Latent Interest-Group Discovery and Management by Peer-to-Peer Online Social Networks

We address management of latent, emerging interest groups (IGs), spanning both unsupervised, distributed IG discovery and any cast-query forwarding, in dynamic peer-to-peer(P2P) on-line social networks. The P2P network has at least one layer of super-peers (Diaspora pods) that support a group of ordinary peers/clients. There are a number of challenges here, including: i) semantic processing at scale to disambiguate word meanings in queries, ii) unsupervised estimation of the number of active IGs, iii) detection of IG churn and emergent IGs, iv)design of optimal query forwarding to maximize query resolution and minimize the required number of hops, while achieving practical local cache searching and network communications. In this preliminary study, we assume a common, fixed keyword lexicon for query formation and latent IG characterization. We propose unsupervised, dynamic, on-line clustering that mines the super-peers' query caches in a distributed fashion. Customized Bayesian Information Criterion based model-order selection is employed, independently by each super-peer, to estimate the set of active IGs and to help achieve efficient query forwarding. The proposed method is numerically evaluated against both exhaustive cache search and a random walk strategy.