Application of experience economy and recommendation algorithm in tourism reuse of industrial wasteland

User profiles are descriptions of certain groups of people created by abstracting tags based on analysis of demographic information, travel preferences, social relationships and consumption behaviour. User profile construction requires establishment of a label grading system based on the business conditions. In accordance with the demographic principles and business conditions considered in this study, the labelling system is divided into the following three levels: level 1, which includes demographics, content, behaviour and business attributes; level 2, which includes location, economic strength, travel preference and brand label information, and level 3, which includes age, gender, home address, purchasing power, travel equipment, eating habits, consumption habits and activeness information.

First, user interest models (UIMs) are constructed. Unique identifiers U known as uIDs are established for the users, where U = {uID|〈t₁,w₁〉, 〈t₂,w₂〉, . . . 〈t_i,w_i〉, . . . 〈t_n,w_n〉}. Here, T’ = {t₁, t₂, . . . t_i, . . . t_n} is the set of topics t_n and W = {w₁, w₂, . . . w_i,. . . w_n} is the set of weights w_n, which indicate the extent to which users are interested in a given topic. These data are obtained by weighting and standardizing the number of times that a user browses, comments or signs in to the data-gathering application (e.g. the TripAdvisor website).

Users are then clustered through a clustering algorithm. Users with similar UIMs are utilized to form a user profile, while users with significant variations in the UIMs are assigned to different user profiles. The clustering algorithm proposed in this paper is similar to the k-medoids technique and is detailed in algorithm 1. The metric formula of the tourist UIM, or the Euclidean distance between two UIMs, is expressed as follows: (1) Dist(Ui,Uj)=(wi1−wj1)2+(wi2−wj2)2+...+(win−wjn)2 Dist({U_i},{U_j}) = \sqrt {{{({w_{i1}} - {w_{j1}})}^2} + {{({w_{i2}} - {w_{j2}})}^2} + ... + {{({w_{in}} - {w_{jn}})}^2}}

First, a cluster centre U_cen is selected. Then, the user farthest from U_cen is selected as the second cluster centre. Here, ‘farthest’ is defined by Eq. (1) and is mainly used to measure the differences between user interests. Cluster centres are continuously selected in this manner until there are k members. Clus.S_Ucen represents the score of U_cen as the cluster centre. The clustering score is the sum of the distances from each UIM to U_cen. Clearly, the lower the clustering score, the better the clustering result. To achieve more compact clustering, the algorithm attempts to replace U_cen with other UIMs (i.e. U_i) as the clustering centre and then calculates the clustering score. If the result of using U_i as a cluster centre is more compact, i.e. S_Ui < S_Ucen, U_cen is replaced with U_i. This process is repeated until there are no changes, which means that the most compact cluster centre has been found. Ultimately, the cluster formed through clustering is composed of multiple UIMs. Certain topics with the highest weight values are extracted and labelled with tags from the three levels described above.

Experience economy is based on life and situation, shaping sensory experience and thinking identity, influencing and understanding users’ consumption behaviour. As a major feature of experience economy, ‘relationship’ emphasizes the long-term relationship with customers. Enterprises should consolidate and develop the relationship through repeated transactions so as to grasp the characteristics of users. Based on this, we combine the experience economy implementation tools proposed by Schmitt and user profiles, aimed to offer useful advice for development strategies for industrial wasteland tourism development.

It was found that the experience economy implementation tools applied to industrial tourism mainly encompass visual signs, media, service process, personnel characteristics, spatial surroundings and external associations. The population characteristics related to industrial tourism are as follows: there are more men than women among tourists aged 15–25 and 50–60 years; these individuals are middle- and high-income earners; and they have college degrees or higher.

In this study, a questionnaire survey was designed to investigate industrial tourism experience economy in Handan. The survey was administered at five typical industrial tourism sites in Handan. From 500 respondents, 469 (93.8%) completed and returned the questionnaire appropriately. This close-ended questionnaire contained two sections, experience economy combination and industrial tourism awareness, and had 20 questions in total. The 5-point interval scale was employed, with 5 points indicating ‘strongly agree’ and 1 point indicating ‘strongly disagree’.

The retrieved questionnaire data were analysed using IBM SPSS Statistics V22.0. Principal component analysis was employed, and the eigenvalues of four factors were extracted, explaining 60.17% of the population variance. The results are listed in Table 1.

Table 1 indicates that the 20 variables can be attributed to 4 factors. Apart from object display, visitors, product purchases, government support and industry alliance, the loading values of all other factors were >0.5, indicating that the questionnaire had good convergence.

