Site Loader
Rock Street, San Francisco

AbstractValuable and useful recommendations are hard to receive yet in a tourist destination. The paper describes the elaboration of a Recommender System (RS) capable of providing high-quality personalized recommendations for the stay in a tourist destination. The proposed RS will give recommendations such as places to dine, to relax and possibilities for sports activities. To encompass the daily routine of a person on vacations in a tourist destination, a hybrid RS is proposed that will help unify the best aspects of different recommendation algorithms while simultaneously minimising the drawbacks of the individual algorithms. As input for the algorithm, the information needs of tourists were examined collecting data in an Alpine tourist destination. The outline of the proposed recommender system and the results of the data collections are presented.IntroductionIn our daily lives, we all have to make decisions. We are faced with questions such as which book should I read? What kind of music should I listen to now? Which movie should I watch? Where could I travel next? To which restaurant could I go for dinner? Which bar could be nice for a drink? To make informed decisions, we use our knowledge or previous experiences, ask a friend, search the internet, ask an expert and so on. However, good advice is difficult to find, expensive, or unreliable. Recommender systems fill up this vacuum by providing personalized, high quality, and affordable advice. Another influence is the use of smartphones. Within a tourist destination, smartphones and apps have become important tools for tourists navigating in an unknown place. Smartphones are used for searching for information what to do, for facilitating the stay by checking train schedules or the weather, to communicate with home and on-site, and for entertainment purposes. Also because of the smartphone driven shorter-term nature of organizing a stay’s activities, online recommendations in tourism are in demand.Recommender systems in diverse domains have been applied over the last 25 years; however, they have only recently started to be developed in the tourism domain. Tourism is a multi-trillion dollar industry (total contribution of tourism to global economy in 2016 was 7.61tr USD (, where the task of providing information is traditionally done by tour operators and travel agents. Further, people are increasingly using internet based platforms(,, and to collect information and plan their trips, but also find it very time consuming. There is no provision of easy access to relevant information for users. Thus, RS fill the gap by providing valuable, personalized and accurate recommendations to them. However, for a short-term planning of trips on-site, specialized recommender systems for mobile devices are needed, which can provide more valuable, personalized and accurate recommendations. This will result in saving time and effort at users’ end. The goal of the project described in this paper is to provide tourists of an Alpine destination in Switzerland with personalized recommendations that suit their needs and taste. Later on, the RS will be applied in further destinations. Recommender systems in the tourism domain have been classified into four categories by Borrás et. al. according to the type of recommendations they provide. These categories are i) Destination/Tourist packs, ii) Suggest attractions, iii) Trip planner, iv) Social aspect. Destination/Tourist packs systems recommend destinations as a whole considering the user’s likings and dislikings. Suggest attractions systems recommend popular spots, stopovers, eateries, and events once a user is inside the target destination. The proposed recommender system falls under the category of Suggest attractions. Trip planner systems recommend places considering the user preferences and needs and help users organize them into a route. Social aspects systems connect tourists in a particular tourist destination through popular social networks and bind together tourists socially with the functionality to share pictures, comments, and reviews with the other tourists in the same destination at that time.The task of providing recommendations to tourists is currently done by tour operators and travel agents who are unable to provide high-quality automated online personalized recommendations. Generally they are only able to provide recommendations which roughly fall in such categories as “best sellers”, “most favorite”, and “most liked”. At the most, they might ask a few questions about a person’s inclinations, for example whether a person would prefer a historical city to a nature reserve or whether a person is adventurous or cautious and recommend a destination accordingly. However, they are unable to provide high-quality personalized online recommendations on hotels, restaurants, bars and sports activities available at the recommended destination. The current practice of tourism service providers allows them to make generic recommendations based on the information available about the most popular restaurants or most visited bars. These recommendations are rather like one-size-fits-all, which do not completely fulfill tourist’s real needs. If, for example, the recommendation is a very expensive restaurant with excellent food and environment, the recommendation might not go well with someone who has a limited budget such as students. A low cost