Forecasting Inter-Destination Tourism Flow via a Hybrid Deep Learning Model

Many tourists go multiple destinations in one trip, and this phenomenon is displayed directly by ITF. A lot of research in tourism management are conducted based on ITF, since it implies things like the relationship between destinations, the tourist behaviour preference, and the structure of multi-destination network.

From the perspective of directions, the ITF can be devided to directed ITF and undirected ITF. The directed ITF from destination A to destination B refers to the number of tourists that visit B right after visiting A in the same trip; While the undirected ITF between two destinations A and B refers to the number of tourists that visit both A and B in the same trip.

In previous study, the directed ITF are used to predict tourism flow over time[22], forcast tourism demand[8], detect tourists’ mobility pattern[6], identify boundary effect[15] and identify common routes[4]. Another common way of study using directed ITF is to build the multi-destination network whose edges are directed ITF, and then calculate the node structure indicators(i.e. node-centrality indicators and structural holes indicators) and network structure indicators(i.e. size, network centralization, etc.).The calculated indicators can be used to classify roles of destinations [4, 15] ,study spatial structure of multi-destination network[5] or compare visitation patterns of tourists with different characteristics(e.g. different origins, purpose, visit duration, e.t.c.)[6].

The undirected ITF, on the other hand, is also commonly used for calculating the node structure and network structure indicators, and then doing the same tasks on the basis of calculated indicators as is done on indicators calculated by directed ITF [23, 17, 7], the only difference is that the multi-destination networks built with undirected ITF are symmetric compared to the networks built with directed ITF. In other word, in multi-destination networks built with undirected ITF, the edge weight from destination i to destination j is the same as that from j to i. Besides calculating the indicators, other methods are also applied to make use of undirected ITF. For example, Xu, Li et al. (2021)[25] uses a community detection algorithm on the multi-destination network to identify seven destination communities in South Korea; Hwang, Gretzel et al.(2006)[7] used clique detection to identify strongly-connected destinations in America. In another research by Liu, Huang et al.(2017)[9], the Quadratic Assignment Procedure (QAP) was used to test the relationships between region proximity, grade proximity, tenure proximity, and the destination network determined by tourists’ free choice movements. Plus, although the undirected ITF has no explicit direction, it can also be used to identify the Directionality of multidestination patterns by using conditional probability comparison[7].

From the perspective of granularity, the ITF is divided into two categories, the inter-city level and the inter-attraction level,¹¹1In some articles like [17], the inter-attraction level is referred to as within-destination level and the inter-city level is referred to as inter-destination level. In studies of inter-city level ITF [25, 7, 5], the whole region containing multiple cities is regarded as a multi-destination system to be studied. While in the studies of inter-attraction level ITF[17, 9, 22, 8, 6, 19, 15], a spatially continuous region(e.g. a city, a province, a river delta) containing multiple tourism attractions is regarded as a system to be studied.

In previous studies, the ITF mainly comes from four data sources, i.e., mobile positioning data, GPS data, survey data, and social media data. The mobile positioning data could capture the location footprints of large populations and is precise and fine-grained. However, such data are difficult to acquire due to privacy issues, and only a few researchers who have access to such data can do research based on it[25, 26]. The GPS data are usually collected by location tracking applications installed on volunteers’ phones, which means it can only cover a very limited number of tourists.[6] The survey data is the most widely used way to collect ITF, but conducting surveys is time-consuming and still can only cover a little number of tourists.[5, 4, 19, 7] Another commonly used data is social media data. One kind of social media data is Volunteered Geographic Information (VGI) data used in [8, 22]. It is data published by users and include embedded geographical metadata, like geotagged tweets on Twitter and photos on Flickr. Another kind of social media data used for extracting ITF is tourism notes[17, 23, 15], which, compared to VGI data, are more targeted on tourists. Although being easy to get and covering large crowds, the social media data can be sparse [11],i.e. not all cities and places have plenty of social media data to be used, and people may not report all the places they visit on social media. Thus, predicting missing ITF is important when accurate volume of ITF is not available.

