A Socially Aware Huff Model for Destination Choice in Nature-based Tourism

. Identifying determinants of tourist destination choice is an important task in the study of nature-based tourism. Traditionally, the study of tourist behavior relies on survey data and travel logs, which are labor-intensive and time-consuming. Thanks to location-based social networks, more detailed data is available at a ﬁner grained spatio-temporal scale. This allows for better insights into travel patterns and interactions between attractions, e.g., parks. Meanwhile, such data sources also bring along a novel social inﬂu-ence component that has not yet been widely studied in terms of travel decisions. For example, social inﬂu-encers post about certain places, which tend to inﬂu-ence destination choices of tourists. Therefore, in this paper, we propose a socially aware Huff model to account for this social factor in the study of destination choice. Moreover, with ﬁne-grained social media data, interactions between attractions (i.e., the neighboring effects) can be better quantiﬁed and thus integrated into models as another factor. In our experiment, we calibrate a model by using trip sequences extracted from geotagged Flickr photos within two national parks in the United States. Our results demonstrate that the socially aware Huff model better simulates tourist travel preferences. In addition, we explore the signiﬁcance of each factor and summarize the spatial-temporal travel pattern for each attraction. The socially aware Huff model and the calibration method can be applied to other ﬁelds such as promotional marketing

Abstract. Identifying determinants of tourist destination choice is an important task in the study of naturebased tourism. Traditionally, the study of tourist behavior relies on survey data and travel logs, which are labor-intensive and time-consuming. Thanks to location-based social networks, more detailed data is available at a finer grained spatio-temporal scale. This allows for better insights into travel patterns and interactions between attractions, e.g., parks. Meanwhile, such data sources also bring along a novel social influence component that has not yet been widely studied in terms of travel decisions. For example, social influencers post about certain places, which tend to influence destination choices of tourists. Therefore, in this paper, we propose a socially aware Huff model to account for this social factor in the study of destination choice. Moreover, with fine-grained social media data, interactions between attractions (i.e., the neighboring effects) can be better quantified and thus integrated into models as another factor. In our experiment, we calibrate a model by using trip sequences extracted from geotagged Flickr photos within two national parks in the United States. Our results demonstrate that the socially aware Huff model better simulates tourist travel preferences. In addition, we explore the significance of each factor and summarize the spatial-temporal travel pattern for each attraction. The socially aware Huff model and the calibration method can be applied to other fields such as promotional marketing.
Keywords. nature-based tourism, socially aware Huff model, tourist destination choice, geotagged social media, Flickr