The data were then subjected to multiple linear regression analysis; the results are presented in Table 2.

From Table 2, the t-test significance values of five variables (tourism signs, neighbouring tourist attractions, advertising, hotspot marketing and interactive experience) were <0.05, indicating that these five variables have significant impact on tourist awareness. The ranking of these factors is as follows, from high to low: advertising, hotspot marketing, interactive experience, neighbouring attractions and tourism signs. Based on the above mentioned findings, the development strategies are listed and verified in section 2.4.

As a result of the poor recognition of industrial tourism and the small number of tourists, the major problems facing development of a recommendation system for industrial tourism products are data sparsity and cold starts. In this study, a recommendation system called recommendation-user-tag-project (R-UTP) is proposed, which is based on a tripartite graph and a random walk algorithm. The definition of the user-tag-project (UTP) relationship tripartite graph is given below.

Fig. 1

Experiments were performed to evaluate the algorithm performance and compare it with similar algorithms. Each dataset could be divided into test and training sets, that is, 80% of the data were used as the training set and the remaining 20% were used as the test set. The experimental environment was the Windows 7 operating system with an Intel i5 3.3G CPU and 8GB memory.

In this study, the hit ratio and recall ratio were used to measure the effectiveness of the proposed algorithm. First, a list of the top-N recommended tourism products was obtained, where the products (represented by P) were generated by the algorithm. If a product of interest to the user appeared in the list of top-N recommended tourism products, one hit was recorded. The hit ratio was measured using the following formula: (5) HT=∑upu|P| HT = {{\sum\nolimits_u {p_u}} \over {|P|}}

Correspondingly, the recall ratio refers to the ratio of projects of interest to the user against the projects of interest to all users in the system, as measured by Eq. (6), where |N_u| represents the number of projects of interest to all users in the system: (6) R=∑upu|Nu| R = {{\sum\nolimits_u {p_u}} \over {\left| {{N_u}} \right|}}

The algorithm proposed in this paper was compared with two classical algorithms: the user-based system filtering algorithm UserKNN and the object-based collaborative filtering algorithm ItemKNN. First, the hit ratio variation in accordance with different values of α was investigated.

The R-UTP algorithm and the other two algorithms were executed on the two datasets individually. The experimental results are shown in Figure 2. With the change of α, the hit ratio of the R-UTP algorithm first increased and then decreased, approaching a peak at α = 0.3. As the other two algorithms had no dependence on α, the results did not fluctuate. Figure 2(b) shows the variation of the hit ratio with changes in α for the CiteULike dataset. The overall trend was similar to that of Figure 2(a), except the peak value of α was different, at α = 0.4. Because of the high density of the TripAdvisor dataset, its average hit ratio was higher than that of the CiteULike dataset.

The recall ratio experiment was also performed on the TripAdvisor and CiteULike datasets, with the recall ratios obtained for different |P| values being tested at α = 0.3 and 0.4, respectively. The experimental results are shown in Figure 3.

Fig. 3

Figure 3 shows the change in recall ratio in accordance with the number of recommended projects |P|. In Figure 3 (a) and (b), α is set to 0.3 (TripAdvisor) and 0.4 (CiteULike), respectively. It is obvious that the recall ratio gradually increased with increasing |P|. Compared with the UserKNN and ItemKNN algorithms, the R-UPT algorithm has obvious advantages in terms of the recall ratio parameters. Note that the TripAdvisor dataset has a higher average recall ratio than the CiteULike dataset because of its high data density.

In addition, experiments on the novelty of recommendation were also performed. The novelty of recommendation refers to the maximum possible recommendation of products that are not well known to users, in addition to recommendations based on tourist interests. As industrial tourism is poorly recognized by the public at present, novelty should be an important parameter in the proposed recommendation system. The recommendation novelty was calculated as follows: (7) Novelty=1nL∑i=1n∑pr∈pRik(pr) {\rm{Novelty}} = {1 \over {nL}}\sum\limits_{i = 1}^n \sum\limits_{{p_r} \in p_R^i} k\left( {{p_r}} \right) where k (p_r) represents the number of tag nodes, pRi p_R^i represents the list of recommended tourism products for user u_i and L=|pRi| L = \left| {p_R^i} \right| is the number of tourism products recommend for user u_i.

The experimental results regarding the novelty of recommendation are shown in Figure 4. It is apparent that increasing α corresponded to a monotonous decrease in novelty; this means that the larger the number of labels, the better the recommendation novelty. Obviously, the novelty is also related to the length of the recommendation list (i.e. L). The longer the list, the more likely the algorithm is to generate a new recommendation item.

Fig. 4