarged by generating additional virtual examples. In detail, m random examples are generated and fed to N* for predictions, and then these predictions are used as their true labels. Let denote the label of x_j, and then a new dataset S″= is obtained. Finally, the filtered training set S′ is combined with the newly generated virtual example set S″ to create a larger training set S_R, and the enlarged training set is used for the model with high comprehensibility in the second phase.

One instantiation of the twice-learning framework employs neural network ensemble in the first phase and the C4.5 Rule in the second phase. Neural network ensemble (Hansen and Salamon 1990), which combines multiple neural networks for one target, has the strongest generalization ability. The C4.5 Rule (Quinlan 1993), which generates a set of decision rules in the form of “IF a AND b THEN c” can be essentially understood by the users. This instantiation, referred to as C4.5 Rule-PANE, has been successfully applied to computer-aided medical diagnoses (Zhou and Jiang 2003) and gene expression analysis (Jiang, Li, and Zhou 2006). The pseudocode of C4.5 Rule-PANE is shown in Table 1.

TABLE 1 The Pseudocode of C4.5 Rule-PANE Algorithm (Zhou and Jiang 2003)

Algorithm: C4.5 Rule-PANE
Input: training set S, the number of additional training instance m
Process:
N* = Bagging (S)    /* generate neural network ensemble using Bagging */
S’ = Φ; S’’= Φ
for i = 1 to n  {    /* process the training set by the trained ensemble */
yi’ = N*(xi: (xi, yi) ∈ S)
S’ = S’ ∪ {(xi, yi’)}
}
for i = 1 to m  {    /* generate additional training data set using ensemble*/
xj = RandVec()   /* generate a random feature vector */
yj’ = N*(xj)
S’’ = S’’ ∪ {(xj, yj’)}
}
SR = S’ ∪ S’’      /* construct the training data set for rule learning */
R = Rule Inducer(SR)
Output: rule set R

Note that, to guarantee effectiveness, C4.5 Rule-PANE should satisfy the applicability condition proposed by Zhou and Jiang (2004). The theoretical justification for the working mechanism is also provided in their work. The condition is summarized into an Application Theorem, which is listed as follows:

Application Condition (Zhou and Jiang 2004):

If the neural network ensemble in the first phase can achieve better performance in prediction than C4.5 Rule and if both are trained from the same dataset, C4.5 Rule-PANE can be applied to yield a model with strong generalization ability and comprehensibility.

MINING PRACTICE

In this study, we applied a twice-learning framework, C4.5 Rule-PANE, to mine tourist motive from tourist external and internal characteristics data collected through on-site survey. First, the process of data collection is introduced. Second, the effectiveness and superiority of C4.5 Rule-PANE on the collected data is verified. After that, the C4.5 Rule-PANE algorithm is used to construct the predictive model, and then the achieved findings are reported and discussed.

TABLE 2 The Detail Information of the Attributes

ID	Att_Name	Possible Values
1	Gender	Male/Female
2	Age	18–24/25–34/35–44/45–54/55–65
3	Education	High School or below/Junior College/University/Postgraduate
4	Occupation	Students/Teachers/White-collar workers/Blue-collar workers/Managers or Executives/Retirees/Others
5	Personal monthly income (RMB)	3000 or below/3001–5000/5001–10000/10001or above
6	Life-cycle stage	Single/Married, with no children/Married, with children/Others
7	Being well-respected	1(not important at all)˜5(very important)
8	Fun and enjoyment in life	1(not important at all)˜5(very important)
9	Security	1(not important at all)˜5(very important)
10	A sense of belonging	1(not important at all)˜5(very important)
11	Excitement	1(not important at all)˜5(very important)
12	Self-fulfillment	1(not important at all)˜5(very important)
13	Self-determination	1(not important at all)˜5(very important)
14	Convenience	1(not important at all)˜5(very important)

Data Collection

Data were collected through on-site surveys in Nanjing, China, from October to November 2012. The targeted respondents were tourists aged 18–65 years. Each respondent was required to complete a 15-item questionnaire referring to their personal characteristics (i.e., input attributes) and motive for tourism (i.e., target concept). Demographics and personal values are incorporated to represent tourists’ extrinsic and intrinsic characteristics, respectively. A five-point Likert-type scale is used, ranging from 1 (very unimportant) to 5 (very important), to evaluate the importance of each item of personal values perceived by the respondents in their daily life. The target concept refers to tourist motive to search for “relaxation and refreshment” (denoted by “physical”) versus “broaden knowledge and horizon” (denoted by “cultural”). The detail information of the attributes is tabulated in Table 2. Finally, 121 valid questionnaires were collected. Each valid questionnaire i(i = 1,2, … 121) is converted into a 14-dim vector of attributes x_i = (x_i1, x_i2, …, x_i14)^T and a concept label y_i, where x_i^j indicated the answer to the jth question in questionnaire i, and .