Introduction
Nature tourism, i.e., tourism that is based on the natural attractions of an area, has gone through rapid growth over the past two decades (Balmford et al., 2009), especially for national parks in the United States, according to visitation statistics by National Park Service. 1 Identifying and evaluating relevant determinants of tourist flows is important. On the one hand to promote tourism, and on the other hand it helps to protect natural lands. Prior work on nature-based tourism relies on manually collected travel logs and survey data, which are time-consuming, labor-intensive, and limited in temporal coverage (Puustinen et al., 2009;Nahuelhual et al., 2013). The emergence of locationbased social networks (LBSNs) and volunteered geographic information (VGI), such as Flickr, Instagram, Facebook etc., together with geotagging technology, provides more fine-grained spatial and temporal data, which equips us with a new lens to understand travel patterns as they relate to natural attractions.
Additionally, LBSNs play an increasingly important role in travel decision making process (Leung et al., 2013). For example, places like Horseshoe Bend, Devil's Bathtub, etc., once being hidden gems, are now receiving a large number of visitors annually. Social media has been regarded as the main culprit for the sudden and overwhelming popularity of these places (Djossa, 2019). More specifically, geotagged photos posted by social media influencers (SMIs) can rapidly attract new visitors (Glover, 2009). These influencers are usually users with a large number of followers and have established credibility in certain fields that can shape attitudes of tourists and thus influencing their travel preferences (Freberg et al., 2011;Li, 2016). Intuitively, a scenic photo posted by a user with 50k followers has a much broader potential influence than a user with 50 followers. Therefore, we argue that social factors brought by increasingly used social media need to be taken into account as a new norm to complement traditional destination choice models. To justify such an argument, we specifically explore this social effect in nature-based tourism destination choices, because tourists tend to share geotagged photos on social media platforms along their trips (Tasse et al., 2017). Moreover, existing work has shown that fine-grained spatio-temporal data collected from social media can 1 https://irma.nps.gov/Stats/ be used to quantify visitation rates (Wood et al., 2013), to estimate visitor flows (Orsi and Geneletti, 2013;Kim et al., 2019), and to detect popular sub-regions and temporal activity patterns (Heikinheimo et al., 2017). These studies illustrate the capability of using LBSNs to capture temporal variations in tourist visiting, with some places (e.g., dive resorts) being more attractive in summer and others (e.g., ski resorts) in winter. In addition, interactions between places (i.e., the neighboring effects) can be better quantified with social media data. For example, we can estimate the interactions between Horseshoe Bend and those attractions in its surroundings (Glen Canyon, Antelope Canyon, Grand Canyon, etc.), based on which we can further explore how they affect potential travel decisions of tourists to Horseshoe Bend, thereby uncovering interesting travel patterns that are difficult to detect using traditional data.
To explore how social factors and neighboring effects contribute to tourist destination choice in natural attractions, we propose a socially aware version of the well-known Huff model (Huff, 1964), which was originally used to calculate the probability of a customer shopping at each retail store.
The contributions of this work are as follows: • We propose a socially aware Huff model, which incorporates both social factors and neighboring effects, to estimate the probability of tourists visiting specific places.
• The proposed method is calibrated on a data set containing 10-year geotagged Flickr photos in two national parks, whose results outperform the baseline Huff model.
• We explore the spatial and temporal variability of model parameters that are associated with attractiveness, distance, and neighboring effect in the socially aware Huff model.
The remainder of this paper is organized as follows. Section 2 introduces related work on tourism, geosocial media, and the Huff model. Section 3 briefly introduces the Flickr data set used for the study and explains the trip reconstruction process. A socially aware Huff model is introduced in section 4 together with the model calibration method. In section 5, we present the model calibration results and explain the spatial and temporal variability of the parameters used in the socially aware Huff model. Finally, we summarize our findings and discuss future directions in section 6.
2 Related Work 2.1 Tourism and Geo-social Media The use and role of social media has been widely discussed in tourism research (Leung et al., 2013). Work by Zheng et al. (2012) used Flickr data to discover regions of attractions (RoAs) and explored tourists movement patterns in relation to the RoAs. Similarly, Hu et al. (2019) extracted popular attractions and tour routes using a graph-based network in New York City from Twitter data. Majid et al. (2013) proposed a context-aware personalized travel recommendation system and evaluated it based on a Flickr data set. Li et al. (2018) used Flickr data to compare the spatial overlap of tourists' and locals' destinations in ten US cities. Work by Mou et al. (2020) analyzed spatiotemporal distribution and changes of inbound tourism flow in Shanghai with Flickr data.
In the past decade, social media has evolved into an important player in tourism advertising and promotion (Bakr and Ali, 2013). Litvin et al. (2008) showed that travelers are increasingly influenced by electronic Word-of-Mouth (eWOM) from social media. Parsons (2017) echoed the similar idea that Instagram influences tourist decision-making, especially for younger generations. Jalilvand and Samiei (2012) examined the influence of eWOM and showed that it has a significant impact on tourist attitudes towards visiting Isfahan, Iran. Tham et al. (2020), however, conducted interviews with tourist decision-makers in Australia and revealed that social media's role appears to have only moderate-low influence on destination choice. In line with such research, we include a social influence factor and examine the impact of social media on destination choices. We quantify the social impact that influencers could bring to a place by measuring the place attractiveness given the travel preference of tourists.

Huff Model
There have been many research efforts towards tourist destination choice and sequential tourist flows (Nico-lau and Más, 2008;Wu et al., 2012;Yang et al., 2013). The Huff model (Huff, 1964) is one of them, though it was originally developed to predict retail sales and consumer behavior. The Huff model estimates the probability of consumer patronizing retail stores based on two factors: attractiveness of a store and travel cost, which can also be applied to tourism research. Misui and Kamata (2016) adopted the Huff model to show the effect of travel time on visiting probability to spa destinations in Japan. Similarly, Nicolau (2008) studied tourist sensitivities to distance and price for destination choice in Spain using a national tourist choice behavior survey data. Yang et al. (2013) conducted a logistic model to study the inter-dependencies among destination choices when two or more destinations are included in a trip, accounting for the future dependency in the multi-destination choice behaviors. Recently, more work has shown the importance of the temporal factor that is missing in the original Huff Model. Gong et al. (2020) included weekday and weekend variations when calculating visiting probability of shopping areas using taxi trajectory data in Shenzhen and New York. Liang et al. (2020) proposed a T-Huff model and proved that it outperforms the original static Huff model when estimating temporal store visits using SafeGraph POI visits data. Likewise, we include the temporal factor in our study given the availability of social media data.

