1. Introduction
Recently, there has been a change in the tourism trend towards the Tourism 2.0 paradigm due to the increase in travel experience worldwide and the spread of information acquisition and sharing through the Internet. The Tourism 2.0 paradigm seeks to create new value for the tourism industry by valuing the vitalization of information and communication and participatory cultural experiences, rather than simple travel and consumption activities [
1]. In addition, Tourism 2.0 aims to give new value to the new tourism industry by developing information and technology, providing various experiences and cultures to tourists, and promoting the development of the local economy and environment by pursuing sustainable tourism. The free/foreign independent tour (FIT) is one of the representative tours of the Tourism 2.0 paradigm, and the demand for non-face-to-face services is increasing due to COVID-19 [
2]. Due to FIT, tourists search, reserve and pay, and share information through online travel platforms and mobile applications, and the influence of social media and online platforms on the overall tourism industry is increasing [
3]. FIT is required to develop intelligent tourism service tools with positive effects such as an improved understanding of tourist destinations and revisit rates, the development of tourism information technology, and improved marketing efficiency [
4]. Travel planning and guidebook applications and tourism recommendation services, which are tourism service tools for FIT in the Tourism 2.0 paradigm, have advantages such as improved customer experiences, time and cost savings, data analysis, and marketing utilization [
5]. For tourism service tools, research is being conducted on a system that recommends tourist destinations according to the relevance of tourists or tourist destinations [
6]. Collaborative filtering (CF)-based recommendation, which recommends tourist destinations according to tourist relevance, is a method of finding similarities between users using information between users and items [
7]. Content filtering (CB)-based recommendation, which recommends tourist destinations according to their relevance, is a method for determining the similarity between items using item information [
8]. Handling new data without information is difficult for filtering-based tourist destination recommendation systems (RS) because of the required data size and cold start problem. Therefore, research on an RS using artificial intelligence (AI) is being conducted [
9]. An AI-based RS is recognized as a key element capable of targeting marketing and personalized recommendations by providing appropriate information or products that users may prefer by subdividing them by user inclination and type [
10].
Tourism service tools include a CF-based RS method that recommends tourist attractions according to their relevance to tourists and a CB-based RS method that recommends tourist attractions according to their relevance to tourist destinations [
11]. The CF-based RS assumes that similar tourists have similar preferences for a specific tourist destination based on interaction data between tourists and tourist destinations [
12]. Because CF is performed based on interactions between tourists and tourist destinations, recommendations can be made even if the similarity between tourist destinations is not high. However, new tourist destinations cannot be applied because no tourist information is available [
13]. To solve the cold start problem, a CB-based RS that recommends tourist destinations according to their relevance can also recommend similar tourist destinations using tourist destination information [
14]. CB can provide recommendations without interaction data between tourists and tourist destinations. However, its performance deteriorates compared with CF-based RS when sufficient data are collected [
15]. An RS using AI can dynamically utilize the data size or characteristics because they use tourism patterns rather than the similarity between items, and they have achieved remarkable recognition performance with the development of algorithms and computation power [
16]. However, RS based on tourism patterns using AI have limitations in dynamic situations because they do not reflect real-time changes in external factors and distance information, such as temperature or precipitation [
17].
In this study, we proposed a real-time recommendation system for tourism (R2Tour) to respond to dynamic situations by reflecting real-time changing external factors and distance information and recommend customized tours according to the type of tourist. R2Tour recommends the top five nearby tourist destinations by learning a real-time context that includes real-time situational information and a tourist profile that includes the type of tourist using a machine learning (ML) model. The real-time context includes weather information, such as temperature and precipitation, and sequential information, such as season and location. The tourist profile includes gender, age, companion, and tour type. In this study, electric vehicle information (EVGPS) provided by the Korea Electric Power Corporation Knowledge Data Network (KEPCO KDN), the Visit Korea data lab, and Korea Meteorological Administration data were used to evaluate R2Tour. The results of evaluating R2Tour using information on tourist attractions, situations, and tourist profiles on Jeju Island, created by combining EVGPS, Visit Korea Data Lab, and Korea Meteorological Administration data, were verified with accuracy of 77.3%, micro-F1 0.773, and macro-F1 0.415. Our updated code, dataset, and trained models are available at
https://github.com/junhoy00n/Real-time-Recommendation-for-Tourism (accessed on 3 March 2023).
