Using Machine Learning in the delivery of ads
•
0 likes•325 views
In this presentation, I explained that a simple machine learning model can be implemented in some situations even if we do not have the perfect conditions. We go over the first iteration of a click prediction model used in the delivery of CPC ads.
Recommended
More Related Content
Similar to Using Machine Learning in the delivery of ads
Similar to Using Machine Learning in the delivery of ads (20)
More from Ruth Garcia Gavilanes
More from Ruth Garcia Gavilanes (10)
Recently uploaded
Recently uploaded (20)
Using Machine Learning in the delivery of ads
- 1. Using Machine Learning in the delivery of ads Ruth Garcia NDR Conference – June 07, 2018 Iasi, Romania
- 2. What is Skyscanner? Skyscanner is a leading global travel search site offering a comprehensive and free flight search service as well as online comparisons for hotels, car hire and now trains.
- 3. Data Science at Skyscanner Decision Science:1. Ensuring we have the right data; insights and information to make the most impactful, scientific decisions in every aspect of our operations. Building Data Products:2. Leveraging our vast wealth of data to build more contextual, relevant products for Travellers and Travel Suppliers (our Partners) D.S Eng.
- 4. Machine Learning at Skyscanner • Destination recommendations • Itinerary mashups • Advertising Recommendations Advertising Mashups
- 6. Not always annoying ads
- 7. Advertising schemes Brand awareness Traffic Action Machine Learning efforts
- 8. Advertising schemes Brand awareness Traffic Action
- 9. Advertising Schemes Brand awareness Traffic Action Machine Learning efforts
- 11. Ads at Skyscanner = No need to worry about how ads are delivered ? Make sure, ads are delivered
- 12. Delivery of Ads Data ownership ? Technical difficulties Black box algorithms Skyscanner Ads Manager Solution
- 13. Click prediction algorithm Click prediction algorithm Rankingalgorithm Candidates
- 14. Cloud services and tools at Skyscanner Languages AWS services Batch S3
- 15. Overview of an Advertising System
- 16. Expectation vs. reality Tensorfiow Optimization technique• Embeddings• Crossed columns• Hashing• Expectation Flexible but with the risk of not being fast enough. Features: User history,• User features,• Route features ,• Ad features with,• colors, text
- 17. Expectation vs. reality Reality Not flexible but fast and easier to implement. Tensorfiow Optimization technique• Embeddings• Crossed columns• Hashing• Features: User history,• User features,• Route features ,• Ad features with,• colors, text
- 18. Challenges in building the model Goal: model, that estimates the likelihood of whether an impression will result in a click or not.
- 19. Challenges: Which model to use? Model Possibilities (easy to read in node.js): • Logistic regression • Random Forest : gets lost • Neural networks: too slow hard to put it in json Solvers: • Logistic regression: Liblinear, sag • SGDClassifier Train all data at once Train data in batches Saves memory Gridsearch for hyperparameteres
- 20. Challenges: Categorical values Categorical values: One hot• encoding One hot encoding grouping less frequent features• Hashing trick• One hot encoding• • One hot encoding grouping • Less frequent features Creatives C1 C2 C3 C1 C2 C3 1 0 0 0 1 0 0 0 1 Creatives C1 C2 C3 C4 C1 C2 C3 C4 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1
- 21. Challenges: Categorical values id features 123 creative1, advertiser2,mobile, etc. 321 creative2, advertiser4,mobile, etc. id Feat_1 Feat_2 Feat_3 …. Feat_k 123 1 0 1 …. 0 321 1 0 0 …. 1 Hashing Trick: Same hashing function in training, testing and production• We gain
- 22. Machine Learning Performance Precision at 1: based on target groups. Mean Reciprocal Rank AUC: if caring about ranking Log-Loss: if caring about the value of CTR Other metrics:
- 23. Challenges: How to monitor performance? Updating model based on different Sampling methods.
- 24. The road ahead: Balancing exploitation and exploration Choose ad based on ONLY CTR Choose ad based on OTHER criteria
- 25. The road even further ahead: quality Post-click, colors, images, text, landing page
- 26. Conclusions 1. Machine Learning can bring many benefits to the product the challenge is to prove it 2. Bringing ML into production could hard at the beginning 3. Cooperation between data scientists and engineers is crucial D.S Eng.