By analyzing ITF, previous studies have found several factors that would influence its volume. Since the factors concerned for inter-city level ITF and inter-attraction level ITF vary a lot, here we mainly discuss the latter, which is the focus of our study. From the perspective of tourists’ characteristics, [17] shows that tourists from different original countries result in different ITF patterns; [6] shows that factors like religion, visiting times(i.e., repeated visitor, first visitor), age and purpose of tourists also influence ITF. And [26] finds that the ITF patterns vary in different duration of stay and lengths of tour trajectory of individual tourists. From the perspective of tourism attractions’ characteristics, [1] finds that there is a visitor flow spillover effect on neighboring attractions, meaning that the distance between attractions would influence the ITF between them; [9] shows that the proximity of two attractions in their region, level, and time of getting the level would influence the ITF between them; [15] finds that the provincial-administrative boundary would have a boundary-shielding effect on the ITF between attractions that belong to different provinces; and [3] finds that the diversity of multi-attractions and the perceived costs would influence the ITF. However, none of the study integrate all of the characteristics of attractions mentioned above together; and the popularity of attractions was not considered; Besides, although regression models like logistic regression, polynomial regression, and QAP are used to regress the ITF; their main purpose was to explain the influence of independent variables instead of predicting ITF precisely, thus, their model’s performance in predicting is limited.

The prediction of ITF is similar to the commuting flow prediction problem, which is to predict realistic commuter flows among locations, given their characteristics and the distance among them, and without any knowledge about the real flows. Traditionally, the gravity[12] and radiation models[32] are the most commonly used ones for this task. The gravity model is based on the assumption that the number of travelers between two locations increases with the locations’ population while decreasing with the distance between them; the Radiation model replaces the distance feature in Gravity model with the intervening opportunities, which considers more variables. However, the traditional formula-based models typically have a strict structural form and a limited number of input variables, which limits their ability to predict commuting flows.

With a better ability to capture nonlinear relationships, machine learning models are used for the commuting flow prediction problem. For example, Morton et al.[14] find that the XGBoost model performs better on this problem compared to traditional models; [16] compare the gravity model, neural networks, and random forest model on Twitter data for predicting commuter flow, and find that the Random Forest model has the best performance. [20] put forward a deep gravity model with good interpretability that improved the performance of the traditional gravity model a lot.

Recently, another type of learning model, Graph Neural Network (GNN) based models, has been used for the commuter flow prediction problem. Liu et al. [10] put forward a GAT(GNN with Attention) model, GMEL, that generates two embeddings for every node on a geo-adjacency network, and then uses Dismult as the score function. The model effectively captured the spatial correlation from geographic contextual information. [28] put forward a GCN(Graph Convolutional Netxork) model, SI-GCN, with GCN as encoder and Bilinear as score function, which reaches a good performance on the T-Drive dataset. Yin et al.[29] put forward a GCN model with a multilayer perceptron as the score function and a random forest as a refinement layer on the task, which effectively improved the prediction accuracy. Most recently, Yin et al. [29] put forward a GCN model, ConvGCN-RF, with multilayer perceptron as a score function and a random forest as a refinement layer on the task, which effectively improved the prediction accuracy.

However, applying the above models directly on the ITF prediction problem has some drawbacks. For non-GNN-based models like deep gravity or Random Forest, they cannot capture the interaction relationships of multiple(i.e., more than two) attractions as a system. For example, if we already know that two attractions, $A_{i}$ and $A_{j}$, has a strong interaction, which means many people choose to visit both of them in one trip; and another attraction, $A_{k}$, is on the way from $A_{i}$ to $A_{j}$. Then we may guess many visitors would also visit $A_{k}$ as a “drop by visit” when they go from $A_{i}$ to $A_{j}$. Thus, the volume of ITF between $A_{i}$ and $A_{k}$ may also be large. As for the GNN-based models, the GMEL must be applied to continuous space, while the tourism attractions are discretely scattered areas in the city. The SI-GCN model can use the coordinates as input features to take the relative positions of attractions into consideration, but this method could not take consideration of the distance feature explicitly. While the RFGCN, which uses an MLP as the score function, could not explicitly consider the corresponding features of $A_{i}$ and $A_{j}$ jointly when predicting the ITF between them.