Application Verification

In order to evaluate the applicability as well as the generalization performance of C4.5 Rule-PANE in this study, four widely used machine learning methods—naïve Bayes, neural network, neural network ensemble, and C4.5 Rule—are compared with C4.5 Rule-PANE before analyzing the current data. These four baseline methods have been generally considered as the powerful classification methods in empirical tourism research. All the methods are implemented using Waikato Environment for Knowledge Analysis (WEKA; Witten and Frank 2000), and the parameters of these methods are set as the default value in WEKA. Ten-fold cross validation is conducted to evaluate the performance of these methods.

TABLE 3 p-Values of the Pairwise t-Test between Any Two Compared Methods (Significance Level 95%)

Methods	Naïve Bayes	C4.5 Rule	Neural Network	Neural Network Ensemble	C4.5 Rule-PANE
Naïve Bayes	—	—	—	—	—
C4.5 Rule	1.12E-02	—	—	—	—
Neural Network	6.86E-04	1.45E-07	—	—	—
Neural Network Ensemble	4.52E-04	5.07E-08	3.97E-01	—	—
C4.5 Rule-PANE	9.62E-06	1.36E-10	6.80E-02	1.99E-01	—

Figure 2 shows the average error rates for all the compared methods in this study. Pairwise two-tailed t-test at 95% significance level is conducted over the results, and the corresponding p-values between each pair of algorithms are listed in Table 3, where the entries showing significance are boldfaced. Figure 2 and Table 3 indicate the following:

1. Neural network ensemble performs significantly better than C4.5 Rule. This proves that the prerequisite of availability is satisfied.

   FIGURE 2 Error rates of all the compared methods in analyzing the current dataset.

   

2. C4.5 Rule-PANE is a little better than neural network and neural network ensemble, but significantly better than C4.5 Rule and naïve Bayes, which suggests that the nonlinear mapping learned by C4.5 Rule-PANE is much closer to the ground truth mapping than the other compared methods.

Above all, C4.5 Rule-PANE is more suitable for predictive modeling in this study, compared with the state-of-the-art machine learning methods that have been widely used in empirical tourism research. Therefore, in the following section, we explore tourist motive by using C4.5 Rule-PANE.

Discovering Important Attributes

The whole original dataset with extrinsic and intrinsic attributes is used as the training dataset for C4.5 Rule-PANE. Through the gain ratio (Quinlan 1993) values produced by C4.5 Rule-PANE, the importance of the personal characteristics in predicting tourist motive is ranked. Gain ratio can be regarded as an importance measure of each individual attribute in predicting the target concept. It measures the ability to discriminate among the examples belonging to different concept classes, based on the value of this attribute. The higher the gain ratio value is, the greater the generalization ability of the attribute. The gain ratio of each individual attribute produced by C4.5 Rule-PANE in this study is presented in Figure 3, and the importance of the personal characteristics in predicting what the tourist is mainly seeking during the tour is clearly indicated.

FIGURE 3 Gain ratios of the attributes obtained from C4.5 Rule-PANE.

Surprisingly though, by comparing with Figure 4, which represents the ranking of attributes provided by C4.5 Rule, different results are found. Without being preprocessed by neural network ensemble, the ranking list provided by C4.5 Rule shows that all of the personal attributes are of no importance in predicting tourist motive. Because the generalization ability of C4.5 Rule-PANE is much better than C4.5 Rule, the importance ordering of attributes disclosed by C4.5 Rule-PANE is much closer to the ground truth.

FIGURE 4 Gain ratios of the attributes obtained from C4.5 Rule.

Obtaining and Discussing Decision Rules

Because all the attributes may interact with each other to influence the final results, considering only the individual generalization ability would not lead to the accurate prediction for an individual. Thus, how these attributes interact with each other to produce the final predictions is further investigated, in order to find what personal characteristics account for a tourist motive.