Data and Trip Reconstruction
In this section, we introduce the data set used for the study in section 3.1 and explain how we reconstruct trips from the geotagged photos step by step in section 3.2.

Data and Study Area
In this study, we collected geotagged Flickr photos of tourist attractions within national parks using Flickr's public API. 2 Two national parks -Acadia National Park and Yosemite National Park -have been selected from the top ten most visited national parks in the United States over the past decade, as reported by National Park Service. These geotagged Flickr photos were collected from January 1, 2010 to December 31, 2019. Each photo is associated with its metadata including photo ID, owner ID, taken date, latitude, longitude, title, and the number of views. The total numbers of the geotagged Flickr photos and unique users in the data set are summarized in Tab. 1.

Identifying attractions using HDBSCAN
Spatial clustering is widely applied to point pattern analysis such as hot spot detection. One of the most popular clustering methods is DBSCAN (Ester et al., 1996), which is a density-based clustering algorithm. It requires two parameters: search radius ( ) and minimum number of points (minPts) within the search radius. Despite its broad applications, it is difficult to determine the in the original DBSCAN algorithm due to varying density distributions of points. In this paper, we adopt HDBSCAN (Campello et al., 2013;McInnes et al., 2017), a hierarchical density-based clustering algorithm, which addresses the aforementioned issue by using flexible values to identify attractions from the geotagged Flickr photos.
To identify a proper value for MinClusterSize, we compare different clustering results with the geographic distribution of top attractions listed on TripAdvisor 3 in the two national parks. Based on this comparison, Min-ClusterSize is set to 1% of the total number of photos, which is 349 and 504 for Acadia and Yosemite National Parks, respectively.
After applying the HDBSCAN algorithm, 13 clusters are extracted from Acadia National Park and 21 clusters from Yosemite National Park. We calculate the centroids of the clusters and label each cluster with the nearest attraction listed on TripAdvisor or Google Maps to its centroid coordinates. Fig. 1 shows the distribution of geotagged photos, clustering results, as well as the identified attraction names of Acadia and Yosemite National Parks.

Extracting trip sequences from geotagged photos
With attractions being identified, each photo in the cluster is labeled with an attraction name (or cluster ID). To extract trip sequences, we first group all photos by their owner ID and then sort them by the date taken. We consider a trip as a temporally-ordered sequence of photographed locations taken by the same user. Given the possibility that one user could make several trips to the area over the years, we set a time threshold λ t to distinguish these trips. If the time difference between two consecutive photos from the same user is larger than λ t , we separate them into two different trips. Here, we set λ t to 4 days, which is the average length of a stay in both national parks according to National Park Statistics. 4, 5 Thus, if a user took a photo at attraction A, attraction B, and then attraction C within the λ t constraint, we are able to capture this trip sequence as [A, B, C] based on the timestamp of each geotagged Flickr photo. For our data set, 1,949 trip sequences were extracted from the clustered geotagged photos in Acadia National Park, and 3,426 trip sequences from Yosemite National Park.

Calculating visiting probabilities from trip sequences
With the trip sequences extracted, we are able to construct a flow matrix based on the trip segments from all trip sequences. For example, [A, B] and [B, C] are two trip segments from the trip sequence [A, B, C]. The visiting probability is calculated proportional to the total number of outgoing trips for each attraction in the flow matrix. A monthly visiting probability matrix is also calculated in order to capture temporal factors in later computations. Fig. 2 visualizes the overall trip flows in the two national parks using flowmap.blue. 6 Further details of each attraction are provided in Tab. 8 and Tab. 9.

