activ explor learn rank clickthrough data filip radlinski thorsten joachim address task learn rank document search enginelog user behavior previou work problem reli onpass collect clickthrough data contrast show anact explor strategi provid data lead fasterlearn specif develop bayesian approach selectingrank present user interact result informativetrain data result trec web corpu assynthet data demonstr direct explor strategi quicklylead user present improv rank onlin learningset find activ explor substanti outperformspass observ random explor 