Link prediction in Foursquare network

TL;DR

This paper evaluates link prediction methods in the Foursquare bipartite network, finding that network-based inference outperforms traditional collaborative filtering and ranking methods, especially when incorporating metadata.

Contribution

It compares multiple link prediction techniques on Foursquare data and demonstrates the effectiveness of network-based inference, highlighting the benefit of using metadata.

Findings

Network-based inference performs best among tested methods.

Incorporating metadata improves prediction accuracy.

Traditional collaborative filtering is less effective than network-based inference.

Abstract

Foursquare is an online social network and can be represented with a bipartite network of users and venues. A user-venue pair is connected if a user has checked-in at that venue. In the case of Foursquare, network analysis techniques can be used to enhance the user experience. One such technique is link prediction, which can be used to build a personalized recommendation system of venues. Recommendation systems in bipartite networks are very often designed using the global ranking method and collaborative filtering. A less known method- network based inference is also a feasible choice for link prediction in bipartite networks and sometimes performs better than the previous two. In this paper we test these techniques on the Foursquare network. The best technique proves to be the network based inference. We also show that taking into account the available metadata can be beneficial.