Practical Lessons from Predicting Clicks on Ads at Facebook

ADKDD'14
Facebook

1. Practical Lessons from Predicting Clicks on Ads at Facebook 2014/ 1/ 27 ( Tue. ) Chang Wei-Yuan @ MakeLab Lab Meeting Facebook ADKDD ‘14

2. Outline Introduction Method Decision tree feature transforms Logistic regression for linear classifier Data freshness Online data joiner Experiment Conclusion Thought 2

3. Introduction 3

4. Outline Introduction Method Decision tree feature transforms Logistic regression for linear classifier Data freshness Online data joiner Experiment Conclusion Thought 4

5. Outline Introduction Method Decision tree feature transforms Logistic regression for linear classifier Data freshness Online data joiner Experiment Conclusion Thought 5

6. 6

7. Outline Introduction Method Decision tree feature transforms Logistic regression for linear classifier Data freshness Online data joiner Experiment Conclusion Thought 7

8. Decision tree feature transforms 8

9. Outline Introduction Method Decision tree feature transforms Logistic regression for linear classifier Data freshness Online data joiner Experiment Conclusion Thought 9

10. Logistic regression for linear classifier Stochastic Gradient Descent (SGD) algorithm the tunable parameters are optimized by grid search Bayesian online learning scheme for profit regression 10

11. One advantages of LR over BOPR is that the model size is half the smaller model size may lead to better cache locality and thus faster cache lookup 11

12. Outline Introduction Method Decision tree feature transforms Logistic regression for linear classifier Data freshness Online data joiner Experiment Conclusion Thought 12

13. Data freshness Click prediction systems are often deployed in dynamic environments where the data distribution changes over time. 13

14. 14

15. 15 These f indings indicate that it is worth retraining on a daily basis .

16. Batch one option would be to have a recurring daily job that retrains the models, possibly in batch Concurrency the training can be done via concurrency in a multi-core machine with large amount of memory 16

17. Outline Introduction Method Decision tree feature transforms Logistic regression for linear classifier Data freshness Online data joiner Experiment Conclusion Thought 17

18. Online data joiner The boosted decision trees can be trained daily, but the linear classier can be trained in near real -time by online learning . O nline J oiner g enerates real -time training data used to train the linear classifier via online learning 18

19. Online data joiner The boosted decision trees can be trained daily, but the linear classier can be trained in near real -time by online learning . O nline J oiner g enerates real -time training data used to train the linear classifier via online learning H ow to label for a new instance ? 19

20. Online data joiner The boosted decision trees can be trained daily, but the linear classier can be trained in near real -time by online learning . O nline J oiner g enerates real -time training data used to train the linear classifier via online learning perform a distributed stream-to-stream join on ad impressions and ad clicks 20

21. 21

22. Outline Introduction Method Decision tree feature transforms Logistic regression for linear classifier Data freshness Online data joiner Experiment Conclusion Thought 22

23. Experiment Number of boosting trees 23

24. Experiment Boosting feature importance 24

25. Experiment Boosting feature importance 25

26. Experiment Historical features 26

27. Experiment Historical features 27

28. Outline Introduction Method Decision tree feature transforms Logistic regression for linear classifier Data freshness Online data joiner Experiment Conclusion Thought 28

29. Conclusion This has inspired a promising hybrid model architecture for click prediction . b oosted decision trees and a linear classier online learning method with real-time training data 29

30. Outline Introduction Method Decision tree feature transforms Logistic regression for linear classifier Data freshness Online data joiner Experiment Conclusion Thought 30

31. Thanks for listening. 2014 / 1 / 27 ( Tue. ) @ MakeLab Lab Meeting [email protected]