A social circle in a user’s ego network is a group of interconnected people that have common attributes between themselves and the user. In order to automatically detect such circles, an unsupervised/semi-supervised learning approach is designed in this paper to optimize the circle and user profile similarity parameters by using the network structure and user profile information. Researchers in information retrieval, social media, ego networks, and machine learning areas will want to study this work.

The proposed approach models circles to be latent graph variables that are inferred from the user’s ego network. Within each graph and driven by edges that are “likely to form within circles and unlikely to form outside of them,” the model computes the edge forming probability in a circle using the profile similarity of the edge’s two user nodes, “rewarding edges that appear within circles ... and penalizing edges that appear outside of circles” via a tradeoff constant. Then the proposed machine learning algorithms calculate the profile similarity parameter. If the learning algorithm only relies on the user node attributes and edge information, the algorithm is unsupervised. If the learning algorithm starts with user-labeled members of circles, the algorithm becomes semi-supervised.

The proposed algorithms have been evaluated on a social network “dataset of 1,143 ego networks and 5,541 ground-truth circles obtained from Facebook, Google+, and Twitter.” Compared to baseline methods considering network structure, profile information, or both, the proposed model has achieved decent performance and scalability in finding disjoint, overlapping, and hierarchically nested circles.