A Socially Aware Huff Model
In this paper, we leverage the Huff model (Huff, 1964) with multi-destination travel behaviors being taken into account (Stouffer, 1940;Um and Crompton, 1990). Fig. 3 is used to illustrate the neighboring effect in a multi-destination trip. In Fig. 3(a), a tourist at Origin O has two destination choices A and B, with equal distance to origin O. In this case, destination B should be preferred since it has more future choices in its neighborhood compared with destination A. Furthermore, Fig. 3(b) illustrates the effect of attractiveness. When destination A and B have the same number of future choices in their neighborhood and the same distance to Origin O, then intuitively the destination with more attractive future choices in its neighborhood would be preferred. Orpana and Lampinen (2003) used the term "store centralities" to model the effect of its neighboring outlets on a store's utility. The term can be interpreted as the possibility of interaction between a store and its neighbors (Hansen, 1959). The "centrality" concept is applied to model tourist destination choice as well, with the assumption that people tend to travel to places with more attractive future choices.

The Original Huff Model
The original Huff Model (Huff, 1964) is designed to estimate the probability of customers at each origin patronizing a given store among all stores as their destination choices. It takes two factors into account: attractiveness and distance. Attractiveness can be computed as a function of many attributes of a store, including the store size, number of parking spaces, customer reviews, etc. The classic form of the Huff model can be expressed as: where P ij represents the probability of a customer at location i visiting store j; A j is the measure of attractiveness of store j; D ij is the distance between location i and store j; and n indicates the total number of stores in the data set. The parameters α and β (α >0, β <0) are associated with the attractiveness and distance factors, respectively.

Socially Aware Huff Model
In this paper, we propose a socially aware Huff model to include social factor and neighboring effect, based on the assumptions that: (1) People tend to choose more attractive travel destinations; (2) People tend to choose closer travel destinations; (3) People tend to choose travel destinations with more beneficial future choices. Based on the original Huff model shown in Eq. 1, the socially aware Huff model can be expressed as: where P ijt represents the probability of a tourist at location i visiting attraction j at time t; A jt is the attractiveness of attraction j at time t; D ij is the distance between origin i and attraction j; C jt is the term used to describe the neighboring effect of attraction j, relative to other attractions at time t; and n indicates the total number of attractions in the area. The parameters α, β and θ are associated with the attractiveness, distance, and neighboring effect factors, respectively.
In the following, we explain how we quantify the three terms, i.e., A jt , D jt , and C jt , mathematically. Previous research has shown that the number of geotagged photos and the number of unique users can be used to represent the attractiveness of a place (Kádár and Gede, 2013;Leung et al., 2017). Here, we include three proxies to estimate the attractiveness A jt for later comparison. Log transformation is performed to address a right-skewed distribution of values. The three types of attractiveness A (l) jt , l = 1, 2, 3, can be expressed as: where M jt is the number of photos at attraction j at time t.

A
(2) jt = log(U jt + 1) where U jt is the number of unique users at attraction j at time t.

A
(3) where V kjt is the number of views for photo k at attraction j at time t. We use the product of the number of photos and the average number of photo views per user at attraction j to include a social influence factor. Given the fact that social media influencers (SMIs) have more followers than others, thus the photos they post would have more views and greater social impact, we include photo views per user here to account for potential existence of SMIs who upload photos at an attraction. We hypothesize that the attraction with more photo views per user is more attractive.
The term C jt , measuring the neighboring effect, can be modeled as: where K is the total number of nearest neighboring attractions being considered. C jt reflects the assumption that people tend to travel to places with more promising future choices in a multi-destination trip. We consider K-nearest neighbors of attraction j, calculating their attractiveness A (l) kt at time period t, and weight A (l) kt by their distance to attraction j, D kj . A higher C jt value is assigned to attractions with closer and more attractive neighbors. Finally, we define the term D ij as the estimated driving distance using the Distance Matrix API 7 from Google Maps.

Calibration Method
Parameters of the Huff model need to be calibrated before further studying the travel patterns. Here, we use the linear regression calibration method -Ordinary Least Squares (OLS), which estimates one set of parameters α, β, and θ, that best fit the model based on observations. The estimation process is executed by minimizing the sum of squared residuals in a linear model. OLS calibration returns fixed values for the parameters and assumes that they are homogeneous across the study area. The general form of OLS regression can be expressed as: where y is the dependent variable; x i is the i th independent variable; n is the number of independent variables; β i is the regression coefficient for the i th independent variable; and is the random error.
To conduct OLS, the socially aware Huff model in Eq. 2 is rewritten in a log-transformed-centered form, according to Nakanishi and Cooper (1974), in order to obtain the least square estimate of parameters: whereP it ,Ã jt ,D i and,C t are the means of P ijt ; A jt ; D ij and C jt over attraction j, respectively. For each origin attraction i, the model will estimate one best fit parameter set (α i , β i , and θ i ). 7 https://cloud.google.com/maps-platform/routes

