Efficient learning by implicit exploration in bandit problems with side observations