63

Beyond Bandit Feedback in Online Multiclass Classification

Abstract

We study the problem of online multiclass classification in a setting where the learner's feedback is determined by an arbitrary directed graph. While including bandit feedback as a special case, feedback graphs allow a much richer set of applications, including filtering and label efficient classification. We introduce Gappletron, the first online multiclass algorithm that works with arbitrary feedback graphs. For this new algorithm, we prove surrogate regret bounds that hold, both in expectation and with high probability, for a large class of surrogate losses. Our bounds are of order BρKTB\sqrt{\rho KT}, where BB is the diameter of the prediction space, KK is the number of classes, TT is the time horizon, and ρ\rho is the domination number (a graph-theoretic parameter affecting the amount of exploration). In the full information case, we show that Gappletron achieves a constant surrogate regret of order B2KB^2K. We also prove a general lower bound of order max{B2K,T}\max\big\{B^2K,\sqrt{T}\big\} showing that our upper bounds are not significantly improvable. Experiments on synthetic data show that for various feedback graphs, our algorithm is competitive against known baselines.

View on arXiv
Comments on this paper