Using Friendship Ties and Family Circles for Link Prediction

Abstract

Social networks can capture a variety of relationships among the participants. Both friendship and family ties are commonly studied, but most existing work studies them in isolation. Here, we investigate how these networks can be overlaid, and propose a feature taxonomy for link prediction. We show that when there are tightly-knit family circles in a social network, we can improve the accuracy of link prediction models. This is done by making use of the family circle features based on the likely structural equivalence of family members. We investigated the predictive power of overlaying friendship and family ties on three real-world social networks. Our experiments demonstrate significantly higher prediction accuracy (between 15% and 30% more accurate) compared to using more traditional features such as descriptive node attributes and structural features. The experiments also show that a combination of all three types of attributes results in the best precision-recall trade-off.