This is a clear accept: all reviewers liked the paper, and I agree with their recommendation, as the paper provides a nice combination of fixed share (with specialists) with graph predictions. The authors are encouraged to include the lower bound. Also, the strength of the paper could be emphasized very clearly by comparing to applying meta-algorithms, such as those of [12-or rather its journal version, 13, 23] (these algorithms are specifically equipped to combine tracking with a large structured predictor class, at the price of a log T increase in complexity). Finally, I'd like to mention that two of the three reviewers were experts in proving regret bounds.