First, we build an interaction Graph G. We get the features (like location, area, ticket price, etc.) of each tourism attraction. For categorical features, we use ordinal encoding to convert them to numerical forms. Then we normalize all the features into the scale of 0 to 1 and concatenate them to get the feature vector of every attraction. The edges for G is gotten by measures mentioned in the definition of Interaction Graph in section 3.

The model we use in this research is similar to the Spatial Interaction Graph Convolutional Network (SI-GCN) model proposed by [28]; the difference is that we add an extra Random Forest Layer as a refinement layer on top of the SI-GCN architecture so that the distance feature can be considered explicitly; Thus, we call our model SI-GCN-RF(Spatial Interaction Graph Convolutional Network with Random Forest refinement) As Figure 1 shows, the SI-GCN-RF model consists of three parts, the Encoder, the Decoder, and the Refinement layer.

Since the GNN-based models incorporate both graph structure and feature information, they lead to complex models and predictions made by GNNs is hard to explain [30]. However, non-GNN-based models like random forests can be easily explained by existing explainable AI frameworks such as SHAP. So to account for how different features influence the volume of ITF, we first train a Random Forest(RF) model to predict ITF from existing features, than we use SHAP to explain the trained RF model.

SHAP applies a game theoretic approach to explain the output of machine learning models. SHAP values are used to determine feature importance and whether it influences the predicted result negatively or positively. The interpretation of the SHAP value $\phi_{j}$ for variable value j is: the value of the $jth$ variable contributed to the prediction of a particular instance compared to the average prediction for the dataset. For example, if we find that for variable ”distance”, when its value is large, its SHAP value is negative; and when its value is small, its SHAP value is positive. We can determine that the model has learned that the distance feature has a negative correlation with the predicted ITF.

We compare the proposed model SI-GCN-RF and all baselines in terms of the prediction performance (i.e., MAPE, MSE and CPC). Table 2 reports the model comparison results. Since the MSE is an absolute metrics and thus has a strong bias towards instance with larger true value, it is least representative when evaluating the model’s performance. So we mainly use MAPE and CPC when analyzing the performance in the following parts. From the results, we observed that the models with a Bilinear scoring function after node embedding have better performance with CPC all larger than 0.85 and MAPE all less than 1.3, compared with other models whose CPC are all less than 0.7 and MAPE larger than 1.95. That is because, the Bilinear scoring function for node embeddings successfully extracts the collaborate influence of corresponding features and also maintains the symmetric feature of ITF. Among these, the SIGCN-RF with both massage passing in Interaction graph, self-updating, Bilinear scoring function and Random Forest Refinement layer has the best performance, with an MAPE of 0.46 and CPC of 0.89. The performance of SIGCN-RF(no edge) is worse than SIGCN-RF with MAPE larger by 0.2, which means the interaction graph successfully captures the interaction intensity features of multiple attractions, thus increases the prediction accuracy. When comparing the SI-GCN with SIGCN-RF, we see that the Random Forest Layer successfully improves the performance of the model by decreasing the MAPE by over 0.7. That is because, the random forest layer has the distance feature input explicitly, while the SI-GCN can only consider the distance by implicitly using the coordinate features as node vectors. For non-GNN based models, the random forest model has a significantly better performance than the deep gravity with CPC increased by over 0.3, which is also one of the reasons of why we choose RF for explaining features. Another interesting thing is that the GCN-RF model with MLP as its scoring function has the worst performance, which may because that the MLP is not suitable for predicting undirected ITF with volume from $A_{i}$ to $A_{j}$ equal to that from $A_{j}$ to $A_{i}$. This phenomenon may also partly explain the limited performance of the deep gravity model, whose backbone is also MLP.

A visualized result of the predicted ITF compare with the true ITF is presented in figure 5 with quantile classified symbology. Besides, we notice that the relative prediction accuracy on larger true ITF values are better, as is shown in figure 7.