restaurant will suit them better as they might not want to spend money on expensive food.The dynamics of a RS in the tourism domain are generally different form the RSs in other domains such as songs, movies, books or news. A person might need recommendations in quick succession for books, movies, and songs, perhaps every week or every day. News recommender systems have to work with an even shorter frequency. Six-hour-old news will not be interesting to a reader who is interested in knowing the current news. On the other hand, a user of a tourism RS might need recommendations only when he is on holidays which could be once a year or once in six months. When he goes back to the normal routine life after holidays, he will be inactive till his next holidays. This allows RS in the tourism domain to utilize User-User Collaborative Filtering (UUCF) more effectively than in other domains as the rating matrix will not change as quickly as in other domains. The problem with UUCF is that they do not scale well. Every time a recommendation is needed, a subset of the intermediate matrices associated with UUCF needs to be computed rendering the approach computationally expensive. In the proposed system, the intermediate matrices will be updated once a day when the system load is low as intermediate matrices do not become obsolete as quickly as in other fields of application. Although, we will still be sacrificing slight accuracy over huge computational efficiency, the effects of inaccuracy are limited to only collaborative filtering based approaches. As the proposed RS is a hybrid recommender system (see Fig. 1), the effects of slight inaccuracy can be compensated by other algorithms.The aim of the proposed RS is to recommend restaurants, bars, hotels and sports activities in an Alpine tourist destination. The recommendations will answer such questions as: Where should I have my lunch or dinner? To which bar should I go to tonight? Which sports activity should I go to tomorrow morning? Which hotel should I book? A user who is on holidays in the target destination is considered an active user needing recommendations many times a day during his holidays.ApproachFigure 1 shows the architecture diagram of the proposed RS. The source of data is a mobile app specifically built to provide personalized recommendations. The app provides the interface through which the users of the recommender system provide explicit feedback in the form of ratings and likings. The app also collects implicit user feedback by monitoring users’ clicks and the viewing duration. As explicit user feedback data is rare compared to implicit feedback data, we intend to use the implicit feedback data more in our proposed recommender system. We use a crude representation of the data, that is, a viewing duration of less than 2 seconds is assumed to be a dislike while more than 2 seconds is assumed to be a like. Thus, if a recommended item was clicked, the viewing duration will decide about the liking or disliking of the item by the user.The data obtained from the app is then stored in the knowledge base which is shared with the recommender system. The recommender system will utilize the data in the knowledge base to provide personalized recommendations. The computed recommendations are then written back to the knowledge base. The appManager queries the knowledge base and asks for recommendations for a particular user. It then sends the computed personalized recommendations to the user. The user sees the recommended items and decides to either click it or not to click it. If he clicks and views the data, then the viewing duration is recorded. If the user decides to consume the item and physically visits the item recommended to him, then he can choose to either rate it, like/dislike the item or provide no feedback at all. This completes the RS cycle.The recommender system is intended to be a hybrid recommender system where the intention is to utilize the strong aspects of different techniques and to minimize the individual drawbacks of the employed techniques. For example, the UUCF is better at providing recommendations which can be both diverse and serendipitous, but they suffer from the cold start problem. As this is a work in progress, the classical UUCF and Item-Item Collaborative Filtering (IICF) algorithm have been implemented.Together with collaborative filtering, content based filtering is also utilized to capture other aspects of user preferences. Content based filtering will provide better recommendations when the collaborative filtering matrices are obsolete. To use the content based filtering, the identification of the correct tags is of utmost importance. The tags should capture all the aspects of the items, and simultaneously they should be expressive enough to capture user preferences. The item tags and categories were identified through a study conducted which is explained in the next section. The recommendations from the different recommendation algorithms will be combined together using weighted linear combination.Category definitionData were collected to define and structure the information for the content-based recommender system.Data collection 1The aim of the first data collection was to reveal the typical information needs of the tourists in an Alpine tourist destination in Switzerland. Further, the important drivers of these information needs should be determined. These drivers were intended to