The objectives and contributions of this thesis are summarized as follows.
Development of a customized travel destination RS according to real-time context changes: It is common for existing tourism service tools to recommend destinations based on the relationship between tourists and tourist attractions or tourism patterns. However, recommendations are complex when there is insufficient information or situations that change in real time. In this paper, we try to overcome these limitations by proposing a real-time RS that recommends customized travel destinations according to external factors, distance information, and types of tourists.
Travel pattern analysis and prediction using ML models: In this study, an ML model is trained using traveler profiles and contextual information to analyze and predict travel patterns. Through this, we can understand travelers’ preferences and behavior patterns and contribute to implementing a personalized RS.
Data collection and analysis for smart advertising system development: In this study, the important process is to collect and analyze the data necessary to implement the RS. Through this, it is possible to understand the traveler’s preferences and behavioral patterns and derive useful information that can be utilized in the tourism industry, such as smart advertising systems.
The remainder of this paper is organized as follows. The related work is mentioned in
Section 2 and consists of a filtering-based RS, ML model, and AI-based RS. The architecture of R2Tour is presented in
Section 3. The experiment is mentioned in
Section 4; the description of the data and environment used in the experiment and the experimental results are discussed; finally, the conclusions and future work are addressed in
Section 5.
3. Real-Time Recommendation System for Tourism
Handling new data without information is challenging for filtering-based tourist destination RS because of the required data size and the cold start problem. Therefore, research on RS using AI is being conducted [
37]. RS using AI can dynamically utilize the size or characteristics of data because they use tourism patterns, ignoring the similarity between items, and achieve remarkable recognition performance with the development of algorithms and computation power [
38]. However, RS based on tourism patterns using AI have limitations in dynamic situations because they do not reflect real-time changes in external factors or distance information, such as temperature or precipitation. In this paper, we propose a tourist destination RS using R2Tour to respond to dynamic situations by reflecting real-time changing external factors and distance information and recommending customized tours according to tourist types. As shown in
Figure 1, R2Tour recommends the top five nearby tourist destinations by learning a real-time context, including real-time situation information, and tourist profiles, including tourist types, through an ML model. The following section describes the configuration of the tourist attraction corresponding to the dependent variable R2Tour, and the real-time context and tourist profile corresponding to the independent variable.
3.1. Tourist Attraction Used for R2Tour
R2Tour’s independent variable, tourist attraction, consists of information on the central tourist attraction. The top five nearby tourist attractions are obtained according to the current location, which uses the EVGPS and the Korea data lab. The central and related tourist destination information of the Visit Korea data lab consists of information on the main monthly tourist destinations and nearby tourist attractions. The EVGPS measured the movement routes in 1-min units using 52 electric vehicles in the Jeju Smart City pilot city project for three years at KEPCO KDN. The numbers of driving records for each vehicle are listed in
Table 2. The central-related tourist attraction information of the Visit Korea data lab used for tourist attractions uses information on the top 10 central tourist attractions and the top five tourist attractions by month, corresponding to Jeju. As shown in
Figure 2, the EVGPS sets the error range (50 m) based on the latitude/longitude lines because it is classified as driving owing to positioning errors, even if no actual movement exists. The EVGPS was used in the visiting state if no movement for more than 10 min occurred within the error range. If no movement occurred for more than 200 min, it was excluded from the experiment as a long-term parking state. In addition, the EVGPS classified as visited was used as the visit to the nearest tourist destination within the error range based on the Visit Korea data lab.