Software and Data Availability
Data used in this paper can be accessed with the public Flickr API. 8 The query used to access the data, code and interactive data visualization (Fig. 2) are available on GitHub. 9 The workflow underlying this paper was partially reproduced by an independent reviewer during the AGILE reproducibility review and a reproducibility report was published at https://doi.org/10. 17605/OSF.IO/4CPM3.

Overall Calibration Results
In this section, we examine the overall calibration results for the two national parks and discuss the necessity of incorporating social factors and neighboring effects in the socially aware Huff model.

K-Nearest Neighbors
First we need to decide on the number of neighbors K in order to calculate the centrality C jt in Eq. 6. Values of K = 2, 3 and 5 are considered. For both Acadia and Yosemite National Parks, K = 2 gives the best performance, with the lowest mean squared error (MSE) and highest R 2 . More details are shown in Tab. 7. The following calibrations are subsequently all computed with K = 2 as the number of nearest neighbors in the centrality term C jt .

Social Influence
In Tab. 2, we show the calibration results using different measurements of the attractiveness factor, A (l) jt , expressed in Eq. 3, Eq. 4 and Eq. 5. Based on R 2 and Akaike information criterion (AIC), we observe that for both national parks, the attractiveness A (3) jt performs the best (highest R 2 and lowest AIC values), compared with the other two measurements (i.e., the number of photos and the number of unique users).  ∆AICi is a measure of each model i to the model with the minimum AIC. Models with ∆AICi < 2 can also be considered to have substantial support (Burnham and Anderson, 2002).
The results of Akaike weights (w i ), which can be interpreted as the probability that model i is the best model (Anderson et al., 2000), also show the same conclusion. Hence, we select A jt to estimate the attractiveness of an attraction, where the combination of photo views, the total number of photos, as well as the number of users are taken into account. The results indicate that including a social factor (i.e., the more photo views and potential social impact SMIs could bring to a place) can better simulate tourist preferences.

Temporal and Neighboring Effect Factors
To examine the overall performance of the temporal factor and neighboring effect in the socially aware Huff model (SA model), we compare it with SA model without the neighboring effect (SA model w/o N), SA model without the temporal factor (SA model w/o T), and the original Huff model (Huff model), whose results are shown in Tab. 3. The proposed SA model that includes both temporal factor and neighboring effect has the highest R 2 and lowest AIC values for Acadia National Park. As for Yosemite National Park, the performance of SA model and SA model w/o N are similar in terms of R 2 and AIC, while both models fit to the data better than SA model w/o T and Huff model. However, we cannot conclude that SA model w/o N is significantly better than SA model or vice versa based on ∆AIC i and w i values. The reason why we get similar performance for SA model and SA model w/o N may be due to the geographic distribution of attractions in Yosemite National Park (see Fig. 4). Most attractions are clustered (i.e., they have similar neighbors) at the center of the park, thus the neighboring effect may not be as significant as those of Acadia National Park. More details about this will be discussed in section 5.2. In general, the SA model w/o N performs better than the SA model w/o T, since the temporal factor provides more fine-grained data and for national parks, to include temporal variation when estimating visiting  patterns is crucial. Overall, the experimental results indicate the necessity to incorporate both social and temporal effects into the Huff model.

