Some remarks

Some remarks on the task:

  • The context is a set of 136 binary features
  • You usually have to choose you favorite arm in a set of 30 arms
  • The clickthrough rate (CTR) is not constant over time, you can try to take advantage of it.
  • There is a total of 30 millions of records. 10 millions are used for the first part of the challenge.
  • The record will be used for evaluation of your algorithm only 1/30 of the time (data are collected with a random policy). To fulfill the time constraint this means you should be able to choose an arm within 5 ms and to update your policy within 50 ms when receiving feedback.
  • The total number of documents (ie articles displayed) in the first part of the challenge is 246 (timestamp 1317513291 to 1317945293).  The total number of documents in the full dataset is 652 (timestamp 1317513291 to 1318809293).
  • Data are presented in chronological order. So between two  consecutive users possible choices tends to be same but evolves over time.
  • Evaluation is done on our servers to simulate online learning and avoid the use of some offline methods on the whole set of data. If for test purpose you want more datas consider the R6 data set from Yahoo!
  • In phase 1, winners will be known at the beginning of June, these winners are strongly encouraged to present their work at the workshop.
  • Phase 2 results will be known only at the workshop, it will be the same procedure of evaluation but with more (and new) data. Participants cannot submit any new algorithm, we will use their best submission of phase 1.
  • If you dislike Java, you can contact us and rewrite the evaluator under GPL license (specs will be provided). After reviewing your code, we may accept it and allow a new language of coding for all the participants. Time limit will remain the same whatever the language.
  • After the challenge the data will be made available through the Yahoo! Webscope program.