When using SHAP to explain the influence of different attractions’ features on ITF in the RF model, we get the SHAP values for every attraction feature in table 1, in every prediction iteration. The result is shown in figure 7, with the features sorted by their maximum SHAP values.

From the result shown in figure 7, we have four observations. First, popularity, quality, and distance are the main influential factors in the predicted volume of ITF for the RF model. Second, the comment number of attractions, which represents the popularity of attractions, and the level and average ranking, which represent the quality of attractions, have a explicit positive relation with ITF. This is because people tend to go to attractions that are popular and have a good reputation. Third, the distance of attractions has a negative relation with ITF, the nearer the two attractions are, the larger the ITF between them is. This verified that there is a spill-over effect of neighboring tourism attractions. Fourth, high ticket price usually leads to low ITF. That may be because, the attractions with a high ticket price are usually amusement parks or Zoos & Arboretums as is shown in figure 9, compared to free attractions, fewer visitors may be willing to pay a high price to those places. Also, compared to historical sites or city sightseeing spots, these attractions show fewer unique characteristics of the city, thus are less possible to attract visitors.

We also examine the SHAP for features that shows no implicit influential tendencies in figure 7 (i.e. type, coordinates, area, adname, mean_time) in detail. Note that in the RF model, when predicting the ITF between two attractions, their feature vectors are concatenated and then fed into the model, thus, a feature is represented by both the feature of $Attraction_{i}$ and $Attraction_{j}$, (like the $ticket\_price$ feature is represented by both $ticket\_price_{i}$ and $ticket\_price_{j}$), we include the SHAP value of the same feature for both $Attraction_{i}$ and $Attraction_{j}$ when analyzing a particular feature like ticket_price. When we examine the SHAP value for feature ”type” in figure 9, we can find that historical sites have an explicit positive influence on ITF, whereas amusement parks, exhibition center & museum has an explicit negative influence. When observing the SHAP value for coordinate features as is shown in figure 10, we find an interesting phenomenon. When the longitude of attraction is between 116.396E and 116.438E; and the latitude is between 39.93N to 39.98N, the coordinate will have a positive effect on the ITF. If we map these areas on the map, as is shown in figure 12, we can see that this area is the NorthEastern area of downtown Beijing, with a lot of neighboring Hutongs (narrow lanes or alleyways in a traditional residential area of a Chinese city) there.

Another interesting phenomenon is that if we examine the SHAP value for feature ”level”, we can see from figure 12 that only the 5A attractions will have an explicit positive influence on ITF. In China, the 5A attractions are rare attractions with high value, and they are ranked by experts strictly. Their high quality is contributing to a large ITF that travels to them.

However, the adname, estimated visiting time, and area of attractions have no explicit tendencies found.

Our proposed SI-GCN-RF model can be modified to predict the directed ITF. To achieve this goal, the structure of the encoder can be the same, with the Interaction Graph G as a directed binary graph, and the $N(v)$ in the encoder meaning the neighboring vectors who have an edge connected to the node $v$. In the problem of predicting directed ITF, the scoring function in the decoder can be modified to be a Dismult:

where the $f(u,v)$ is the predicted ITF from node u to node v; $E^{u}$ and $E^{v}$ are the embedding vector of node u and v respectively, and $R$ is a learnable square matrix. Since $R$ is not necessarily a symmetric matrix, the Dismult decoder can allow the flow from node $u$ to node $v$ different than that from node $v$ to node $u$. We have not tested this extended model due to the data limitation. Tourism notes are not necessarily written in the order of visiting, thus are more suitable for undirected ITF. The test of the extended model or other better models on directional ITF prediction is left to future works.

We should note that SHAP can only explain explicit features like the distance between $A_{i}$ and $A_{j}$ and the features defined on individual attractions like the ticket price. However, it cannot consider the implicit features of the structure of the interaction graph nor the joint effect of corresponding features.

Recently, several explainable AI methods on GNN have been put forward[30], and they may be able to explain the prediction models of ITF with GNN involved. We also left it to be a future research focus.