Regional Variability of Parameters
After examining the globally fitted parameters for the entire park, we further explore the regional variability of the parameters. The intuition underlying this experiment is that the relative impacts of attractiveness (α), distance (β), and neighboring effect (θ) can be different across regions in the park. Therefore, calibration is conducted for each attraction in the two national parks using observed visiting probabilities calculated from the trip sequence data. Each attraction is treated as an origin to estimate how visitors choose their next destination starting from this origin attraction. The OLS calibration gives one set of parameters (α, β, and θ) per origin, reflecting how attractiveness, distance and neighboring effect, respectively, contribute to the visiting probabilities. The results for attractions within Acadia and Yosemite National Parks are shown in Tab. 4 and Tab. 5. Only these significant origin attractions with more than 30 observed trips are included in the table.
In general, a large absolute estimation for α, β, or θ indicates a significant influence of attractiveness, distance, or neighboring effect to the destination choices, respectively. Tab. 4, demonstrates the parameter esti-mations for attractions in Acadia National Park. Here, we see a relatively small estimation of α for Cadillac Mountain, which is the top 1 traveler favorite attraction in Acadia National Park ranked by TripAdvisor. This means that compared with visitors at Boulder Beach, Sand Beach, and Jordan Pond etc., after visiting Cadillac Mountain, they are less interested in the attractiveness of an attraction to visit the next destination. The absolute values of β estimations are greater for Bass Harbor and Bubble Rock, which indicate that visitors tend to choose closer next destinations starting from these two attractions. A larger θ estimation indicates a higher probability of visitors choosing destinations with closer and more attractive neighbors, most likely clustered attractions, such as the Boulder Beach, Thunder Hole and Sand Beach cluster (see Fig. 4(a)).
Tab. 5 includes the parameter calibration results for attractions in Yosemite National Park. We see relatively larger estimations of α for attractions clustered at the center of the park, El Capitan Meadow, Lower Yosemite Fall, Sentinel Bridge, etc., which are shown in the red box of Fig. 4(b). The largest absolute value of β estimation for Tuolumne Grove indicates that visitors at this attraction are very likely to choose a closer next destination. Since Yosemite National Park is roughly 15 times larger than Acadia National Park in area, absolute values of β estimations are generally smaller compared with those of attractions in Acadia National Park. This indicates visitors are less sensitive to distance and are willing to travel further in  Yosemite National Park. The estimation of parameter θ is greater for dispersed attractions like Olmsted Point and Tuolumne Grove, as can be seen in Fig. 4(b). This means visitors at these two attractions are more attracted to clustered attractions (i.e. attractions with closer and more attractive neighbors), most likely the Tunnel View and Glacier Point clusters as shown in the map. In Tab. 5, we also see significant negative θ values, which reveal that visitors tend to travel to less clustered attractions (i.e., attractions with further and less attractive neighbors), especially for visitors at Four Mile Trailhead, Bridalveil Falls, Sentinel Bridge, etc., that are already in a clustered region. For origin attractions with θ closer to 0, it means that neighboring effect is not an important factor for visitors to choose their next destination at these places.

Temporal Variability of Parameters
To further explore the temporal variability of the model parameters, we divide the trips in Yosemite National attractions in non-summer months. Meanwhile, we see also larger absolute θ estimations for Curry Village in non-summer months and Bridalveil Falls in summer months. A positive θ estimation means visitors are more attracted to clustered attractions (i.e., attractions with closer and more attractive neighbors) and negative value means the opposite.

Conclusions and Future Work
In this work, we explore the visiting probabilities of attractions within two national parks using a socially aware Huff model, in which both social factors and neighboring effects are taken into account. To calibrate the model parameters, we use the observed trip sequences extracted from the geotagged Flickr photos within Acadia and Yosemite National Parks from the year 2010 to 2019. For the social factor, we have shown that incorporating the number of photo views when evaluating the attractiveness of a place achieves a better result than simply using the number of photos or number of users alone. The calibration results also demonstrate that the socially aware Huff model considering a temporal factor in place attractiveness and neighboring effects is more accurate than the original Huff model in predicting visiting probabilities of attractions within both national parks.
We further explore the visiting patterns of each attraction within the two national parks based on model parameters of attractiveness, distance, and neighboring effect factors. In general, visitors in Acadia National Park are more sensitive to the distance factor and neighboring effects when choosing their next destination, while visitors in Yosemite National Park are more sensitive to the attractiveness factor. We have also shown that there is a regional and temporal variability of the model parameters.
This work is based on our three assumptions that: (1) People tend to choose more attractive travel destinations; (2) People tend to choose closer travel destinations; (3) People tend to choose travel destinations with more beneficial future choices. In fact, there could be many other factors contributing to the visiting probability of a place. Taking the social factor alone as an example, traveling under social influence or taking copycat photos (Picheta, 2021) has become an emerging trend. In future work, we plan to consider more comprehensive measurements to evaluate the social impact of geotagged photos in order to better capture the travel patterns. Furthermore, this work is only studied using existing attractions within national parks. We plan to look at more than just two parks in the future, explore new methods to discover emerging travel destinations and study their interactions with the surroundings.