TABLE 4 Decision Rules Learned from C4.5 Rule-PANE

Rule ID	Precondition	Decision Result
R1	IF Age = 35–44 AND Life-cycle stage = Married, with children AND Occupation = White-collar worker AND Being well-respected ≤ Average	physical
R2	IF Age = 25–34 AND Life-cycle stage = Married, with no children AND Education ≥ University AND Security ≥ Average	physical
R3	IF Gender = Female AND Education = High school or below AND Excitement ≤ Unimportant AND Convenience ≥ Important	physical
R4	IF Life-cycle stage = Single AND Education = University AND Self-determination ≥ Important AND Self-fulfillment = Very Important AND Security ≥ Important	cultural
R5	IF Age = 18–24 AND Occupation = Student AND Self-fulfillment ≥ important AND Fun and enjoyment in life ≤ Average	cultural
R6	IF Gender = Male, Income = 10001 or above, AND Excitement ≤ Unimportant AND Self-fulfillment ≥ Important	cultural

Because of the high comprehensibility of C4.5 Rule-PANE, the multivariate relationship from the input attributes to the target concept is constructed by the learned decision rules, presented in Table 4. These rules characterize the typical groups of tourists who seek “relaxation and refreshment” or “broaden knowledge and horizon” during pleasure traveling in the form of “IF a AND b THEN c,” where a, b, c are logical expressions. From these rules, the underlying truth of tourist motive can be further disclosed and interpreted.

Table 4 shows that, for the tourists who seek “relaxation and refreshment” (denoted by physical), three typical groups are found:

1. R₁: The first group is middle-aged people holding jobs in white-collar fields, married with children. Because the dual pressures of working as well as raising children, people at this stage usually use tourism as a way to relieve themselves. Because “being well-respected” is not the major concern for them, this category of tourist would probably pay more attention to physical rather than spiritual needs during a tour. Therefore, the pursuit of relaxation and refreshment would probably be the driving force of their tourism.

2. R₂: The second group is high-educated young people, married with no children. Most people in this group might have the ability to search for information on the Internet, such as the local culture and history of their destinations. In view of this, “broaden knowledge and horizon” might not be the major driving force in their tourism. Given their great concern for “security,” this group might probably view tourism as a way of pursuing ease and comfort.

3. R₃: The third group is women with lower level of education. Choi et al. (2008) indicated that uneducated female tourists usually have no interest in the local culture of their tourism destinations. Fun and relaxation would probably be their great concern during a trip. This argument is confirmed in this rule. As the rule presented, this group of tourists pay more attention to “convenience” and less attention to “excitement,” which indicates that the natural scenery sightseeing would be preferable to them, and relaxation and refreshment would probably be what they are seeking during a tour.

For the tourists who seek to “broaden knowledge and horizon” (denoted as cultural), three typical groups are found as well:

1. R₄: The first group is high-educated and unmarried people. Given their great emphasis on “self-fulfillment” and “self-determination” presented in the rule, this group of people would use tourism as a way to enrich and improve themselves. Because they pay much attention to security, they would not use tourism to pursue excitement and adventure. Therefore, a safe cultural tour that might broaden their knowledge and horizon would probably be their reason for a trip.

2. R₅: The second group is young students aged from 18 to 24. Most of these would be college students or graduate students. Because they attach much importance to self-fulfillment and less importance to fun and enjoyment in life, this group of tourists would probably pursue tourism as a way to enrich and improve themselves by means of increasing knowledge and experience at their destinations.

3. R₆: The third group is male with high income. Because they attach much importance to self-fulfillment and little importance to excitement, this group of people would probably hope to have an easy and comfortable experience with intellectual enjoyment during the tour. To broaden their knowledge and horizon in order to gain a sense of self-fulfillment would probably be the driving force for them to travel.

RELATED WORK

Machine learning has been successfully applied to many disciplines, such as information science (Li, Li, and Zhou 2009; Zhou, Chen, and Dai 2006), medicine (Li and Zhou 2007), biology (Jiang, Li, and Zhou 2006), social science (Zhang and Zhang 2012, 2014), and other fields. The first application of machine learning techniques in tourism research was in the late 1990s, and its superiorities in model construction and prediction accuracy have made it an ideal data analysis tool in this field. Typical machine learning methods employed in tourism research include neural network, Bayesian method, and decision rule/tree.

