Exploration & Exploitation Challenge 2011
Download as PPT, PDF0 likes718 views
The document outlines a competition focused on website optimization through an exploration and exploitation challenge. Participants needed to maximize engagement by predicting which visitor-option pairs would likely result in clicks, using a dataset of 20 million anonymized records. The competition was divided into two phases, with phase 2 showing significant performance improvement, particularly from the Inria team, which achieved a notable uplift in scoring compared to random selections.
Exploration & Exploitation Challenge 2011
- 1. Exploration & Exploitation Challenge http://explo.cs.ucl.ac.uk /
- 2. Schedule 14.00 - Website optimisation at Adobe Challenge presentation Phase 2 results 14.25 - INRIA team 14.50 - Orange Labs team
- 4. Website optimisation For a given visitor v, choose content to display on a webpage. 1 out of N options. Objective: maximise engagement=clicks Input: (v,1) ... (v, N) pairs Output: index of the pair for which a click is most likely
- 5. The data 20 million anonymised records: (visitor feature, option index, click indicator) 120 continuous and nominal features 6 options, all with same CTR (0.24%) For a given v , only one (v,o) pair out of N will be in the data
- 6. The task Input : batch of 6 visitor-option pairs Output : index of the pair most likely to be associated to a click If click, get reward of 1 Only observe reward for selected pair Maximise cumulated reward (=score)
- 7. Evaluation Submit ClickPredictor.jar Phase 1: live leaderboard from 14 Mar to 6 May 500,000 batches (~ 6 weeks) no initial knowledge about the data live leaderboard logs: reward, time, memory at each iteration Phase 2: only one submission from 13 May to 1 Jun 2,810,084 batches (~ 34 weeks) phase 1 data has been revealed
- 8. Resources Sun Grid Engine at UCL 100ms per batch 4GB per node -> 3.5GB JVM -> 1.75GB
- 9. Phase 1 #1 Olivier Nicol INRIA, SequeL 2170 #2 Christophe Salperwyck Orange Labs 2072 #3 Aurélien Garivier CNRS / Telecom ParisTech 2047 #4 Olivier Cappé CNRS / Telecom ParisTech 2031 #5 Jérémie Mary INRIA, SequeL 1987 #6 Tanguy Urvoy Orange Labs 1714 #7 Martin Antenreiter MUL 1669 #8 Ronald Ortner MUL 1644
- 10. Phase 1 #1 INRIA, SequeL 2170 #2 Orange Labs 2072 #3 CNRS / Telecom ParisTech 2047 #4 MUL 1669 Random 1177
- 11. Phase 1
- 12. Phase 1
- 13. Phase 2 #1 INRIA, SequeL 11529 #2 Orange Labs 10419 #3 CNRS / Telecom ParisTech 9990 #4 MUL 8049 Random 5598
- 14. Congratulations!
- 15. Phase 2
- 16. Uplift = s/r - 1 where s is the score of the algorithm and r is the score of random 0% if not using visitor features 106% for the INRIA algorithm Uplift
- 17. Phase 2 Rank Name Affiliation Total time Score Uplift #1 Olivier Nicol INRIA 3h 40m 11529 106% #2 Christophe Salperwyck Orange 29h 50m 10419 86% #3 Tanguy Urvoy Orange 4h 10179 82% #4 Aurélien Garivier CNRS 1h 17m 9990 78% #5 Martin Antenreiter MUL 20h 8049 44% Random 1h 12m 5598 0%
- 18. Stochastic algorithms Batches presented in the same order, but elements in a batch presented in different orders at each evaluation Luck?
- 19. Resources 78h theoretical max running time (100ms per batch) the INRIA algorithm only took 3h 40m 1.75GB of memory available, 20GB of data
Editor's Notes
- #2: These are my notes
- #3: Questions -> interrupt me
- #5: The challenge is about finding good algorithms to do that. We can’t evaluate algorithms live, but on offline data (in an online fashion).
- #6: Simulated data that has the characteristics of the actual data that can be observed, but which is such that all options have same CTR.
- #7: Batches: - All visitors in a batch are different and have never been seen before. - There can be several clicks or no click in a batch. - All options might not be represented in a batch. Remarks - Need to learn a mapping from (visitor, option) to reward, and need to optimise the cumulated reward: exploration and exploitation trade-off. - Visitor responses might change through time, making it essential to keep learning their interests. - Because the CTRs for each option are the same, it is necessary to use the visitor features if we want to make better predictions than random
- #19: 2 remarks. First, it’s not sure that Christophe’s algorithm is better than Tanguy’s.