serve as tags which will be used by the algorithm. Drivers such as the context, the duration of the stay, travel companions, socio-demographic factors, needs, attitudes, and expectations were analyzed.Participants were recruited on site, with the help of the project partners; hotels and apartments where the tourists were living during their stay. Seven tourists actively participated, providing in-depth data about the information they needed during their stay. As an incentive, participants received a free one-day ski pass.Method: Mobile ethnography was applied to collect the data. The method applied is based on self-report data. It allows participants to capture the information they consider to be relevant regarding the research question and thus facilitates to receive an extensive but relevant feedback.Materials and measures and procedure. In an initial personal briefing, participants filled in a form, giving information about their country of origin, sex, age, relation to their fellow travelers (e.g., family), if they are first-time visitors or regulars in the Alpine tourist destination, and why they have chosen the destination. Further, they had to indicate their most urgent questions and information needs before and at the beginning of their stay. In the initial briefing, participants were also informed about the aim of the data collection, to capture their experiences with moments of information needs. Further, they were given instructions about downloading and using a tool for capturing the self-report data. The smartphone app “ExperienceFellow” was applied. ExperienceFellow works like a diary on the smartphone. Participants were also informed about the voluntary individual interviews which were planned to be conducted after the self-report phase of the data collection, to precise the self-captured information. Participants captured data during 3-4 days. They were instructed to evaluate each captured moment on a very bad – bad – neutral – good – very good scale. Photos, videos, and texts could be added to captured moments.Data collection 2 To complement data collection 1, the aim of data collection 2 was likewise to identify and contextualize the moments of information needs of tourists.Participants and method. Nine tourists from various countries with different sociodemographic characteristics participated in in-depth qualitative interviews in the Alpine tourist destination. As an incentive for their participation, a beverage was offered to them.Materials and measures. The interviewees answered questions regarding their information needs before and during their journey to the destination, at the beginning of their stay, and during their stay as well as special moments/experiences. To learn about details about the interviewees, they were asked to answer questions about their country of origin, mother tongue, age, fellow travelers, frequency of visits to the destination, and type of accommodation.Definition of information need drivers: Taking into account the existing literature and data of the local DMOs about their guest structure, the qualitative data of the sixteen participants of the data collections 1 and 2 were analyzed. The drivers of the information needs of the tourists, i.e., the tags for the recommender system’s algorithm, were defined by tourism and information technology experts from academia and practice, by condensing and relating the data available. To structure the data, the common concept of the customer journey was applied, the guest’s experience from getting inspired and searching for information, to booking and traveling, up to the experience after traveling.Results: For instance, age was identified as not an important driver of tourists’ information needs. In contrast, the fellow travelers were an important determinant. For example, families have different information needs in an Alpine tourist destination than single travelers. Another important driver identified was, whether tourists were first-time visitors or regulars.In addition to the drivers defined, useful recommendations per drivers and subgroups of drivers were specified. Thereby, the phase of the tourists’ stay (e.g., at the beginning of the stay or during the stay) was considered. One further point considered, when providing inputs for the algorithm, was the flexibility in recommendations. For example, if the recommender system recognises that a tourist loves eating meat, it should not recommend restaurants offering meat every day of a tourist’s stay but recommend the tourist to go for having traditional Swiss fondue, one day.ConclusionA recommender system in the tourism domain is proposed. The data in the tourism domain has its own peculiarities which makes it different from the data in the other domains. Utilizing these peculiarities, the scaling problem of user-user collaborative filtering can be minimized. Both explicit and implicit user feedback data is used in a hybrid recommendation system. A hybrid recommendation system is used because it combines the strong aspects of the different recommendation approaches while simultaneously minimizing their drawbacks. Data were collected in an Alpine tourist destination to capture expressive tags for the algorithm, i.e., the drivers of distinct information needs and based on these needs recommendation opportunities.

Post Author: admin


I'm Dora!

Would you like to get a custom essay? How about receiving a customized one?

Check it out