Location-based Social Networks (LBSNs) can be considered as a special Online Social Network category. Actually, an LBSN has the same OSN’s properties, but qualifies location as the core object of its structure.
Recently, advances in broadband wireless networks and location sensing technologies led to the emergence of smart mobile phones, tablets etc. that allowed ubiquitous access to the Web. In this new era, users can benefit by getting ubiquitous access to location-based services from anywhere via mobile devices. Moreover, users can share location-related information with each other to leverage the collaborative social knowledge by using LBSNs.
LBSNs allow users to see where their friends are, to search location-tagged content within their social graph, and to meet others nearby. LBSNs consist of the new social structure made up of individuals connected by the interdependency derived from their locations in the physical world as well as their location-tagged media content, such as photos, video, and texts. LBSNs are a type of social networking in which geographic services and capabilities such as geo-coding and geo-tagging are used to enable additional social dynamics. It presents three layers, namely, the user, the location, and the content layers. It is obvious that someone can exploit information from each layer independently to leverage recommendations. For instance, we can compute the geographical distance (i.e. Euclidean distance) between each pair of places in the location layer. Moreover, we can calculate the similarity among users based on the social network that exists in the user layer. Regarding the content layer, we can compute similarity among the information objects (i.e. video, tags etc.) based on their metadata. Please also notice the ternary relation among entities (i.e. user, location, content), which goes through all layers.
Acquiring this abundant contextual information, LBSNs can improve the quality of services on: (1)generic (non-personalized) recommendations of social events, locations, activities and friends, (2) personalized recommendations of social events, locations, activities and friends, and (3) user and group mobility behavior modeling and community discovery.
Generic Recommendations compute the same recommendation list (location, activity,event etc.) for all users, regardless the personalized preferences of each individual user. The most simple recommender systems are those based on counting frequencies of occurrences or co-occurrences of some given dimension. For example, a simple recommender system could just count the number of check-ins per
place, rank them and recommend those places with the larger number of check-ins.
A location recommender, for any user who travels in a specific city (e.g. New York), can first count each location’s frequency of check-ins. Then, it can recommend the top-n locations by sorting these locations in decreasing order of their scores and selecting the n most popular. Notice that an interesting location can be defined as a cultural place, such as the Acropolis of Athens (i.e., popular tourist destinations), and commonly frequented public areas, such as shopping streets, restaurants, etc. As far as the activity recommendations is concerned, an activity recommender can provide a user with the most popular activities that may take place at a given location, e.g. dinning or shopping. A target user can provide to the system the activity she wants to do and the place she is (e.g. coffee in New York). Then, the system provides a map with coffee places, which are nearby the user’s location (i.e. EuroPan Cafe in location A) andwere visited many times (i.e. 17 times) from 16 people. All the aforementioned recommendations can guide a user in an unknown place of visit.
The personalized recommender systems rely on past “check-in” history of users. Then, they correlate them with other users that have similar preferences and suggest to them new locations, activities and events. In particular, a personalized recommender exploits the time that someone has visited a location and her explicit ratings or comments on that location and predicts her interest in unvisited places. As there are three approaches that have emerged in the context of recommender systems: collaborative filtering (CF), content-based Filtering (CB) and hybrid methods. In the following, we briefly discuss, the special characteristics of each approach in the LBSN field.
CF methods recommend those locations, activities and events in a city to the target user, that have been rated highly by other users with similar preferences and tastes. In most CF approaches, only the locations and users’ ratings are accessible and no additional information, i.e. locations or users, is provided. User-based CF, employs users’ similarities for the formation of the neighborhood of nearest users. User-based CF is an effective approach in terms of accurate recommendations. However, it cannot scale-up easily due to the high computation of user similarity matrix. In contrast, location-based CF algorithm employs locations’ similarities for the formation of the neighborhood of nearest users, reducing the problem of scalability. In any case, a pitfall of both user-based and location-based CF is the cold start problem: new locations have received only few ratings, so they cannot be recommended; new users have performed only few visits, so there can be hardly found other users similar to them.
CB methods assume that each user operates independently. As a result, it exploits only information derived from location features. For example, a restaurant may have features such as cuisine and cost. If a user, in her profile, has set her preferable cuisine to be Chinese, then the Chinese restaurants will be presented to her. Apparently, the limitation of these systems lies upon the fact that other people’s preferences are not considered. In particular, it exploits a set of attributes that describes the location and recommend other locations similar to those that existin the user’s profile. This way, the cold start problems, faced by CF methods, for new locations and new users are alleviated. However, the pitfall of CB is that there is no diversity in the location and activity recommendations.
The combination of social with geographical data, is becoming a way of handling shortcomings when only one type of data is taken into consideration. For example, the social graph (i.e. trust/friend connections) is not dealing with location analysis, whereas collaborative filtering maintains a user profile mainly based on rating data. The idea of a hybrid approach suggests that by using both data (i.e. social and rating data) it is possible to overcome each other’s shortcomings and make the recommendation result to be more accurate. A hybrid system is where geographical data are combined with social data to provide location
and activity recommendations. GPS location data, user ratings and user activities to propose recommendations to interested users along with appropriate explanations.