Adapting the nearest neighbour rule to contextual bandit problems leads to an efficient algorithm with no assumptions about data generation.
