We study an online multi-task learning setting, in which instances of related\ntasks arrive sequentially, and are handled by task-specific online learners. We\nconsider an algorithmic framework to model the relationship of these tasks via\na set of convex constraints. To exploit this relationship, we design a novel\nalgorithm -- COOL -- for coordinating the individual online learners: Our key\nidea is to coordinate their parameters via weighted projections onto a convex\nset. By adjusting the rate and accuracy of the projection, the COOL algorithm\nallows for a trade-off between the benefit of coordination and the required\ncomputation/communication. We derive regret bounds for our approach and analyze\nhow they are influenced by these trade-off factors. We apply our results on the\napplication of learning users' preferences on the Airbnb marketplace with the\ngoal of incentivizing users to explore under-reviewed apartments.\n