Data source description and visualization tasks
This paper takes West Lake as an example to verify the effectiveness of the proposed method. The experimental data were provided by the West Lake administration, including the real-time number of tourists and the registration data of tourists when they enter the park, including basic tourist information, historical visit records and satisfaction evaluations. After data cleaning, a total of 5000 available data points were obtained. The above data were used to test the effectiveness of the tourist evacuation route recommendation method proposed in this paper and to perform the following visualization tasks.
1.
Visualization analysis of real-time carrying capacity warning for the West Lake scenic area.
2.
Visualization analysis of the style clustering results of West Lake scenic spots.
3.
Recommendation result analysis of evacuation routes for tourists in the West Lake scenic area.
4.
Visualization analysis of the evacuation efficiency of West Lake scenic spots (by adjusting the PP and PR parameters).
5.
Visualization analysis of tourist portraits of West Lake scenic spots.
Comparison experiments
To verify the effectiveness of the proposed PER-GCN in “The proposed evacuation route recommendation method” section, the PER-GCN algorithm is compared with the GCN (original algorithm) and collaborative filtering algorithm based on users (uCF) (classical algorithm) for comparison experiments. The experiments are run 30 times under an NVIDIA GeForce GTX 1060 5 GB environment. All algorithms are cross validated on the West Lake dataset with ten iterations, and the mean and variance are taken as the final experimental results of each algorithm with the precision rate and recall rate as the evaluation indices. The statistical results are shown in Table 3.
Table 3 Experimental results comparison.
To further test the significance of the comparison results, we used the pairwise (one-sided) Wilcoxon’s signed-rank test with a significance level \(\alpha \) = 0.05 to compare the PER-GCN with uCF and GCN. The test results are shown in Table 4.
Table 4 Wilcoxon’s signed-rank test was performed with a significance level of \(\alpha \) = 0.05.
The results in Tables 3 and 4 show that the PER-GCN is significantly better than the GCN and uCF. The reasons are analysed as follows.
1.
Although uCF is a classical method, it also has an obvious sparsity problem. With data scale expansion, the proportion of data rated by users in the overall database decreases, the sparsity of the user-rating matrix increases and becomes more serious, leading to significant decreases in accuracy when calculating the nearest neighbours of users or items, thus making the recommendation quality of the recommendation system drop sharply and the accuracy of mining information insufficient.
2.
GCN uses the matrix decomposition technique, which is one of the solutions to the sparsity problem. The method expresses the relationship between nodes with the adjacency matrix and then decomposes the matrix to obtain the required embedding vector. Graph embedding represents the nodes in the graph as low-dimensional, real-valued, dense vector forms so that the obtained vector forms can represent and reason in the vector space; such vectors can be used in specific downstream tasks. Representation of the entire graph as a low-dimensional, real-valued, dense vector form is used to classify the entire graph structure. This algorithm reasonably increases the effectiveness of the algorithm. The recommendation algorithm based on GCN performs a secondary fusion operation on the existing results of GCN again. A simple dimensionality reduction method is adopted to set conditions to remove some users who have not participated in the rating or have rated very few times to reduce the dimensionality of the user-rating matrix and obtain a more accurate user-rating matrix. Then, the user-rating matrix was added to the model for fusion for a second time to further improve result accuracy.
3.
Based on the advantages of the GCN, our method further considers the similarity of tourism styles. It uses the GCN to calculate tourists’ preferred scenic spots based on their historical travel routes, and then mining a list of scenic spots with similar styles through tourism style similarity calculation will further obtain tourists’ preference information. The GCN-PER algorithm will help alleviate cold start and sparsity issues, thus improving the recall and accuracy compared to the GCN algorithm.
Visualization analysis results
The real-time bearing capacity warning of the West Lake scenic spot
In the visualization system, different colours are used to show the tourist flow and travel style of the scenic spots. For tourist flow, light green, dark green, yellow, orange and red were used to show the real-time tourist density. The colour from light green to red with tourist density increased, as shown in Fig. 4. For example, real-time carrying capacity warning shows that the tourist flow of Zhejiang Art Museum, Qu Yuanfeng, and Yuewang Temple scenic spots exceeds the carrying capacity. The real-time carrying capacity warning is displayed in red, while there are fewer tourists from Wansong Academy and Jingci Temple with similar travel styles. In our proposed method, tourists to the Zhejiang Art Museum, Qu Yuanfeng, and Yuewang Temple scenic spots are recommended to go to Wansong Academy and Jingci Temple. For each scenic spot in Fig. 4, clicking on a certain scenic spot will display its ID and name.
Figure 4
The visualization result for West Lake’s real-time carrying capacity.
Visualization analysis of the clustering results of scenic spot styles
According to the method introduced in “Experimental results and visualization analysis” section, K-means was used to obtain clustering results for 36 scenic spots in West Lake, as shown in Fig. 5a. Based on the style scores of historical buildings, museums, art galleries, parks, natural landscapes, religious buildings, and modern landscapes, the location of 36 scenic spots and their style are displayed in the scatter plot shown in Fig. 5b. The seven types of scenic spots (classic sights museums temples mountains modern land art gallery park) are represented in orange, yellow, purple, red, cyan, green, and blue, respectively. The colours are clear between each category, and the difference is obvious so that tourists can easily observe the distribution of similar scenic spots. It is also allowed to select any style of attraction on the map to view tourist portraits.
Figure 5
The visualization results of tourism style in the West Lake scenic area.
The information extracted from the clustering results of the tourism style in the West Lake scenic area is used to obtain the style feature information and make labels for the scenic spots; then, using the scenic spot labels to mine the traveller’s preferred travel style, it can be found that Qu Yuan Feng He and Wan Song Shu Yuan, Yue Wang Temple and Jing Ci Temple have similar styles. Therefore, when the number of tourists in popular scenic spots exceeds the carrying capacity, the other scenic spots with similar styles will be recommended. Compared to the limited access and traditional evacuation methods, the proposed personalized recommendation method based on tourist preference can bring a better travel experience.