3.2. Real-Time Context Used for R2Tour
Circumstantial information, such as temperature, precipitation, or season, is a significant factor while visiting tourist destinations, and it changes in real time according to the location [
39]. However, existing AI-based RS do not reflect dynamic situations that change in real time because they use tourism patterns. R2Tour learns tourism patterns by reflecting the real-time context and tourism patterns to respond to dynamic situations, addressing the limitations of existing RS. The real-time context, a dependent variable of R2Tour, was extracted from the Korea Meteorological Administration data based on tourist attractions, generated by combining the Visit Korea data lab and EVGPS. The real-time context includes seasonal information on the visit date or measurement date, as well as temperature and precipitation information using Meteorological Agency data.
3.3. Tourist Profile Used for R2Tour
R2Tour responds to dynamic situations and uses tourist profiles as a dependent variable to segment customized tour recommendations according to tourist types. Because a tourist profile comprises information on age, sex, companion, and tour type, recommending segmented tourist destinations is possible. Tourist profiles were analyzed based on mobile communication data and keywords on social media. It uses current information for each tourist destination in the Visit Korea data lab. The value corresponding to the highest value based on the data analysis was used as the tourist profile of the destination. The tourist attraction extracted from EVGPS uses similar profile information (error range: −1 to +1) based on the measurement date and time.
5. Conclusions
Recently, there has been a change in the tourism trend towards the Tourism 2.0 paradigm due to the increase in travel experience worldwide and the spread of information acquisition and sharing through the Internet. Tourism 2.0 aims to give new value to the new tourism industry by developing information and technology, providing various experiences and cultures to tourists, and promoting the development of the local economy and environment by pursuing sustainable tourism [
41]. FIT is one of the representative tours of the Tourism 2.0 paradigm; travel planning and guidebook applications and tourism recommendation services, which are tourism service tools for FIT in the Tourism 2.0 paradigm, have advantages such as improved customer experiences, time and cost savings, data analysis, and marketing utilization. For tourism service tools, the tourism RS is filtering-based research that recommends tourist destinations according to the relevance of tourists or tourist destinations [
42]. Because handling new data without information is challenging for filtering-based RS owing to the required data size and cold start problem, research on RS using AI is being conducted [
43]. AI-based RS use tourism patterns, ignoring the similarities between items. It can dynamically utilize the data size or features and achieve remarkable recognition performance by developing algorithms and computation power [
44]. However, RS based on tourism patterns using AI have limitations in dynamic situations because they do not reflect real-time changes in external factors or distance information such as temperature or precipitation.
In this study, we proposed a tourist destination RS using R2Tour to respond to dynamic situations by reflecting real-time changing external factors and distance information and recommending customized tours according to tourist types. R2Tour recommends the top five nearby tourist destinations by training an ML model with the real-time context, real-time situation information, and tourist profiles that include tourist types. R2Tour trains tourism patterns to recommend tourist destinations according to real-time context and tourist profile information, not the interaction between tourists and tourist destinations. Therefore, even if a tourist has no preference for a new tourist destination, it is possible to recommend it according to tourism patterns such as weather, distance, and tourism type. In addition, unlike existing AI-based RS, it learns tourist destinations according to weather and distance information, so it is possible to respond to changing situational information in real time. We evaluated R2Tour using the Jeju tourism dataset created by combining EVGPS, Visit Korea data lab, and Korea Meteorological Administration data, and it was verified with accuracy of 77.3%, micro-F1 of 0.773, and macro-F1 of 0.415 in LightGBM. R2Tour responded to dynamic situations and recommended customized tours according to the type of tourist. In addition, classifying tourist activity patterns using XGBoost according to situational information based on EVGPS achieved accuracy of 80.6%, micro-F1 of 0.806, and macro-F1 0.73. The results of the experiment confirmed that real-time situation information, such as weather, season, and time zone information, is a significant factor while visiting tourist destinations. In the future, R2Tour can be expanded to other tasks such as recommending tour schedules [
45], travel routes [
46], and nearby tourist destinations based on location information. Moreover, R2Tour can be installed in vehicles to recommend nearby tourist destinations or expanded to tasks for the tourism industry, such as a smart target advertising system. The applicability of R2Tour has advantages such as improved customer experiences, time and cost savings, data analysis, and marketing utilization.