The neural network method was first introduced to tourism research by Pattie and Snyder (1996). Through the comparison of neural networks with several classical time series prediction methods, they indicated that neural network is a valid alternative to classical techniques in tourist prediction. Since then, the neural network method has been widely applied in tourism research, and its superiority of prediction accuracy over many classical prediction methods has been confirmed. For example, Law and Au (1999) used neural network to predict Japanese travel demand for Hong Kong and found that neural network outperforms multiple regression, naïve, moving average, and exponent smoothing. Burger et al. (2001) showed that the neural network method is the preferable prediction model over the naïve, decomposition, exponential smoothing, autoregressive integrated moving average, multiple regression, and genetic regression models. Kon and Turner (2005) pointed out that neural network generally outperforms the classical time series and multiple regression models in tourism prediction research. However, due to its “black-box” nature, the neural network method cannot provide the explanation of the predictions from the model representation, therefore, the method has little practical significance from the economic perspective (Song and Li 2008). With the continual applications of neural network in tourism prediction, the limitation of it has been widely recognized by the tourism academics.

The Bayesian method has been widely used in predicting tourist attitudes and behaviors. Hsu, Tsai, and Wu (2009) proposed a mechanism to combine a structural equation model (SEM) with a Bayesian network to predict tourist loyalty level and achieved effective predictions. Melián-González, Moreno-Gil, and Araña (2011) used Bayesian model averaging (BMA) to investigate the relationship between the satisfaction of gay tourists and their perceptions of the condition in terms of the valuable resources. The analysis results would be beneficial to explain a destination’s competitiveness in the gay tourist segment. With the comparison of these predictions obtained by Bayesian methods with those obtained by other classical predictive models in these studies, the effectiveness of Bayesian network in predicting tourist attitudes and behaviors has been greatly confirmed. However, this paradigm highly relies on the choice of the Bayesian prior and model assumption and usually requires a large number of labeled training data for parameter estimation. So it might not achieve good performance when the prior knowledge of the problem to be solved is too insufficient to help select the Bayesian prior and determine the model assumption, and when the size of the training data is small. Besides, Bayesian inference is relatively slow, and, hence, it might not be of practical use in some applications that require quick prediction response.

Decision rule/tree has been used to characterize groups of tourists in the tourism market. Emel, Taskin, and Akat (2007) learned decision rules to characterize various groups of outgoing tourists in the domestic tourism market. Li et al. (2010) incorporated decision rule into a self-organizing, map-based segmentation process. With the obtained rules, several groups of outbound tourists in Hong Kong were profiled. Rong et al. (2011) learned decision rules to characterize two groups of outbound tourists, viz., the online experience sharers and travel website browsers, based on their demographic and travel information attributes. A number of decision rules were obtained, which would be helpful for tourism managers in defining new target customers. Compared with decision rule, the use of decision tree has not appeared in tourism research until Kim, Timothy, and Hwang (2011), who used decision tree to investigate the influential factors of Japanese tourists’ shopping preferences and intentions to revisit Korea. However, due to the limitation on the size of training data and the potential noise contained in the training data, the symbolic learning algorithm can achieve only reasonably good prediction accuracy, although its symbolic nature can help the user understand the discovered knowledge.

CONCLUSIONS

Prediction of tourist decision-making processes is of great importance in guiding the development of tourism marketing strategies. To make effective predictions of tourist decision-making processes, a learner with strong generalization ability and comprehensibility, as well as ability to accommodate the characteristics of data (such as the data is small and noisy prone), is required. To address the issue, this article introduces a twice-learning framework, C4.5 Rule-PANE, to generate a model that is accurate in prediction and essentially comprehensible to the data miners from a small yet noisy dataset. Mining practice on tourist motive prediction indicates that the twice-learning framework is effective, and the obtained comprehensible model discloses the importance of personal external and internal attributes in predicting tourist motive, as well as the relationship between a tourist’s personal attributes and his/her motive for pleasure traveling, which would be of great significance for tourism practitioners in developing targeted marketing and production strategies.

Note that, as well as demographic and personal-values attributes, more personal attributes such as tourist experience and self-concept would also be relevant to tourist motive and would be incorporated into the input attributes for prediction. Moreover, applying twice-learning methods, such as C4.5-Rule PANE, for predictive modeling in other empirical tourism research problems would be interesting future work.

FUNDING

The research is supported by the National Science Foundation of China under the Grant No. 41301145 and No. 41271149. The research is also supported by A Project Funded by the Priority Academic Program Development of Jiangsu Higher Education Institutions.

Additional information

Funding

The research is supported by the National Science Foundation of China under the Grant No. 41301145 and No. 41271149. The research is also supported by A Project Funded by the Priority Academic Program Development of Jiangsu Higher Education Institutions.