Recommender systems could benefit from not only recommending the most fiting items, but also in what quantity the user should consume them. For example, a personalized travel recommender system could indicate not just which city one should travel to, but also how much time to spend there. We present a data-driven solution to this problem based on mining trips from location-based social networks. To determine the recommended duration of stay at a destination, we consider how long travelers typically stay at diferent cities and how much time the current user generally spends visiting cities.
favorite artists appear repeatedly in a long music playlist. In this case, the assumption is that an item should be watched, listened to or experienced as a whole, not only parts of it.
In some scenarios, it is important to decide not just which items should be recommend, but how
much of an item should be consumed. For example, a destination recommender should not only
recommend where to go, but also the optimal duration of one’s stay at each location. The duration
can vary depending on the relevance of the item and other domain-related factors, such as the type of
traveler [
In this paper, we examine the problem of determining the personalized amount of recommended item consumption in recommender systems, since previous approaches have not solved this problem convincingly. We then present a way to derive the duration of stays in the domain of destination recommendation. Finally, we discuss the generalizability of our method and draw conclusions.
There is very limited related work with regard to determining the amount of item consumption in recommender systems for travel and tourism.
Melià-Seguí et al. have investigated the typical duration of stays for tourists visiting diferent
point-of-interest categories using a Foursquare data set; for example, they considered the average
amount of time that users spent in restaurants [
There are several approaches to recommending travel packages, such as the Tourist-Area-Season
Topic (TAST) Model [
When recommending a sequence of travel-related items, such as an itinerary for a city visit, the
problem of how much time to spend at individual locations arises as well. For example, De
Choudhury et al. have analyzed Flickr photo streams to reconstruct paths of tourists in a city [
DERIVING THE DURATION OF STAY IN A DESTINATION RECOMMENDER SYSTEM Having analyzed the related work, we will now sketch our ideas as to how to resolve the problem in the domain of destination recommendation. Our proposed solution addresses the question of making personalized recommendations regarding the duration of a tourist’s visit to a city by considering two factors: the typical time that all tourists spend in that city and the particular user’s average length of stay at a given destination. Initially, we need to know the distribution of the durations of people’s stay at a destination, since there can be substantial diferences between destinations as to how long one needs to explore it. For example, a smaller city can be covered within a day or two, whereas a major metropolis might require more time. The second aspect is the pace at which the particular traveler visits destinations. Some tourists want to immerse themselves deeply into a culture and therefore stay at each location for a longer time, whereas others want to visit as many diferent places as possible during their holidays. To quantify these behaviors, we need a database of previous trips to establish a distribution of how long people stay at a specific destination, such as a country or a city.
We employ our previously proposed approach to mine trips from a data set stemming from
Foursquare [
The next step is to determine the pace at which our particular user typically travels, i.e., the distribution of the duration the individual traveler’s past blocks. To obtain this information, we can either ask the traveler to provide some information about past trips directly, or we can request access to the individual’s mobility paterns from her profile on a LBSN. Once we have this information about past trips, we can derive the user’s pace by comparing it to the quantiles of all other travelers who have visited the same destinations. This essentially establishes a collaborative filtering method to derive the duration of stays from actual user behavior. Example. To visualize our approach, we show how the algorithm would calculate the personalized duration of a sample user’s visit to Washington, D.C. To that end, we calculate the quantiles of the previously visited cities. In our example, the user made three previous visits, spending 16 days in Tokyo, 10 days in Jakarta, and 7 days in London. We have visualized the distributions of the durations of blocks in the three cities in Figures 2, 3, and 4. The durations of these trips reveal that our user is a relatively slow-paced traveler compared to others, as her lengths of stays are toward the right side of the distributions. The trip to Tokyo was at 97%, the stay in Jakarta at 81%, and the visit to London at 96% of the cumulative distribution function. To aggregate the user’s pace over the previous trips, we can calculate the mean percentile, which is 91%. We can then find that percentile in the distribution of visits to Washington, D.C., where the 91th percentile of the distribution is at 10 days (see Figure 5). Therefore, this would be the recommended duration of stay.
In this paper, we have identified and examined the problem of determining the amount of item consumption in recommender systems. To solve this problem, additional information about the domain and the user’s preferences is required. We showcased an approach to determining the personalized duration of a stay in a city, based on the analysis of mobility data from location-based social networks. The underlying method is, however, generalizable to similar problems, given the availability of appropriate data. We argue that such data are indeed often available, especially in commercial recommender systems. In the tourism sector, airlines and hotel portals have a long history of user data and, which they could easily leverage when making recommendations. After all, the proposed approach can be used in any recommender systems domain, where the amount of the recommendation maters and where information about the distribution of the quantity is available for both all users and the particular user of interest.
In the future, we plan to extend our analysis using more trips from diferent LBSNs and to assess our approach by using ofline evaluations that involve cross-validation as well as user studies.