A fundamental challenge in contextual bandits is to develop flexible,
general-purpose algorithms with computational requirements no worse than
classical supervised learning tasks such as classification and regression.
Algorithms based on regression have shown promising empirical succes