Visualization analysis of recommendation results based on different parameter settings
By adjusting the control parameters PR and PP (PR is the weight of route complexity, PP is the weight of scenic spot carrying capacity) in Equation (15), the recommended results of 36 scenic spots in West Lake are analysed.
Figure 6
The recommendation results based on different PP and PR settings.
Figure 6 shows three bar charts displaying the recommendation scores for each scenic spot by adjusting PR and PP and comparing them with PR = PP = 0. When PR and PP are set to 0, only tourist preferences are considered to calculate the recommendation score. When PR is increased, the recommended score further considers the route complexity, and when PP is increased, the recommended score further considers the carrying capacity. Figure 6 shows the recommendation score for a tourist located in a scenic spot with ID 18 (Leifeng Xizhao). When PR = PP = 0, the scenic spot with ID 35 (Lingyin Temple) has a high recommended score (see green bar), but when PR is set to 0.3, the recommended score is significantly decreased (see blue bar). The reason is that Lingyin Temple is far from Leifeng Xizhao, which can be found in Fig. 5b. When PP = 0 and PR = 0.3, the scenic spot with ID 35 (Lingyin Temple) has a higher recommended score (see red bar). The reason is that Lingyin Temple has a greater carrying capacity, which can be found in Fig. 4. Therefore, when PP and PR are set to 0.3, the recommendation results account for both route complexity and carrying capacity. Therefore, the height of the orange bar is between the heights of the blue and red bars; that is, the recommendation scores are in the middle of the two.
Visualization analysis of evacuation efficiency based on cooperation degrees and evacuation batches
When both PP and PR values are set to 0.1, the evacuation efficiency is analysed based on the willingness of tourists to cooperate and the evacuation batch, and a river map is drawn accordingly, as shown in Figs. 7 and 8.
Figure 7
Changes in the river map with different settings of tourist willingness to cooperate.
Figure 8
Changes in the river map with different evacuation batch settings.
As shown in Figs. 7 and 8 the higher the number of cooperating tourists (willingness to cooperate), the more unstable the river map; the fewer evacuation batches, the more unstable the river map. The reasoning is as follows:
1.
The willingness to cooperate reflects, to a certain extent, the current flow of people during evacuation activities (willingness to cooperate * number of batches = number of people moving). However, when the willingness to cooperate is high, the flow of people fluctuates greatly, making the river map more unstable. Therefore, in our method, a certain number of tourists unwilling to cooperate can be tolerated, which is more in line with the actual situation and helps to improve evacuation efficiency.
2.
For evacuation batches, it is equivalent to the update interval in the model. When the number of people is fixed, more batches indicate fewer people in a single evacuation activity, which can make the river map smoother and have fewer fluctuations.
Based on the above analysis, it can be concluded that the optimal parameters in the West Lake tourist evacuation recommendation system should be set as follows: when the willingness of tourists to cooperate is within the range of 60% -75%, and the batch setting is 10 or more times, the evacuation efficiency is higher.
Visualization analysis of tourist portraits in the West Lake scenic area
In this visualization system, radar chart is used to display the tourist portraits of each scenic spot and each tourist, as shown in Fig. 9. The tourist portraits can be switched to view all scenic spots and tourists by clicking the location (or ID) of the scenic spot in Fig. 3(2) or (3) and choosing a tourist ID number in Fig. 3(4).
In Fig. 9, there are five indicators to describe the tourist portrait, including sex, age, tourism type, athletic ability and consumption ability. The data come from tourists’ registration information and their historically visited scenic spots. In Fig. 9, the red area shows the tourist portrait of tourists, and the green area shows that of scenic spots, where the five scenic spot indicators are obtained by calculating the average value of tourist information that has visited the scenic spot. The range of sex is [0,1], where 0 represents male and 1 represents female. The age range is [15,65], and each grid represents a 10-year increase. The range of consumption capacity is [1000,6000], and each grid represents an RMB1000 increase. The consumption capacity refers to the acceptable travel expenses for tourists in West Lake. The range of athletic ability is [5000, 15,000], referring to the daily steps recorded in their mobiles. The scope of tourism type ranges from knowledge type to experience type, obtained from the style vectors of scenic spots and tourists in Table 1 and 2. Generally, the styles of the historical buildings, museums, art galleries, religious buildings, parks, natural landscapes, and modern landscapes are developed from knowledge type to experience type, and the value is from 1 to 0. Therefore, the highest score in each row of Tables 1 and 2 represents the tourism style of each scenic spot and tourist. Therefore, the tourist portrait of tourist (ID 2) is male, 53 years old, acceptable travel expenses are 3500, and the average daily steps are 9773. The tourism type is knowledge oriented. Based on that, the Tomb of Su Xiaoxiao and Hangzhou Garden tourist portraits can be analysed as follows:
1.
The tourist portrait of “Su Xiaoxiao Tomb” has the following characteristics: the tourism type tends to be knowledge style, the tourists are relatively older in age, the majority are male, the consumption ability belongs to the luxury type, and athletic ability is moderate.
2.
The tourist portrait of “Hangzhou Garden” has the following characteristics: the tourism type tends to be experience style, the age and gender are relatively average, the consumption ability belongs to the economic type, and athletic ability is relatively weak.
Figure 9
The tourist portraits of different scenic spots (taking the Tomb of Su Xiaoxiao and Hangzhou Garden as examples).
The visualization results of tourist portraits are beneficial for the West Lake administration to analyse the characteristics of scenic spots and tourists and provide specific customer services for different scenic spots.
