LBRW: A Learning based Random Walk for Recommender Systems

1. Introduction

Recently, with the advances in GPS-enabled devices, Wireless Networks and ubiquitous computing, online social networks or Location-Based Social Networks (LBSNs) such Foursquare, Gowalla and FaceBook Places have witnessed a rapid expansion largely among users and have attracted important researchers' efforts, to investigate spatial, temporal and social aspects of user patterns. LBSNs allow users to “check-in” at geographical locations¹ and share this information with friends.

This information offers a potential knowledge about users' preferences on geographical locations. Therefore, it can help advertisers to build a personalized and efficient recommender system, in order to guide users exploring new locations. Typically, recommender systems are inevitable tools to filter and present relevant information to the users, helping them in their decision-making process. For instance, we have a set of users and a set of locations. Each user can rate the visited locations by a multiple-scale rating. The recommender system has to provide rating prediction for unvisited venues, and recommend venues that are already rated.

The most challenging aspect of this kind of recommender systems is the correlation between the facets of check-ins activities: user interests and “mobility homophily”, understanding to what extent these facets can be exploited to identify and recommend personalized new locations to be discovered by the user.

To tackle the aforementioned challenging aspect, our work contributes by an approach for both location prediction and location recommendation. We develop a Learning Based Random Walker (LBRW) that obviously constructs a network which its structure and characteristics are inferred from LBSN data, particularly user interests and “mobility homophily”; with the aim to improve the location recommendation quality.

In Section 4, we analyze the Foursquare² data and we discover that it exhibits the following interesting characteristics.

• •

  New Locations or Previously Unvisited Locations: We discovered that 75.28% of users have check-ins in previously unvisited locations, while the average percentage of check-ins in unvisited locations is 71.41% (see Figure 5 and Figure 6, Section 4.1.).

• •

  Friends and Similarity: In our data, 90% of users have 3.16% and 1.04% of common check-ins with friends and with all users respectively. This indicates that similarity in check-ins with friends exists; still this similarity is limited (see Figure 3 and Figure 4, Section 4.1.).

• •

  Matrix Sparsity: From analyzing our data set, we have concluded that the user-location matrix density is about 7.81*10^-4, and the ratings matrix density is about 3.81*10^-2.

Figure 5.

Power-law distribution of users count versus check-ins in new venues count on a log-log scale

Figure 6.

Visited venues versus number of users with new venues on a log-log scale

Figure 3.

Number of users with social link versus check-ins similarity percentage

Figure 4.

Users count versus check-ins similarity percentage