Impact of Global Network on Localized Link Prediction

The problem of link prediction and its variants have driven significant research efforts over the past decade. Real world applications, ranging from social networks (Facebook friend recommendation) to national security (detecting links in terrorist networks), have motivated interest in this problem. Link prediction can be reframed into other varieties, such as missing link prediction or link recommendation. We will explore the latter framing. After considerable literature review, I found a lack of research into the impact of network structure when link prediction is constrained to a subset of the total network. One possible context of this situation is using performing link recommendation for a group of individuals at a networking event. Newly edges are formed exclusively between attendees, so what is the impact of utilizing external network information including nodes currently not present at the event. Does this information improve accuracy or does this external network play only a small factor in a local instance of link prediction? With this motivation, we set out to modify existing link prediction algorithms for this special case and apply them to an example network. We utilized unsupervised and supervised random walk, as well as some similarity metrics and linear classifiers in our aim of measuring this impact.