This paper proposes a general interaction model called iLink for social networks to accomplish tasks through peer collaboration. iLink models social networks as graphs with nodes having property vectors and probabilistic, dynamic topology. It develops a learning framework to understand how social networks create artifacts and build new applications by modeling node/link properties, developing topic models from interactions, and learning these models over time from message passing. The paper describes applying this model in a system called FAQtory for social search and message routing.